The K reference

Guido B 1

Joan D 1

Michaël P 2

Gilles V A 1

http://keccak.noekeon.org/

Version 3.0 1 STMicroelectronics

January 14, 2011 2 NXP Semiconductors
The K reference

2 / 69
Contents

1 K specifications 7
1.1 Conventions and notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.1.1 Bitstrings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.1.2 Padding rules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.2 The K - f permutations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.3 The sponge construction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.4 The K sponge functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.5 Security claim for the K sponge functions . . . . . . . . . . . . . . . . . . 9
1.6 Parts of the state . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

2 The K - f permutations 13
2.1 Translation invariance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.2 The Matryoshka structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.3 The step mappings of K -f . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.3.1 Properties of χ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.3.2 Properties of θ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.3.3 Properties of π . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.3.4 Properties of ρ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.3.5 Properties of ι . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.3.6 The order of steps within a round . . . . . . . . . . . . . . . . . . . . . 23
2.4 Differential and linear cryptanalysis . . . . . . . . . . . . . . . . . . . . . . . . 23
2.4.1 A formalism for describing trails adapted to K -f . . . . . . . . . 23
2.4.2 The Matryoshka consequence . . . . . . . . . . . . . . . . . . . . . . . . 24
2.4.3 The column parity kernel . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.4.4 One and two-round trails . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.4.5 Three-round trails: kernel vortices . . . . . . . . . . . . . . . . . . . . . 26
2.4.6 Beyond three-round trails: choice of π . . . . . . . . . . . . . . . . . . . 27
2.4.7 Truncated trails and differentials . . . . . . . . . . . . . . . . . . . . . . 29
2.4.8 Other group operations . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.4.9 Differential and linear cryptanalysis variants . . . . . . . . . . . . . . . 29
2.5 Solving constrained-input constrained-output (CICO) problems . . . . . . . . 30
2.6 Strength in keyed mode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.7 Symmetry weaknesses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

3 Trail propagation in K -f 33
3.1 Relations between different kinds of weight . . . . . . . . . . . . . . . . . . . . 33
3.2 Propagation properties related to the linear step θ . . . . . . . . . . . . . . . . 35
3.3 Exhaustive trail search . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.3.1 Upper bound for the weight of two-round trails to scan . . . . . . . . . 36

3 / 69
The K reference CONTENTS

3.3.2 Constructing two-round trails . . . . . . . . . . . . . . . . . . . . . . . 37

3.3.3 Extending trails . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
3.3.4 Linear and differential trail bounds for w ≤ 8 . . . . . . . . . . . . . . . 40
3.4 Tame trails . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.4.1 Construction of tame trails . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.4.2 Bounds for three-round tame trails . . . . . . . . . . . . . . . . . . . . . 42
3.4.3 Bounds for four-round tame trails . . . . . . . . . . . . . . . . . . . . . 44

4 Analysis of K -f 45
4.1 Algebraic normal form . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
4.1.1 Statistical tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
4.1.2 Symmetric trails . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
4.1.3 Slide a acks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
4.2 Solving CICO problems algebraically . . . . . . . . . . . . . . . . . . . . . . . 48
4.2.1 The goal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
4.2.2 The supporting so ware . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
4.2.3 The experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
4.2.4 Third-party analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
4.3 Properties of K - f [25] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
4.3.1 Algebraic normal statistics . . . . . . . . . . . . . . . . . . . . . . . . . 51
4.3.2 Differential probability distributions . . . . . . . . . . . . . . . . . . . . 52
4.3.3 Correlation distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
4.3.4 Cycle distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
4.4 Distinguishers exploiting low algebraic degree . . . . . . . . . . . . . . . . . . 59
4.4.1 Zero-sum distinguishers . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
4.4.2 Pre-image a acks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62

5 Design rationale summary 63

5.1 Choosing the sponge construction . . . . . . . . . . . . . . . . . . . . . . . . . 63
5.2 Choosing an iterated permutation . . . . . . . . . . . . . . . . . . . . . . . . . 64
5.3 Designing the K - f permutations . . . . . . . . . . . . . . . . . . . . . . . 64
5.4 Strength estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65

4 / 69
Introduction

In this document we specify, analyze and motivate the design of the cryptographic primitive
K . Based on the sponge construction, K inherits many of its features. We gath-
ered all our analysis on sponge functions in a separate document titled Cryptographic sponge
functions [8]. Reading it is a requisite for understanding the usability and security properties
of K and the security requirements for K - f , the permutation used in K .
Other documents and files are of interest to the readers of this K reference.

• This document comes with a set of files containing results of tests and experiments,
available from http://keccak.noekeon.org/.

• The implementation aspects are covered in a separate document K implementation

overview, which treats so ware and hardware techniques and results, with or without
protection against side-channel a acks [10].

• Also, K T , an open-source so ware aimed at helping analyze K [9].

This document is organized as follows. Chapter 1 contains the formal specifications of

K . The subsequent three chapters are dedicated to the K - f permutations:

• Chapter 2 explains the properties of the building blocks of K - f and motivates the
choices made in its design.

• Chapter 3 is dedicated to trail propagation in K -f.

• Chapter 4 covers all other analysis of K -f.

Finally, Chapter 5 summarizes the design choices behind K and contains our estimation
of the safety margin of K .

Acknowledgments
We wish to thank (in no particular order) Charles Bouillaguet and Pierre-Alain Fouque for
discussing their results later published in [13] with us, Dmitry Khovratovich for discussing
with us the results published in [29] and for his analysis in [1], Jean-Philippe Aumasson for
his analysis in [1] and [2], Joel Lathrop for his analysis in [34] and Willi Meier for his analysis
in [2], Anne Canteaut and Christina Boura for their analysis in [15, 14, 16], Christophe De
Cannière for his analysis in [16], Paweł Morawiecki and Marian Srebrny for their analysis
in [36], Dan Bernstein for his analysis in [3], Ming Duan and Xuejia Lai for their analysis in
[26], Yves Moulart, Bernard Kasser and all our colleagues at STMicroelectronics and NXP
Semiconductors for creating the working environment in which we could work on this. Fi-
nally we would like to thank Agentschap voor Innovatie door Wetenschap en Technologie (IWT)
for funding two of the authors (Joan Daemen and Gilles Van Assche).

5 / 69
The K reference CONTENTS

6 / 69
Chapter 1

K specifications

K (pronounced [kɛtʃak]) is a family of sponge functions [8] that use as a building block a
permutation from a set of 7 permutations. In this chapter, we introduce our conventions and
notation, specify the 7 permutations underlying K and the K sponge functions.
We also give conventions for naming parts of the K state.

1.1 Conventions and notation

We denote the absolute value of a real number x is denoted by | x |.

1.1.1 Bitstrings
We denote the length in bits of a bitstring M by | M |. A bitstring M can be considered as a
sequence of blocks of some fixed length x, where the last block may be shorter. The number
of blocks of M is denoted by | M | x . The blocks of M are denoted by Mi and the index ranges
from 0 to | M | x − 1.
We denote the set of all bitstrings including the empty string by Z2∗ and excluding the
empty string by Z2+ . The set of infinite-length bitstrings is denoted by Z2∞ .

1.1.2 Padding rules

For the padding rule we use the following notation: the padding of a message M to a se-
quence of x-bit blocks is denoted by M ||pad[ x ](| M |). This notation highlights that we only
consider padding rules that append a bitstring that is fully determined by the bitlength of M
and the block length x. We may omit [ x ], (| M|) or both if their value is clear from the context.
K makes use of the multi-rate padding.
Definition 1. Multi-rate padding, denoted by pad10∗ 1, appends a single bit 1 followed by the
minimum number of bits 0 followed by a single bit 1 such that the length of the result is a multiple of
the block length.
Multi-rate padding appends at least 2 bits and at most the number of bits in a block plus
one.

1.2 The K - f permutations

There are 7 K - f permutations, indicated by K - f [b], where b = 25 × 2ℓ and ℓ
ranges from 0 to 6. K - f [b] is a permutation over Z2b , where the bits of s are numbered

7 / 69
The K reference 1. K specifications

from 0 to b − 1. We call b the width of the permutation.

The permutation K - f [b] is described as a sequence of operations on a state a that
is a three-dimensional array of elements of GF(2), namely a[5][5][w], with w = 2ℓ . The
expression a[ x ][y][z] with x, y ∈ Z5 and z ∈ Zw , denotes the bit in position ( x, y, z). It
follows that indexing starts from zero. The mapping between the bits of s and those of a is
s[w(5y + x ) + z] = a[ x ][y][z]. Expressions in the x and y coordinates should be taken modulo
5 and expressions in the z coordinate modulo w. We may sometimes omit the [z] index, both
the [y][z] indices or all three indices, implying that the statement is valid for all values of the
omi ed indices.
K - f [b] is an iterated permutation, consisting of a sequence of nr rounds R, indexed
with ir from 0 to nr − 1. A round consists of five steps:
R = ι ◦ χ ◦ π ◦ ρ ◦ θ, with

4 4
θ : a[ x ][y][z] ← a[ x ][y][z] + ∑ a[ x − 1][y′ ][z] + ∑ a[ x + 1][y′ ][z − 1],
′
y =0 ′
y =0
ρ : a[ x ][y][z] ← a[ x ][y][z − (t + 1)(t + 2)/2],
( )t ( ) ( )
0 1 1 x
with t satisfying 0 ≤ t < 24 and = in GF(5)2×2 ,
2 3 0 y
or t = −1 if x = y = 0,
( ) ( ) ( ′)
′ ′ x 0 1 x
π: a[ x ][y] ← a[ x ][y ], with = ,
y 2 3 y′
χ: a[ x ] ← a [ x ] + ( a [ x + 1] + 1) a [ x + 2],
ι: a ← a + RC[ir ].
The additions and multiplications between the terms are in GF(2). With the exception of
the value of the round constants RC[ir ], these rounds are identical. The round constants are
given by (with the first index denoting the round number)
RC[ir ][0][0][2 j − 1] = rc[ j + 7ir ] for all 0 ≤ j ≤ ℓ,
and all other values of RC[ir ][ x ][y][z] are zero. The values rc[t] ∈ GF(2) are defined as the
output of a binary linear feedback shi register (LFSR):
( )
rc[t] = x t mod x8 + x6 + x5 + x4 + 1 mod x in GF(2)[ x ].

The number of rounds nr is determined by the width of the permutation, namely,

nr = 12 + 2ℓ.

1.3 The sponge construction

The sponge construction [8] builds a function [ f , pad, r ] with variable-length input
and arbitrary output length using a fixed-length permutation (or transformation) f , a padding
rule “pad” and a parameter bitrate r. The permutation f operates on a fixed number of bits,
the width b. The value c = b − r is called the capacity.
For the padding rule we use the following notation: the padding of a message M to a
sequence of x-bit blocks is denoted by M ||pad[ x ](| M |), where | M | is the length of M in bits.
Initially, the state has value 0b , called the root state. The root state has a fixed value and
shall never be considered as an input. This is crucial for the security of the sponge construc-
tion.

8 / 69
1. K specifications The K reference

Algorithm 1 The sponge construction [ f , pad, r ]

Require: r < b

Interface: Z = sponge( M, ℓ) with M ∈ Z2∗ , integer ℓ > 0 and Z ∈ Z2ℓ

P = M ||pad[r ](| M|)
s = 0b
for i = 0 to | P|r − 1 do
s = s ⊕ ( Pi ||0b−r )
s = f (s)
end for
Z = ⌊ s ⌋r
while | Z |r r < ℓ do
s = f (s)
Z = Z ||⌊s⌋r
end while
return ⌊ Z ⌋ℓ

1.4 The K sponge functions

We define the sponge function denoted by K [r, c] by applying the sponge construction
as specified in Algorithm 1 with K - f [r + c], multi-rate padding and the bitrate r.

K [r, c] , [K - f [r + c], pad10∗ 1, r ].

This specifies K [r, c] for any combination of r > 0 and c such that r + c is a width
supported by the K - f permutations.
The default value for r is 1600 − c and the default value for c is 576:

K [c] ,K [r = 1600 − c, c],

K [] ,K [c = 576].

1.5 Security claim for the K sponge functions

For each of the supported parameter values, we make a flat sponge claim [8, Section “The flat
sponge claim”].

Claim 1. The expected success probability of any a ack against K [r, c] with a workload equiv-
alent to N calls to K - f [r + c] or its inverse shall be smaller than or equal to that for a random
oracle plus
( )
1 − exp − N ( N + 1)2−(c+1) .

We exclude here weaknesses due to the mere fact that K - f [r + c] can be described com-
pactly and can be efficiently executed, e.g., the so-called random oracle implementation impossibility
[8, Section “The impossibility of implementing a random oracle”].

Note that the claimed capacity is equal to the capacity used by the sponge construction.

9 / 69
The K reference 1. K specifications

1.6 Parts of the state

In this subsection, we define names of parts of the K - f state, as illustrated in Figure 1.1.
This naming convention may help use a common terminology when analyzing or describing
properties of K -f.
The one-dimensional parts are:

• A row is a set of 5 bits with constant y and z coordinates.

• A column is a set of 5 bits with constant x and z coordinates.

• A lane is a set of w bits with constant x and y coordinates.

The two-dimensional parts are:

• A sheet is a set of 5w bits with constant x coordinate.

• A plane is a set of 5w bits with constant y coordinate.

• A slice is a set of 25 bits with constant z coordinate.

10 / 69
1. K specifications The K reference

Figure 1.1: Naming conventions for parts of the K - f state

11 / 69
The K reference 1. K specifications

12 / 69
Chapter 2

The K - f permutations

This chapter discusses the properties of the K - f permutations that are relevant for the
security of K . A er discussing some structural properties, we treat the different map-
pings that make up the round function. This is followed by a discussion of differential and
linear cryptanalysis to motivate certain design choices. Subsequently, we briefly discuss the
applicability of a number of cryptanalytic techniques to K -f.

2.1 Translation invariance

Let b = τ ( a) with τ a mapping that translates the state by 1 bit in the direction of the z
axis. For 0 < z < w we have b[ x ][y][z] = a[ x ][y][z − 1] and for z = 0 we have b[ x ][y][0] =
a[ x ][y][w − 1]. Translating over t bits gives b[ x ][y][z] = a[ x ][y][(z − t) mod w]. In general,
a translation τ [t x ][ty ][tz ] can be characterized by a vector with three components (t x , ty , tz )
and this gives:

b[ x ][y][z] = a[( x − t x ) mod 5][(y − ty ) mod 5][(z − tz ) mod w] .

Now we can define translation-invariance.

Definition 2. A mapping α is translation-invariant in direction (t x , ty , tz ) if

τ [t x ][ty ][tz ] ◦ α = α ◦ τ [t x ][ty ][tz ] .

Let us now define the z-period of a state.

Definition 3. The z-period of a state a is the smallest integer d > 0 such that:

∀ x, y ∈ Z5 and ∀z ∈ Zw : a[ x ][y][(z + d) mod w] = a[ x ][y][z] .

13 / 69
The K reference 2. The K - f permutations

It is easy to prove the following properties:

• The z-period of a state divides w.

• A state a with z-period d can be represented by the lane size w, its z-period d, and its d
first slices a[.][.][z] with z < d. We call this the z-reduced representation of a.

• For a given w, the z-period defines a partition on the states.

• The number of states with z-period d is zero if d does not divide w and fully determined
by d only, otherwise.

• For w values that are a power of two (the only ones allowed in K ), the state space
consists of the states with z-period 1, 2, 22 up to 2ℓ = w.

• The number of states with z-period 1 is 225 . The number of states with z-period 2d for
d d −1
d ≥ 1 is 22 25 − 22 25 .

• There is a one-to-one mapping between the states a′ with z-period d for any lane length
w that is a multiple of d and the states a with z-period d of lane length d: a′ [.][.][z] =
a[.][.][z mod d].

• If α is translation-invariant in the direction of the z axis, the z-period of α( a) divides

the z-period of a. Moreover, the z-reduced state of α( a) is independent of w.

• If α is injective and translation-invariant in the direction of the z axis, α preserves the

z-period.

2.2 The Matryoshka structure

With the exception of ι, all step mappings of the K - f round function are translation-
invariant in the direction of the z axis. This allows the introduction of a size parameter that
can easily be varied without having to re-specify the step mappings. As in several types
of analysis abstraction can be made of the addition of constants, this allows the re-use of
structures for small width versions as symmetric structures for large width versions. We
refer to Section 2.4.2 for an example. As the allowed lane lengths are all powers of two,
every smaller lane length divides a larger lane length. So, as the propagation structures for
smaller width version are embedded as symmetric structure in larger width versions, we call
it Matryoshka, a er the well-known Russian dolls.

2.3 The step mappings of K -f

A round is composed from a sequence of dedicated mappings, each one with its particular
task. The steps have a simple description leading to a specification that is compact and in
which no trapdoor can be hidden.
Mapping the lanes of the state, i.e., the one-dimensional sub-arrays in the direction of the
z axis, onto CPU words, results in simple and efficient so ware implementation for the step
mappings. We start the discussion of each of the step mappings by pseudocode where the
variables a[ x, y] represent the old values of lanes and A[ x, y] the new values. The operations
on the lanes are limited to bitwise Boolean operations and rotations. In our pseudocode we
denote by ROT( a, d) a translation of a over d bits where bit in position z is mapped to position
z + d mod w. If the CPU word length equals the lane length, the la er can be implemented

14 / 69
2. The K - f permutations The K reference

Figure 2.1: χ applied to a single row

with rotate instructions. Otherwise a number of shi and bitwise Boolean instructions must
be combined or bit-interleaving can be applied [10].
In this section we discuss the difference propagation and input-output correlation prop-
erties of the different mappings. We refer to [8, Sections “Differential cryptanalysis” and
“Linear cryptanalysis”] for an introduction of the terminology and concepts.

2.3.1 Properties of χ
Figure 2.1 contains a schematic representation of χ and Algorithm 2 its pseudocode.

Algorithm 2 χ
for y = 0 to 4 do
for x = 0 to 4 do
A[ x, y] = a[ x, y] ⊕ ((NOT a[ x + 1, y]) AND a[ x + 2, y])
end for
end for

χ is the only nonlinear mapping in K - f . Without it, the K - f round function

would be linear. It can be seen as the parallel application of 5w S-boxes operating on 5-
bit rows. χ is translation-invariant in all directions and has algebraic degree two. This has
consequences for its differential propagation and correlation properties. We discuss these
in short in Sections 2.3.1.1 and Section 2.3.1.2 and refer to [20, Section 6.9] for an in-depth
treatment of these aspects.
χ is invertible but its inverse is of a different nature than χ itself. For example, it does not
have algebraic degree 2. We refer to [20, Section 6.6.2] for an algorithm for computing the
inverse of χ.
χ is simply the complement of the nonlinear function called γ, that is used in R G
[4], P [21] and several other ciphers [20]. We have chosen it for its simple nonlinear
propagation properties, its simple algebraic expression and its low gate count: one XOR, one
AND and one NOT operation per state bit.

15 / 69
The K reference 2. The K - f permutations

2.3.1.1 Differential propagation properties

Thanks to the fact that χ has algebraic degree 2, for a given input difference a′ , the space of
′ ′
possible output differences forms a linear affine variety [19] with 2wr (a ,b ) elements. More-
over, the cardinality of a differential ( a′ , b′ ) over χ is either zero or a power of two. The
corresponding (restriction) weight wr ( a′ , b′ ) = wr ( a′ ) is an integer that only depends on the
input difference a′ . A possible differential imposes wr ( a′ ) linear conditions on the bits of
input a.
We now provide a recipe for constructing the affine variety of output differences corre-
sponding to an input difference, applied to a single row. Indices shall be taken modulo 5 (or
in general, the length of the register). We denote by δ(i ) a pa ern with a single nonzero bit
in position i and δ(i, j) a pa ern with only non-zero bits in positions i and j.
We can characterize the linear affine variety of the possible output differences by an offset
A′ and a basis ⟨c j ⟩. The offset is A′ = χ( a′ ). We construct the basis ⟨c j ⟩ by adding vectors to
it while running over the bit positions i:

• If ai′ ai′+1 ai′+2 ai′+3 ∈ {·100, ·11·, 001·}, extend the basis with δ(i ).

• If ai′ ai′+1 ai′+2 ai′+3 = ·101, extend the basis with δ(i, i + 1).

This algorithm is implemented in K T [9]. The (restriction) weight of a difference

is equal to its Hamming weight plus the number of pa erns 001. The all-1 input difference
results in the affine variety of odd-parity pa erns and has weight 4 (or in general the length
of the register minus 1). Among the 31 non-zero differences, 5 have weight 2, 15 weight 3
and 11 weight 4.
A differential ( a′ , b′ ) leads to a number of conditions on the bits of the absolute value a.
Let B = A′ ⊕ b′ = χ( a′ ) ⊕ b′ , then we can construct the conditions on a by running over each
bit position i:

• ai′+1 ai′+2 = 10 imposes the condition ai+2 = Bi .

• ai′+1 ai′+2 = 11 imposes the condition ai+1 ⊕ ai+2 = Bi .

• ai′+1 ai′+2 = 01 imposes the condition ai+1 = Bi .

The generation of these conditions given a differential trail is implemented in K T

[9].

2.3.1.2 Correlation properties

Thanks to the fact that χ has algebraic degree 2, for a given output mask u, the space of
input mask v whose parities have a non-zero correlation with the parity determined by u
form a linear affine variety. This variety has 2wc (v,u) elements, with wc (v, u) = wc (u) the
(correlation) weight function, which is an even integer that only depends on the output mask
u. Moreover, the magnitude of a correlation over χ is either zero or equal to 2−wc (u)/2 .
We now provide a recipe for constructing the affine variety of input masks corresponding
to an output mask, applied to a single row. Indices shall again be taken modulo 5 (or in
general, the length of the register). We use the term 1-run of length ℓ to denote a sequence of
ℓ 1-bits preceded and followed by a 0-bit.

16 / 69
2. The K - f permutations The K reference

We characterize the linear affine variety with an offset U ′ and a basis ⟨c j ⟩ and build the
offset and basis by running over the output mask. First initialize the offset to 0 and the basis
to the empty set. Then for each of the 1-runs as as+1 . . . as+ℓ−1 do the following:

• Add a 1 in position s of the offset U ′ .

• Set i = s, the starting position of the 1-run.

• As long as ai ai+1 = 11 extend the basis with δ(i + 1, i + 3) and δ(i + 2), add 2 to i and
continue.

• If ai ai+1 = 10 extend the basis with δ(i + 1) and δ(i + 2).

This algorithm is implemented in K T [9]. The (correlation) weight of a mask is

equal to its Hamming weight plus the number of 1-runs of odd length. The all-1 output
mask results in the affine variety of odd-parity pa erns and has weight 4 (or in general the
length of the register minus 1). Of the 31 non-zero mask, 10 have weight 2 and 21 have weight
4.

2.3.2 Properties of θ
Figure 2.2 contains a schematic representation of θ and Algorithm 3 its pseudocode.

Algorithm 3 θ
for x = 0 to 4 do
C [ x ] = a[ x, 0]
for y = 1 to 4 do
C [ x ] = C [ x ] ⊕ a[ x, y]
end for
end for
for x = 0 to 4 do
D [ x ] = C [ x − 1] ⊕ ROT(C [ x + 1], 1)
for y = 0 to 4 do
A[ x, y] = a[ x, y] ⊕ D [ x ]
end for
end for

The θ mapping is linear and aimed at diffusion and is translation-invariant in all direc-
tions. Its effect can be described as follows: it adds to each bit a[ x ][y][z] the bitwise sum of
the parities of two columns: that of a[ x − 1][·][z] and that of a[ x + 1][·][z − 1]. Without θ, the
K - f round function would not provide diffusion of any significance. The θ mapping
has a branch number as low as 4 but provides a high level of diffusion on the average. We
refer to Section 2.4.3 for a more detailed treatment of this.
In fact, we have chosen θ for its high average diffusion and low gate count: two XORs
per bit. Thanks to the interaction with χ each bit at the input of a round potentially affects 31
bits at its output and each bit at the output of a round depends on 31 bits at its input. Note
that without the translation of one of the two sheet parities this would only be 25 bits.

2.3.2.1 The inverse mapping

Computing the inverse of θ can be done by adopting a polynomial notation. The state can
be represented by a polynomial in the three variables x, y and z with binary coefficients.

17 / 69
The K reference 2. The K - f permutations

y z z

Figure 2.2: θ applied to a single bit

Here the coefficient of the monomial xi y j zk denotes the value of bit a[i ][ j][k]. The exponents
i and j range from 0 to 4 and the exponent k ranges from 0 to w − 1. In this representa-
tion a translation τ [t x ][ty ][tz ] corresponds with the multiplication by the monomial x tx yty ztz
modulo the three polynomials 1 + x5 , 1 + y5 and 1 + zw . More exactly, the polynomial rep-
resenting the state is an element of a polynomial quotient ⟨ ring defined by ⟩the polynomial
ring over GF(2)[ x, y, z] modulo the ideal generated by 1 + x5 , 1 + y5 , 1 + zw . A translation
corresponds with multiplication by x tx yty ztz in this quotient ring. The z-period of a state a
is d if d is the smallest nonzero integer such that 1 + zd divides a. Let a′ be the polynomial
corresponding to the z-reduced state of a, then a can be wri en as

1 + zw
a = (1 + zd + z2d + . . . + zw−d ) × a′ = × a′ .
1 + zd
When the state is represented by a polynomial, the mapping θ can be expressed as the
multiplication (in the quotient ring defined above) by the following polynomial :
( ) 4
1 + y5
1 + ȳ x + x4 z with ȳ = ∑ yi = 1+y
. (2.1)
i =0

The inverse of θ corresponds with the multiplication by the polynomial that is the inverse
of polynomial (2.1). For w = 64, we have computed this with the open source mathematics
so ware SAGE [38] a er doing a number of manipulations. First, we assume it is of the form
1 + ȳQ with Q a polynomial in x and z only:
( ) ⟨ ⟩
1 + ȳ( x + x4 z) × (1 + ȳQ) = 1 mod 1 + x5 , 1 + y5 , 1 + z64 .

Working this out and using ȳ2 = ȳ yields

−1
⟨ ⟩
Q = 1 + (1 + x + x 4 z ) mod 1 + x5 , 1 + z64 .

The inverse of 1 + x + x4 z can be computed with a variant of the extended Euclidian algo-
rithm for polynomials in multiple variables. At the time of writing this was unfortunately

18 / 69
2. The K - f permutations The K reference

not supported by SAGE. Therefore, we reduced the number of variables to one by using the
change of variables t = x −2 z. We have x = t192 and x4 z = t193 , yielding:
−1
Q = 1 + (1 + t192 + t193 ) mod (1 + t320 ) .

By performing a change in variables from t to x and z again, Q is obtained.

For w < 64, the inverse can simply be found by reducing Q modulo 1 + zw . For w = 1,
the inverse of θ reduces to 1 + ȳ( x2 + x3 ).
For all values of w = 2ℓ , the Hamming weight of the polynomial of θ −1 is of the order b/2.
This implies that applying θ −1 to a difference with a single active bit results in a difference
with about half of the bits active. Similarly, a mask at the output of θ −1 determines a mask
at its input with about half of the bits active.

2.3.2.2 Propagation of linear masks

A linear Boolean function defined by a mask u at the output of a linear function has non-zero
correlation to a single linear Boolean function at its input. Given the matrix representation
of the linear function, it is easy to express the relation between the input and output mask.
Given b = Ma, we have:
uT b = uT Ma = ( MT u)T a .
It follow that uT b is correlated to vT a with v = MT u with correlation 1. We say that a mask
u at the output of a linear mapping M propagates to v = MT u at its input. We denote the
mapping defined by MT the transpose of M.
As θ is linear, we have v = θ T (u), with u a mask at the output of θ, v a mask at its input
and where θ T the transpose of θ. We now determine the expression for the transpose of θ in
the formalism of [5]. Let b = θ ( a) and

∑ u[x][y][z]b[x][y][z] = ∑ v[x][y][z]a[x][y][z] .
x,y,z x,y,z

Filling in the value of b[ x ][y][z] from the specification of θ in [5] and working this out yields:

∑ u[x][y][z]b[x][y][z] =
x,y,z
( )
∑ u[ x ][y][z] + ∑ u[ x + 1][y′ ][z] + ∑ u[ x − 1][y′ ][z + 1] a[ x ][y][z]
x,y,z y′ y′

It follows that:

v = θ T (u) ⇔ v[ x ][y][z] = u[ x ][y][z] + ∑ u[ x + 1][y′ ][z] + ∑ u[ x − 1][y′ ][z + 1] (2.2)

y′ y′

In polynomial notation the application of θ T is a multiplication by

( )
1 + ȳ x4 + xz4 .

2.3.3 Properties of π
Figure 2.3 contains a schematic representation of π and Algorithm 4 its pseudocode.
Note that in an efficient program π can be implemented implicitly by addressing.

19 / 69
The K reference 2. The K - f permutations

Algorithm 4 π
for x = 0 to 4 do
for(y =) 0 to(4 do) ( )
X 0 1 x
=
Y 2 3 y
A[ X, Y ] = a[ x, y]
end for
end for

Figure 2.3: π applied to a slice. Note that x = y = 0 is depicted at the center of the slice.

20 / 69
2. The K - f permutations The K reference

The mapping π is a transposition of the lanes that provides dispersion aimed at long-
term diffusion. Without it, K - f would exhibit periodic trails of low weight. π operates
in a linear way on
(0 1) the coordinates ( x, y): the lane in position ( x, y) goes to position ( x, y) MT ,
with M = 2 3 a 2 by 2 matrix with elements in GF(5). It follows that the lane in the origin
(0, 0) does not change position. As π operates on the slices independently, it is translation-
invariant in the z-direction. The inverse of π is defined by M−1 .
Within a slice, we can define 6 axes, where each axis defines a direction that partitions the
25 positions of a slice in 5 sets:

• x axis: rows or planes;

• y axis: columns or sheets;

• y = x axis: rising 1-slope;

• y = − x axis: falling 1-slope;

• y = 2x axis: rising 2-slope;

• y = −2x axis: falling 2-slope;

The x axis is just the row through the origin, the y axis is the column through the origin, etc.
There are many matrices that could be used for π. In fact, the invertible 2 by 2 matrices
with elements in GF(5) with the matrix multiplication form a group with 480 elements con-
taining elements of order 1, 2, 3, 4, 5, 6, 8, 10, 12, 20 and 24. Each of these matrices defines
a permutation on the 6 axes, and equivalently, on the 6 directions. Thanks to its linearity,
the 5 positions on an axis are mapped to 5 positions on an axis (not necessarily the same).
Similarly, the 5 positions that are on a line parallel to an axis, are mapped to 5 positions on
a line parallel to an axis.
For π we have chosen a matrix that defines a permutation of the axes where they are in a
single cycle of length 6 for reasons explained in Section 2.4.6. Implementing π in hardware
requires no gates but results in wiring.
As π is a linear function, a mask u at the output propagates to the mask v at the input
with v = π T (u) (see Section 2.3.2.2). Moreover, we have π T = π −1 , yielding u = π (v). This
follows directly from the fact that π is a bit transposition and that subsequently its matrix is
orthogonal: MT M = I.

2.3.4 Properties of ρ
Figure 2.4 contains a schematic representation of ρ, while Table 2.1 lists its translation offsets.
Algorithm 5 gives pseudocode for ρ.

Algorithm 5 ρ
A[0,
( )0] =( a[)0, 0]
x 1
=
y 0
for t = 0 to 23 do
A[ x,
( )y] =
( ROT)( a(
[ x,)
y], (t + 1)(t + 2)/2)
x 0 1 x
=
y 2 3 y
end for

21 / 69
The K reference 2. The K - f permutations

Figure 2.4: ρ applied to the lanes. Note that x = y = 0 is depicted at the center of the slices.

x=3 x=4 x=0 x=1 x=2

y =2 153 231 3 10 171
y =1 55 276 36 300 6
y =0 28 91 0 1 190
y =4 120 78 210 66 253
y =3 21 136 105 45 15

Table 2.1: The offsets of ρ

The mapping ρ consists of translations within the lanes aimed at providing inter-slice
dispersion. Without it, diffusion between the slices would be very slow. It is translation-
invariant in the z-direction. The inverse of ρ is the set of lane translations where the constants
are the same but the direction is reversed.
The 25 translation constants are the values defined by i (i + 1)/2 modulo the lane length.
It can be proven that for any ℓ, the sequence i (i + 1)/2 mod 2ℓ has period 2ℓ+1 and that any
sub-sequence with n2ℓ ≤ i < (n + 1)2ℓ runs through all values of Z2ℓ . From this it follows
that for lane lengths 64 and 32, all translation constants are different. For lane length 16,
9 translation constants occur twice and 7 once. For lane lengths 8, 4 and 2, all translation
constants occur equally o en except the translation constant 0, that occurs one time more
o en. For the mapping of the (one-dimensional) sequence of translation constants to the
lanes arranged in two dimensions x and y we make use of the matrix of π. This groups
the lanes in a cycle of length 24 on the one hand and the origin on the other. The non-zero
translation constants are allocated to the lanes in the cycle, starting from (1, 0).
ρ is very similar to the transpositions used in R G [4], P [21] and S -
R U [20]. In hardware its computational cost corresponds to wiring.
As ρ is a linear function, a mask u at the output propagates to the mask v at the input
with v = ρT (u) (see Section 2.3.2.2). Moreover, we have ρT = ρ−1 , yielding u = ρ(v). This
follows directly from the fact that ρ is a bit transposition and that subsequently its matrix is
orthogonal: MT M = I.

2.3.5 Properties of ι
The mapping ι consists of the addition of round constants and is aimed at disrupting sym-
metry. Without it, the round function would be translation-invariant in the z direction and
all rounds would be equal making K - f subject to a acks exploiting symmetry such as
slide a acks. The number of active bit positions of the round constants, i.e., the bit positions

22 / 69
2. The K - f permutations The K reference

in which the round constant can differ from 0, is ℓ + 1. As ℓ increases, the round constants
add more and more asymmetry.
The bits of the round constants are different from round to round and are taken as the
output of a maximum-length LFSR. The constants are only added in a single lane of the state.
Because of this, the disruption diffuses through θ and χ to all lanes of the state a er a single
round.
In hardware, the computational cost of ι is a few XORs and some circuitry for the gener-
ating LFSR. In so ware, it is a single bitwise XOR instruction.

2.3.6 The order of steps within a round

The reason why the round function starts with θ is due to the usage of K - f in the sponge
construction. It provides a mixing between the inner and outer parts of the state. Typically,
the inner part is the part that is unknown to, or not under the control of the adversary. The
order of the other step mappings is arbitrary.

2.4 Differential and linear cryptanalysis

In this section we discuss the differential and linear cryptanalysis aspects that have deter-
mined our choice of step mappings. For a more in-depth discussion on the propagation of
differential and linear trails, we refer to Chapter 3.

2.4.1 A formalism for describing trails adapted to K -f

The propagation of differential and linear trails in K - f is very similar. Therefore we
introduce a formalism for the description of trails that is to a large extent common for both
types of trails. Differential trails describe the propagation of differences through the rounds
of K - f and linear trails the propagation of masks. We will address both with the term
pa erns.
As explained in Section 2.3.1, for a given difference a at the input of χ, the set of possible
output differences is a linear affine variety. For a given mask a at the output of χ, the set
of input masks with non-zero correlation to the given output mask is also a linear affine
variety. Hence, to make the pa ern propagation similar, for differential trails we consider
the propagation from input to output and for linear trails we consider the propagation from
output to input.
A difference at the input of χ is denoted by ai and we call it a pa ern before χ (in round i).
A difference at the output of χ is denoted by bi and we call it the pa ern a er χ. Similarly, a
mask at the output of χ is denoted by ai and we call it a pa ern before χ. A mask at the input
of χ is denoted by bi and we call it the pa ern a er χ. In both cases we denote the linear
affine variety of possible pa erns a er χ compatible with ai by B( ai ).
Thanks to the fact that χ is the only nonlinear step in the round, a difference bi a er χ fully
determines the difference ai+1 before χ of the following round: we have ai = π (ρ(θ (bi ))).
We denote the linear part of the round by λ, so:
λ = π◦ρ◦θ .
Similarly, a mask bi a er χ fully determines the mask ai+1 before the χ of the following round.
Now we have ai = θ T (ρT (π T (bi ))) = θ T (ρ−1 (π −1 (bi ))). Here again, we denote this linear
transformation by λ, so in this case we have:
λ = θ T ◦ ρ −1 ◦ π −1 .

23 / 69
The K reference 2. The K - f permutations

Note that the way B( ai ) is formed depends on whether we consider differential or linear
trails. Moreover, the meaning of λ depends on whether we consider differential or linear
trails.
Consider now the set obtained by applying λ to all elements of B( ai ). Thanks to the
linearity of λ this is again a linear affine variety and we denote it by A( ai ).
We now define a ℓ-round trail Q by a sequence of state pa erns ai with 0 ≤ i ≤ ℓ. Every
ai denotes a state pa ern before χ and ai must be compatible with ai−1 , i.e., ai ∈ A( ai−1 ). We
use bi to denote the pa erns a er χ, i.e., ai+1 = λ(bi ). So we have:
χ λ χ λ χ λ
a0 → b0 → a1 → b1 → a2 → b2 → . . . aℓ . (2.3)

The restriction weight of a differential trail Q is the number of conditions it imposes on

the absolute values on the members of a right pair. It is given by

wr ( Q ) = ∑ wr ( ai ) .
0≤i <ℓ

Note that the restriction weight of the last difference aℓ does not contribute to that of the
trail. Hence the weight of any ℓ-round trail is fully determined by its ℓ first differences. For
weight values well below the width of the permutation, a good approximation for the DP of
a trail is given by DP( Q) ≈ 2−wr (Q) . If wr ( Q) is near the width b, this approximation is no
longer valid due to the fact that the cardinality of a trail is an integer. While the mapping ι
has no role in the existence of differential trails, it does in general impact their DP. For trails
with weight above the width, it can make the difference between having cardinality zero or
non-zero.
The correlation weight of a linear trail over an iterative mapping determines its contri-
bution to a correlation between output and input defined by the masks a0 and aℓ . The corre-
lation weight of a trail is given by

wc ( Q ) = ∑ wc ( ai ) .
0≤i <ℓ

Here also the correlation weight of aℓ does not contribute and hence the weight of any ℓ-
round trail is fully determined by its ℓ first masks. The magnitude of the correlation contri-
bution of a trail is given by 2−wc (Q) . The sign is the product of the correlations over the χ and
ι steps in the trail. The sign of the correlation contribution of a linear trail hence depends on
the round constants.
In our analysis we focus on the weights of trails. As the weight of a ℓ-round trail is
determined by its first ℓ pa erns, in the following we will ignore the last pa ern and describe
ℓ-round trail with only ℓ pa erns ai , namely a0 to aℓ−1 .

2.4.2 The Matryoshka consequence

The existence of trails (both differential and linear) and their weight is independent of ι.
The fact that all other step mappings of the round function are translation-invariant in the
direction of the z axis, makes that a trail Q implies w − 1 other trails: those obtained by
translating the pa erns of Q over any non-zero offset in the z direction. If all pa erns in a
trail have a z-period below or equal to d, this implies only d − 1 other trails.
Moreover, a trail for a given width b implies a trail for all larger widths b′ . The pa erns
are just defined by their z-reduced representations and the weight must be multiplied by
b′ /b. Note that this is not true for the cardinality of differential trails and the sign of the
correlation contribution of linear trails, as these do depend on the round constants.

24 / 69
2. The K - f permutations The K reference

2.4.3 The column parity kernel

The mapping θ is there to provide diffusion. As said, it can be expressed as follows: add to
each bit a[ x ][y][z] the bitwise sum of the parities of two columns: that of a[ x − 1][·][z] and
that of a[ x + 1][·][z − 1]. From this we can see that for states in which all columns have even
parity, θ is the identity. We call this set of states the column parity kernel or CP-kernel for short.
The size of the CP-kernel is 220w as there are in total 2b = 225w states and there are 25w
independent parity conditions. The kernel contains states with Hamming weight values as
low as 2: those with two active bits in a single column. Due to these states, θ only has a
branch number (expressed in Hamming weight) of 4.
The low branch number is a consequence of the fact that only the column parities prop-
agate. One could consider changing θ to improve the worst-case diffusion, but this would
significantly increase the computational cost of θ as well. Instead, we have chosen to address
the CP-kernel issue by carefully choosing the mapping π.
We can compute from a 25w-bit state its 5w-bit column parity pa ern. These pa erns par-
tition the state space in 25w subsets, called the parity classes, with each 220w elements. We
can now consider the branch number restricted to the states in a given parity class. As said,
the minimum branch number that can occur is 4 for the CP-kernel, the parity class with the
all-zero column parity pa ern. Over all other parity classes, the branch number is at least
12.
Note that for states where all columns have odd parity, θ adds 0 to every bit and also acts
as the identity. However, the Hamming weight of states in the corresponding parity class is
at least 5w resulting in a branch number of 10w.

2.4.4 One and two-round trails

Now we will have a look at minimum weights for trails with one and two rounds. The
minimum weight for a one-round differential trail ( a0 ) is obtained by taking a difference a0
with a single active bit and has weight 2. For a linear trail this is obtained by a mask a0 with
a single active bit or two neighboring active bits in the same row, and the weight is also 2.
This is independent of the width of K -f.
For the minimum weight of two-round trails we use the following property of χ: if a
difference before χ restricted to a row has a single active bit, the same difference is a possible
difference a er χ. Hence for difference with zero or one active bits per row, χ can behave as
the identity. Similarly, for masks with zero or one active bits per row, χ can behave as the
identity. We call such trails in which the pa erns at the input and output of χ are the same,
χ-zero trails. Note that all pa erns in a χ-zero trail are fully determined by the first pa ern
a0 .
For all widths, the two-round trails with minimum weight are χ-zero trails. For a differ-
ential trail, we choose for a0 a difference with two active bits that are in the same column. Af-
ter χ the difference has not changed and as it is in the CP-kernel, it goes unchanged through
θ as well. The mappings π and ρ move the two active bits to different columns, but in no
case to the same row. This results in a value of a1 with two active bits in different rows. As
the weight of both a0 and a1 is 4, the resulting trail has weight 8. For linear trails, the two
active bits in a0 must be chosen such that a er ρ and π they are in the same column. with
a similar reasoning it follows that the minimum trail weight is also 8. Note that the low
weight of these trails is due to the fact that the difference at the input of θ in round 0 is in the
CP-kernel.

25 / 69
The K reference 2. The K - f permutations

2.4.5 Three-round trails: kernel vortices

From here on, we concentrate on differential trails as the explanation is very similar for linear
trails. We can construct a three-round χ-zero trail where both differences a0 and a1 are in the
CP-kernel. As in a χ-zero trail χ behaves as the identity and a0 is in the CP-kernel, we have
a1 = π (ρ( a0 )). Hence, we can transfer the conditions that a0 is in the kernel to conditions on
a1 , or vice versa.
We will now look for pa erns a0 where both a0 and π (ρ( a0 )) are in the CP-kernel. a0
cannot be a pa ern with only two active bits in one column since π ◦ ρ maps these bits to
two different columns in a1 .
The minimum number of active bits in a0 is four, where both a0 and a1 have two active
columns with two active bits each. We will denote these four active bits as points 0, 1, 2 and
3. Without loss of generality, we assume these points are grouped two by two in columns
in a0 : {0, 1} in one column and {2, 3} in another one. In a1 we assume they are grouped in
columns as {1, 2} and {3, 0}.
The mapping π maps sheets (containing the columns) to falling 2-slopes and maps planes
to sheets. Hence the points {0, 1} and {2, 3} are in falling 2-slopes in a1 and the points {1, 2}
and {3, 0} are in planes in a0 . This implies that projected on the ( x, y) plane, the four points
of a0 form a rectangle with horizontal and vertical sides. Similarly, in a1 they form a paral-
lelogram with vertical sides and sides that are falling 2-slopes.
The ( x, y) coordinates of the four points in a0 are completely determined by those of the
two opposite corner points ( x0 , y0 ) and ( x2 , y2 ). The four points have coordinates: ( x0 , y0 ),
2
( x0 , y2 ), ( x2 , y2 ) and ( x2 , y0 ). The number of possible choices is (25) = 100. Now let us have
a look at their z coordinates. Points 0 and 1 should be in the same column and points 2 and
3 too. Hence z1 = z0 and z3 = z2 . Moreover, ρ shall map points 1 and 2 to the same slice and
bits 3 and 0 too. This results in the following conditions for their z-coordinates:

z0 + r [ x0 ][y2 ] = z2 + r [ x2 ][y2 ] modw ,

(2.4)
z2 + r [ x2 ][y0 ] = z0 + r [ x0 ][y0 ] modw ,

with r [ x ][y] denoting the translation offset of ρ in position ( x, y). They can be converted to
the following two conditions:

z2 = z0 + r [ x0 ][y2 ] − r [ x2 ][y2 ] modw ,

z2 = z0 + r [ x0 ][y0 ] − r [ x2 ][y0 ] modw .

In any case z0 can be freely chosen, and this determines z2 . Subtracting these two equations
eliminates z0 and z2 and results in:

r [ x0 ][y0 ] − r [ x0 ][y2 ] + r [ x2 ][y2 ] − r [ x2 ][y0 ] = 0 mod w . (2.5)

If this equation is not satisfied, the equations (2.4) have no solution.

Consider now w = 1. In that case Equation (2.5) is always satisfied. However, in order
to be χ-zero, the points must be in different rows, and hence in different planes, both in a0
and a1 , and this is not possible for a rectangle.
If ℓ ≥ 1, Equation (2.5) has a priori a probability of 2−ℓ of being satisfied. Hence, we
can expect about 2−ℓ 100 rectangles to define a state a0 with both a0 and π (ρ( a0 )) in the CP-
kernel. So it is not inconceivable that such pa erns exists for w = 64. This would result in
a 3-round trail with weight of 8 per round and hence a total weight of 24. However, for our
choice of π and ρ, there are no such trails for w > 16.
Note that here also the Matryoshka principle plays. First, the z-coordinate of one of the
points can be freely chosen and determines all others. So, given a rectangle that has a solution

26 / 69
2. The K - f permutations The K reference

for Equation (2.5), there are 2ℓ pa erns a0 , one for each choice of z0 . Second, if Equation (2.5)
′
is not satisfied for ℓ but it is for some ℓ′ < ℓ, it implies a pa ern a0 with 2ℓ−ℓ 4 points rather
than 4 for which both a0 and π (ρ( a0 )) are in the kernel.
These pa erns can be generalized by extending the number of active bits: a pa ern a0
with both a0 and π (ρ( a0 )) in the kernel can be constructed by arranging 2e points in a cycle in
the ( x, y) plane and giving the appropriate z-coordinates. In such a cycle each combination
of points {2i, 2i + 1} are in the same sheet and each combination of points {2i + 1, 2i + 2}
are in the same plane. We call such a cycle of 2e ( x, y) positions a kernel vortex V.
For the z coordinates, the conditions that the points {2i, 2i + 1} are in the same column
in a0 and the points {2i + 1, 2i + 2} are in the same column in a1 results in 2e conditions.
Similar to the rectangle case, these conditions only have a solution if the ρ rotation constants
in the lanes of the cycle satisfy a condition. For a given kernel vortex V, we define its depth
d(V ) as:
2e−1
d (V ) = ∑ (−1)i r [point i ]. (2.6)
i =0

Now, the vortex results in a valid pa ern a0 if d(V ) mod w = 0. We call the largest power
of 2 dividing d(V ) the character of the vortex c(V ). If d(V ) = 0, we say its character is
c(V ) = ∞. Summarizing, a vortex V defines a valid pa ern a0 with 2e active bits for lane
length w ≤ c(V ). For constructing low-weight 3-round trails, it suffices to find vortices with
small e and large character: given a vortex V it results in a 3-round trail with weight 12e
for all values of 2ℓ ≤ c(V ) and with weight 12e2ℓ /c(V ) for all values of 2ℓ > c(V ) (using
symmetric trails of period c(V )).
As the length of vortices grows, so does their number. There are 600 vortices of length
6, 8400 of length 8 and 104040 of length 10. The character c(V ) over these vortices has an
exponential distribution: about half of them has character 1, 1/4 have character 2, 1/8 have
character 4 and so on. It follows that as their length 2e grows, there are more and more
vortices that result in valid pa ern a0 with 2e active bits, even for lane length 64.
Moreover, one can construct pa erns a0 containing two or more vortices, provided that
they do not result in a row with two active bits in either a0 or a1 . The character of such a
combination is just the minimum of the characters of its component vortices. Clearly, due
the large number of kernel vortices, it is likely that there are three-round trails with low
weight for any choice of ρ and π. For our choice of π and ρ, the vortex that leads to the
3-round trail with the smallest weight for K - f is one of length 6 and character 64. It
results in a 3-round trail with weight 36.

2.4.6 Beyond three-round trails: choice of π

We will now try to extend this to four-round trails: we try to find pa erns a0 such that a0 , a1
and a2 are in the CP-kernel.
A vortex of length 4, i.e., with e = 2 cannot do the job with our choice of π: a rectangle
in a0 with sheets and planes as sides results in a parallelogram in a1 with falling 2-slopes
and columns as sides and in a parallelogram in a2 with rising 2-slopes and falling 2-slopes
as sides. Hence the four points in a2 cannot group in columns 2 by 2 and therefore it cannot
be in the kernel.
Consider now a vortex of length 6. We choose the points such that the grouping in
columns is {0, 1}, {2, 3}, {4, 5} in a0 , it is {1, 2}, {3, 4}, {5, 0} in a1 and {1, 4}, {2, 5}, {3, 0}
in a2 . The grouping in a1 simply implies that {1, 2}, {3, 4}, {5, 0} are grouped in planes in a0 .
Actually, the first two groupings are similar to the three-round trail case: they determine a

27 / 69
The K reference 2. The K - f permutations

character c(V ) and fix the z coordinates of all points but one. We will now study the impli-
cations of the grouping in a2 on the ( x, y) coordinates. Grouping in columns (sheets) in a2
implies grouping in planes in a1 and subsequently grouping in rising 1-slopes in a0 .
For the z-coordinates this results in 3 additional conditions: points 1 and 4, points 2 and
5 and points 3 and 0 must have the same z-coordinate in a2 . Similar to Equation (2.4) these
conditions are equalities modulo 2ℓ . For each of the equations, the a priori probability that it
is satisfied for a given value of 2ℓ is 2−ℓ . With each of these equations we can again associate
a character: the largest value w that is a power of two for which the equation is satisfied. The
4-round character (i.e. leading to a0 , a1 and a2 all three in the kernel) of the vortex in this
context is now the minimum of the 3-round character (i.e. leading to both a0 and a1 in the
kernel) of the vortex and the characters of the three additional equations. The probability
that the 4-round character is larger than 2ℓ is approximately 2−4(ℓ+1) . It turns out that for
our choice of π and ρ, 8 of the 50 candidate vortices have 4-round character 2 and the others
have all 4-round character 1.
The conditions on the ( x, y) coordinates imply that only vortices are suited that have an
even number of active points in each sheet, each plane and each rising 1-slope. This limits
the number of suitable vortices of length 6 to 50, of length 8 to 300, of length 10 to 4180 and
of length 12 to 53750. To illustrate this, let us now study the number of activity pa erns in
the ( x, y) coordinates of a0 assuming there is only a single active bit in each lane. In total
there are 225 − 1 nonzero pa erns. If we impose the pa ern to be in the CP-kernel, the parity
of each sheet must be even, resulting in 5 independent linear equations. Hence there are
220 − 1 pa erns in the kernel. Additionally requiring a1 to be in the kernel imposes that the
number of points in each plane of a0 must be even. This adds 5 parity conditions. However,
one is redundant with the ones due to a0 as the total parity of the activity pa ern over the
state is even. Hence there are 216 − 1 such pa erns. Additionally requiring a2 to be in the
kernel imposes that the number of points in each rising 1-slope of a0 must be even. This
adds again 5 new parity condition, with one of them redundant and reduces the number of
possible pa erns to 212 − 1. Since π runs through all directions, adding more rounds results
in 28 − 1, and 24 − 1 and finally 0 pa erns. It follows that the range of possible activity
pa erns shrinks exponentially as the number of rounds grows.
This is the main reason for choosing a π that runs through all axes in a single cycle.
Consider a π that would map sheets to rising 1-slopes and rising 1-slopes back to sheets. For
such a π there would be 216 − 1 activity pa erns with a0 , a1 and a2 in the kernel. Moreover,
this number would not decrease for more rounds and periodic χ-zero trails of low weight
might appear.
When trying vortices with length above 6, the conditions on the z coordinates can be more
involved. If in a particular sheet of a2 the number of active points is 2, the condition is the
same as for the case described above: their z coordinates should match. However, if there
are 4, 6 or any even number of active points, there are several ways for them to be grouped in
different columns. In general a character can be computed per sheet and the character of the
complete structure is the minimum of all these characters. The character for a given sheet
can be computed in a recursive way. The probability that an active sheet has character 1 is
1/2. For larger characters, the probability decreases faster with growing number of active
bits in the character.
We have done tests for vortex lengths up to 14 and for constructions making use of two
vortices totaling to about 1 million valid a0 pa erns. The vast majority have character 1, less
than 13000 have character 2, 103 have character 4 and one has character 8. This last one is
based on vortex of length 8 and it results in a 4-round trail with weight 512 in K - f [1600].

28 / 69
2. The K - f permutations The K reference

2.4.7 Truncated trails and differentials

Truncated trails deal with the propagation of activity pa erns rather than differences [30]. A
partition of the state in sub-blocks is defined where the activity pa erns describe whether a
sub-block has no active bits (passive or 0) or has at least one active bit (active or 1). The struc-
ture of the state in K - f suggests several bundlings. In a first order, one may take rows,
columns, lanes, sheets, planes or slices as sub-blocks. We have gone through an exercise of
a empting this but got stuck very soon for each of the choices. The problem is that for every
choice, at least one of the step mappings completely tears apart the sub-blocks. We have also
considered hybrid state definitions, such as the combination of row activities with column
parities. However, in the cases that could be interesting, i.e., states with low weight (with
respect to the truncation considered), this soon lead to the full specification of the difference.
In [29] truncated cryptanalysis was applied to R G [4], where the truncation was
defined by a linear subspaces of the word vectors. In the a ack it made sense as part of the
R G state (the belt) is updated in a linear way. In K - f the round function is
uniformly non-linear and we do not believe that this approach can work.

2.4.8 Other group operations

We have considered differential and linear cryptanalysis while assuming the bitwise addi-
tion as the group operation. One may equivalently consider differential and linear properties
with respect to a wide range of other group operations that can be defined on the state. How-
ever, for any other choice than the bitwise addition, θ becomes a nonlinear function and for
most choices also ι, π and ρ become nonlinear. We do not expect this to lead to be er results.

2.4.9 Differential and linear cryptanalysis variants

There are many a acks that use elements from differential cryptanalysis and/or linear crypt-
analysis. Most are applied to block ciphers to extract the key. We have considered a number
of techniques:
• Higher-order differentials [30]: the algebraic order of the K - f round function is
only two and therefore constructing higher-order differentials seems like an interesting
route to structural distinguishers. We believed and still believe that due to the high
average diffusion it is very difficult to construct higher-order differentials of practical
significance for K - f . However, the low degree of K - f is indeed the cause
for the distinguishers covering the highest number of rounds of K - f to date. We
discuss these in Section 4.4.
• Impossible differentials [41]: we expect the K - f permutations to behave as ran-
dom permutations. If so, the cardinality of differentials has a Poisson distribution with
λ = 1/2 [25] and hence about 60 % of the differentials in K - f will have cardinal-
ity 0, and so are impossible. However, given a differential ( a, b), it is a priori hard to
predict whether it is impossible. Se ings in which one could exploit impossible differ-
entials are keyed modes, where part of the input is fixed and unknown. In this case one
would need truncated impossible differentials. If the number of rounds is sufficient to
avoid low-weight differential trails, we believe this can pose no problem.
• Differential-linear a acks [33]: in these a acks one concatenates a differential over a
number of rounds and a correlation over a number of subsequent rounds. We think that
for reduced-round versions of K - f differential-linear distinguishers are a candi-
date for the most powerful structural distinguisher. The required number of pairs

29 / 69
The K reference 2. The K - f permutations

is of the order DP−2 LP−2 with DP the differential probability of the distinguisher’s
differential and LP the square of the distinguisher’s correlation. If we assume the
differentials are dominated by a single low-weight differential trail, we have DP ≈
2−wr (Qd ) . Additionally, if we assume the correlation is dominated by a single low-
weight linear trail, we have DP ≈ 2−wc (Ql ) . This gives for the number of required
pairs: 22(wr (Qd )+wc (Ql )) . The number of required pairs to exploit a trail in a simple dif-
ferential or linear a ack is of the order 2wc (Q) . Hence, over a number of rounds, the
differential-linear distinguisher is more powerful than a simple differential or linear
distinguisher if wr ( Qd ) + wc ( Ql ) < wc ( Q)/2. Where Q is a trail over all rounds, Qd
a trail of the first n rounds and Ql a trail over the remaining rounds. As we expect in
the K - f variants with large width and a low number of rounds, the minimum
trail weight tends to grow exponentially, and the chaining of two half-length trails is
favored over a single full-length trail.

• (Amplified) Boomerang [39, 28] and rectangle a acks [11]: These a acks chain (sets of)
differentials over a small number of rounds to construct distinguishers over a larger
number of rounds. These are also likely candidates for good structural distinguishers,
for the same reason as differential-linear ones.

• Integral cryptanalysis (Square a acks) [22]: this type of cryptanalysis lends itself very
well to ciphers that treat the state in blocks. It was applied to bit-oriented ciphers in
[42]. Based on the findings of that paper we estimate that it will only work on reduced-
round versions of K - f with three to four rounds.
In this section we have limited ourselves to the construction of structural distinguishers.
We have not discussed how these distinguishers can be used to a ack the sponge function
making use of the permutation. In [8, Section “Some structural distinguishers”], we discuss
the applicability of structural distinguishers when a acking sponge functions.

2.5 Solving constrained-input constrained-output (CICO) problems

In [8, Section “The constrained-input constrained output problem”] we introduce CICO
problems for permutations and show that the resistance of a sponge function against many
a acks is determined by the hardness of solving these problems for their underlying permu-
tation.
There are several approaches to solving a CICO problem for K - f . The most straight-
forward way is to use the K - f specification to construct a set of algebraic equations in
a number of unknowns that represents the CICO problem and try to solve it. Thanks to
the simple algebraic structure of K - f , constructing the algebraic equations is straight-
forward. As a ma er of fact, this is supported by K T [9]. A single instance of
K - f results in (nr − 1)b intermediate variables and about as many equations. Each
equation has algebraic degree 2 and involves about 31 variables. Solving these sets of equa-
tions is however not an easy task. This is even the case for the toy version K - f [25] with
lane length w = 1. We refer to Section 4.2 for the results of some experiments for solving
CICO-solving algebraically.

2.6 Strength in keyed mode

In keyed modes we must consider a ack scenario’s such as explained in [8, Section “Keyed
modes”]. Here we see two main approaches to cryptanalysis. The first one is the exploita-

30 / 69
2. The K - f permutations The K reference

tion of structural distinguishers and the second one is an algebraic approach, similar to the
one presented in Section 2.5. A possible third approach is the intelligent combination of ex-
ploiting a structural distinguisher and algebraic techniques. In our opinion, the strength in
keyed modes depends on the absence of good structural distinguishers and the difficulty of
algebraically solving sets of equations.

2.7 Symmetry weaknesses

Symmetry in the state could lead to properties similar to the complementation property of
DES [35]. Symmetry between the rounds could lead to slide a acks. We believe that the
asymmetry introduced by ι is sufficient to remove all exploitable symmetry from K -f.
We refer to Section 4.1.2 for some experimentally obtained evidence of this.

31 / 69
The K reference 2. The K - f permutations

32 / 69
Chapter 3

Trail propagation in K -f

As explained in [8, Section “Some structural distinguishers”], the existence of differential or

linear trails in K - f with a weight below the width of K - f may result in a structural
distinguisher of K - f . In this chapter we report on our investigations related to lower
bounds for weights of such trails.
In Section 3.1 we define different types of weight and in Section 3.2 we discuss the relevant
properties of θ. In Section 3.3 we discuss the techniques used to come up with the lower
bounds on trail weights and report on the results obtained. Finally, in Section 3.4 we report
on experiments that allows us to get an idea what to expect for higher width values.

3.1 Relations between different kinds of weight

In this section, we recall or define the various kinds of weight we use in our treatment of trail
propagation. We speak of pa erns that may have the shape of a state, slice, plane, sheet, row,
column, lane or even a single bit. Some types of weight are only defined for certain shapes
and we will indicate it when this is the case.
We call a bit equal to 1 in a pa ern an active bit and a bit equal to zero a passive bit.

Definition 4. The Hamming weight of a pa ern a is the number of active bits in the pa ern and
is denoted by || a||.

For a generic definition of the restriction weight and the correlation weight we refer to [8,
Sections “Differential cryptanalysis” and “Linear cryptanalysis”]. These two weights relate
to the properties of χ, which are detailed in Section 2.3.1. Since χ operates on each row
independently, the restriction and correlation weights of a pa ern can be computed row per
row and the results are summed. The weights for all row pa erns are listed explicitly in
Tables 3.1 and 3.2.

Definition 5. The (propagation) weight of a pa ern a, denoted by w( a), is a generic term for either
the restriction weight or the correlation weight of a pa ern before χ. Since χ operates on rows, the
pa ern a must consist of full rows implying that the weight is only defined for states, slices, planes
and rows.

Note that the size of the linear affine varieties B( ai ) and A( ai ) (see Section 2.3.1) is deter-
mined by the propagation weight of ai :

|B( ai )| = |A( ai )| = 2w(ai ) .

33 / 69
The K reference 3. Trail propagation in K -f

Difference a wr ( a ) wrev
r ( a) || a|| ∥ a∥row
00000 0 0 0 0
10000 2 2 1 1
11000 3 2 2 1
10100 3 2 2 1
11100 4 2 3 1
11010 3 3 3 1
11110 4 3 4 1
11111 4 3 5 1

Table 3.1: Weights of all row differences (up to cyclic shi s)

Mask a wc ( a ) wrev
c ( a) || a|| ∥ a∥row
00000 0 0 0 0
10000 2 2 1 1
11000 2 2 2 1
10100 4 2 2 1
11100 4 2 3 1
11010 4 2 3 1
11110 4 2 4 1
11111 4 4 5 1

Table 3.2: Weights of all row masks (up to cyclic shi s)

Definition 6. For a pa ern b a er χ, we define the minimum reverse weight wrev (b) as the min-
imum weight over all compatible a. Namely,

wrev (b) = min w( a ).

a : b∈B( a)

This weight applies to states, slices, planes and rows.

Given an ℓ-round trail Q = ( a1 . . . aℓ ), it is easy to find the ℓ + 1-round trail Q′ =

( a0′ , a1′ . . . a′ℓ ) with minimum
weight such that ai = ai′ for 1 ≤ i ≤ ℓ. In this case, w( Q′ ) =
w( Q) + wrev (λ−1 ( a1 )).
It may also be useful to express the number of active rows in a given pa ern.

Definition 7. For a pa ern a before (or a er) χ, the number of active rows, denoted by ∥ a∥row is
simply the number of rows whose value is non-zero. This weight applies to states, slices, planes and
rows.

The different kinds of weight for all row pa erns are given in Tables 3.1 and 3.2. We now
give some relations between the various kinds of weights. The following bounds relate the
Hamming weight to the weight:
⌊ ⌋
|| a||
ŵ(|| a||) , || a|| − + [1 if || a|| = 1 (mod 5)] ≤ w( a) ≤ 2|| a||,
5
⌈ ⌉ ⌊ ⌋
w( a ) 5w( a)
≤ || a|| ≤ .
2 4

34 / 69
3. Trail propagation in K -f The K reference

The following bounds relate the number of active rows to the weight:

2∥ a∥row ≤ w( a) ≤ 4∥ a∥row ,
⌈ ⌉ ⌊ ⌋
w( a ) w( a )
≤ ∥ a∥row ≤ .
4 2
Given the Hamming weight, the minimum reverse restriction weight can be lower bounded
as follows:


 0 if || a|| = 0 (mod 5),



⌊
|| a|| 1 if || a|| = 1 (mod 5),
⌋ 
wr ( a) ≥ ŵr (|| a||) , 3
rev rev
+ 2 if || a|| = 2 (mod 5),
5 


 2 if || a|| = 3 (mod 5),



3 if || a|| = 4 (mod 5).

Given the Hamming weight, the minimum reverse correlation weight can be lower bounded
as follows: ⌈ ⌉
|| a||
wc ( a) ≥ ŵc (|| a||) , 2
rev rev
.
4
Other relations on the minimum reverse weight follow:

wrev (b) ≤ 2||b||,

⌈ rev ⌉ ⌊ ⌋
w (b) 5wrev (b)
≤ ||b|| ≤ ,
2 4
2∥b∥row ≤ wrev (b) ≤ 4∥(∥row b),
⌈ rev ⌉ ⌊ rev ⌋
w (b) w (b)
≤ ∥b∥row ≤ .
4 2
And finally,
wrev (b) ≤ w(b).
Definition 8. A weight function f ( a) is said to be monotonous if se ing a passive bit of a pa ern
a to active does not decrease f ( a). More formally, let the partial ordering a ≤ a′ be defined as
a[ x ][y][z] = 1 ⇒ a′ [ x ][y][z] = 1. Then, f is a monotonous weight if ∀ a, a′ : a ≤ a′ , we have
f ( a ) ≤ f ( a ′ ).
All weights defined in this section are monotonous. This follows directly from Tables 3.1
and 3.2.

3.2 Propagation properties related to the linear step θ

Definition 9. The column parity (or parity for short) P( a) of a pa ern a is a pa ern P( a) = p[ x ][z]
defined as the parity of the columns of a, namely p[ x ][z] = ∑y a[ x ][y][z]. The parity is defined for
states, slices, sheets and columns. The parity of a state has the shape of a plane, the parity of a slice
has the shape of a row, the parity of a sheet has the shape of a lane and the parity of a column is a bit.
Equivalently, the parity is the result of applying the operator ȳ (see Section 2.3.2), and the
value C in Algorithm 3 is the parity of a state. A column is even (resp. odd) if its parity is 0
(resp. 1). When the parity of a pa ern is zero (i.e., all its columns are even), we say it is in
the CP-kernel, as in Section 2.4.3. Note that the column parity defines a partition on the set
of possible states.

35 / 69
The K reference 3. Trail propagation in K -f

Definition 10. The θ-effect of a state a before θ is a pa ern E( a)[ x ][z] defined as the result of applying
the operator ȳ( x + x4 z) to the state, or equivalently by applying the operator ( x + x4 z) to its parity,
i.e., E( a)[ x ][z] = P( a)[ x − 1][z] + P( a)[ x + 1][z − 1].

In difference propagation, the θ-effect is also the value of D in Algorithm 3. In mask

propagation, the θ-effect is defined by θ T rather than θ itself and the expression becomes
E( a)[ x ][z] = P( a)[ x + 1][z] + P( a)[ x − 1][z + 1] (see Section 2.3.2.2). Note that the θ-effect al-
ways has an even Hamming weight. A column of coordinates ( x, z) is affected iff E( a)[ x ][z] =
1; otherwise, it is unaffected.
The θ-effect entirely determines the effect of applying θ to a state a. For a fixed θ-effect
e[ x ][z], θ just adds a constant pa ern e[ x ][y][z] of bits to the state, constant in each column,
namely e[ x ][y][z] = e[ x ][z] for all y. Since the θ-effect has even Hamming weight, it means
that the number of affected columns is even.

Definition 11. The θ-gap is defined as the Hamming weight of the θ-effect divided by two.

Hence, if the θ-gap of a state at the input of θ is g, the number of affected columns is 2g
and applying θ to it results in 10g bits being flipped.
When a state is in the CP-kernel, the θ-gap is zero. However, the θ-gap is also zero when
the parity is all-one, i.e., when all columns have odd parity before θ.
We have defined the θ-gap using the θ-effect, but it can also be defined using the parity
itself. For this, we need to represent the parity p[ x ][z] differently. We map the ( x, z) coor-
dinates to a single coordinate t as specified in Section 2.3.2.1 (i.e., t goes to ( x, z) = (−2t, t))
and denote the result by p[t].
In this representation, a run is defined as a sequence R of consecutive t-coordinates, R =
{s, s + 1, . . . , s + n − 1}, such that p[s − 1] = 0, p[t] = 1∀t ∈ R and p[s + n] = 0. The
following lemma links the number of runs to the θ-gap.

Lemma 1. The parity p has θ-gap g iff p[t] has g distinct runs.

3.3 Exhaustive trail search

In Section 3.3.2 we will show an efficient method for generating all two-round trails up to a
given propagation weight T2 . We will then show in section Section 3.3.3 how these trails can
be extended to more rounds. In general, we want to generate all ℓ-round trails up to some
propagation weight Tℓ . We start with Section 3.3.1 deriving the minimum value of T2 given
ℓ and Tℓ . Finally, we report on our bounds obtained in Section 3.3.4

3.3.1 Upper bound for the weight of two-round trails to scan

The idea is have bounds on ℓ-round trails by starting from all two-round trails of weight up
to T2 and extending them both forwards and backwards. More precisely, each two-round
trail is extended n rounds backwards and ℓ − 2 − n rounds forwards, for each value of n ∈
{0, . . . , ℓ − 2}. This way, we cover all trails that have a two-round subtrail with weight up to
T2 .

Lemma 2. To list all ℓ-round trails of weight not higher than Tℓ exhaustively,⌊ it⌋is necessary to
start from all 2-round trails with weight up to (and including) T2 , with T2 = 2Tℓ ℓ if ℓ is even, or
⌊ ⌋
2( Tℓ −2)
T2 = ℓ−1 if ℓ is odd.

36 / 69
3. Trail propagation in K -f The K reference

Proof. For a sequence of weights W = w1 , w2 , . . . , wℓ , let δ(W ) = mini=1...ℓ−1 (wi + wi+1 ). We

want to make sure that δ(W ) ≤ T2 for all W such that ∑i wi ≤ Tℓ .
If ℓ is even, Tℓ ≥ ∑iℓ=1 wi = ∑ℓj=
/2
1 ( w2j−1 + w2j ) ≥ 2 δ (W ) and thus δ (W ) ≤ ℓ . Se ing
ℓ 2Tℓ
⌊ ⌋
T2 = 2Tℓ ℓ satisfies the condition.
If ℓ is odd, we can always assume that wℓ ≥ 2 (as in any non-trivial trail) and thus we
can consider the same problem with sequences of ℓ − 1 weights and Tℓ−1 = Tℓ − 2. ⊓
⊔
So we have all the necessary ingredients to make exhaustive search of trails up to a given
weight, within the limits of a reasonable computation time, and we can use that to find trails
with minimum weight.

3.3.2 Constructing two-round trails

In this section we describe an efficient method for constructing all two-round trails ( a0 , a1 )
with w( a0 ) + w( a1 ) ≤ T2 for a given value T2 . For any such trail, we know that there exists a
trail ( a0′ , a1 ) with w( a0′ ) = wrev (λ−1 ( a1 )) ≤ w( a0 ). The quantity wrev (λ−1 ( a)) + w( a) ≤ T2
imposes a lower bound on the weight of a 2-round trail trail that has a as its state before χ of
the second round and we give it the following name.
Definition 12. The propagation branch number of a state a before χ is denoted by Bp ( a) and
given by:
Bp ( a) = wrev (λ−1 ( a)) + w( a) .
Hence, rather than explicitly constructing all two-round trails, we can generate the set of
states a such that Bp ( a) ≤ T2 . We call this set α( T2 ). It contains the second members a1 of
all two-round trails Q with w( Q) ≤ T2 . Each state in this set α( T2 ) then serves as a starting
point for trail extension. For the resulting trails, we know that the subtrail ( ai−1 , ai ) with
ai = a has a weight of at least Bp ( a).
In generating α( T2 ) we use a strategy that exploits the monotonicity of the propagation
weight and the properties of θ in terms of the Hamming weight of its input/output. In Sec-
tion 3.1 we listed equations providing lower bounds on the weight and reverse minimum
weight of a state as a function of its Hamming weight. Hence, || a|| and ||λ−1 ( a)|| gives a
lower bound on Bp ( a).
Definition 13. The Hamming branch number of a state a before χ is denoted by Bh ( a) and given
by:
Bh ( a) = ||λ−1 ( a)|| + || a|| .
We can now just generate all states up to some given propagation branch number by
generating all states up to a sufficiently high Hamming branch number value.
Now let a′ be the state before θ corresponding with a. In difference propagation, we
−1
have a′ = λ−1 ( a) and in mask propagation this is a′ = θ T ( a). It follows that in difference
propagation, we have || a′ || = ||λ−1 ( a)|| and ||θ ( a′ )|| = || a|| and hence Bh ( a) = || a′ || +
||θ ( a′ )||. In mask propagation we can similarly derive Bh ( a) = || a′ || + ||θ T ( a′ )||. Hence we
can instead move our reference from a to a′ and just compute the value of a from a′ . With a
slight abuse of notation we will write Bh ( a′ ) and Bp ( a′ ) to denote Bh ( a) and Bp ( a).
Consider now the set of states a′ with a given parity p, i.e., P( a′ ) = p. As the θ-effect e is
fully determined by the parity, all these states have the same parity effect. It follows that θ
is reduced to the addition of a constant value, facilitating the computation of Bh ( a′ ) and the
deduction of lower bounds on Bp ( a′ ).
Similar to the Hamming branch number of a state, we can define the Hamming branch
number of a parity.

37 / 69
The K reference 3. Trail propagation in K -f

Definition 14. The Hamming branch number of a parity p before θ is defined as the minimum Ham-
ming branch number over all states with the given parity:

Bh ( p) = min Bh ( a′ ) .
a′ :P( a′ )= p

In a state a′ we can use its parity p to partition its columns a′ [ x ][z] in four kinds: odd
(P( a′ )[ x ][z] = 1) and even (P( a′ )[ x ][z] = 0), combined with affected (E( a′ )[ x ][z] = 1) and
unaffected (E( a′ )[ x ][z] = 0). We can use this to easily compute Bh ( p) for any parity p.

• An unaffected odd column has at least one active bit before θ and is preserved a er it.
Hence, it contributes at least 2 to Bh ( a′ ). As this minimal case can be constructed, the
contribution to Bh ( p) is strictly equal to 2.

• An affected (odd or even) column having n active bits before θ has 5 − n bits a erwards,
hence contributes exactly 5 to Bh ( a′ ) and to Bh ( p).

• An unaffected even column can have zero active bits, hence does not contribute to
Bh ( p).

Hence, it turns out that

Bh ( p) = 5|| E( p)|| + 2|| p · (0̄ + E( p))|| = 10g + 2uo ,

with g the θ-gap of p, · the componentwise product, 0̄ the all-1 state and uo the number of
unaffected odd columns.
We can now generate all states a′ up to some given propagation branch number and with
given parity p in two phases.

• In a first phase we generate all states a′ with Bh ( a′ ) = Bh ( p). We call those states
branch-parity-minimal.

• In a second phase, we can generate states a′ that are not branch-parity-minimal by

taking branch-parity-minimal states and adding pairs of active bits in columns such
that the parity is unchanged.

38 / 69
3. Trail propagation in K -f The K reference

The generation of branch-parity-minimal states is done as follows:

• For each unaffected odd column, put a single active bit. There are 5 possibilities: one
for each positions y.

• For each affected even column, put an even number of active bits. There are 24 possi-
bilities.

• For each affected odd column, put an odd number of active bits. There are in 24 possi-
bilities.

The number of branch-parity-minimal states a′ with given parity p is thus 28g 5| p(1+e)| .
From the monotonicity of the weights it follows that in the set of states a′ with given
parity p, the subset of branch-parity-minimal states contain the states that minimize Bp ( a′ ).
Adding a pair of active bits in a single column of a′ leaves its parity intact and thanks to the
monotonicity cannot decrease Bp ( a′ ). From this, we devise the following strategy to generate
all states a′ with given parity p with Bp ( a′ ) ≤ T2 .
For each branch-parity-minimal state a′ with Bp ( a′ ) ≤ T2 do the following:

• Output a′ .

• Iteratively construct states a′ by adding pairs of active bits in each column, as long as
Bp ( a′ ) ≤ T2 . To avoid duplicates the active bits shall have y coordinates with larger
values than any active bits in the columns.

To generate all states a′ for which Bp ( a′ ) ≤ T2 we must do this for all parities with a small
enough θ-gap. Actually, we can compute a lower bound on Bp ( a′ ) given only P( a′ ). We then
have to consider only those parities for which this bound is lower than or equal to T2 . We
compute it in the following way. First, we consider the Hamming branch number Bh ( p) and
assume that |λ−1 ( a)| = Bh ( p) − n and | a| = n for some value n. Then, we use the bounds
found in Section 3.1 and minimize over n. Note that we have checked that the minimum is
always at n = 1, hence:

Bp ( a′ ) ≥ min ŵrev ( Bh ( p) − n) + ŵ(n)

n∈{1...Bh ( p)−1}

= ŵrev ( Bh ( p) − 1) + 2
= ŵrev (10g + 2uo − 1) + 2.

Hence the θ-gap of the parity p imposes a lower bound to the propagation branch number
of a state.
Then, it is possible to determine the maximum θ-gap gmax above which the lower bound
is above T2 . If we further relate gmax to the number of runs in the parity, as in Lemma 1, we
can generate all possible parities we need to consider by generating those with up to gmax
runs.
Notice that both χ and λ are invariant by translation along z. It is thus necessary to keep
only a single member of the states (or parities) in α( T2 ) that are equal modulo the translations
along z.

3.3.3 Extending trails

We can now use the elements in α( T2 ) to recursively generate longer and longer trails up
to some given length and some given weight. We can extend the trails in two directions:
forward and backward. Given a trail Q, extending the trail forward (resp. backward) means

39 / 69
The K reference 3. Trail propagation in K -f

constructing trails Q′ of which Q is a prefix (resp. suffix). It can also be in both directions,
i.e., adding a number of steps as a prefix and another number of steps as a suffix of Q.
In the forward direction, the general idea is the following. Given the last state aℓ−1 of the
trail Q, we characterize the affine space A( aℓ−1 ) as an offset and a basis, i.e.,

A( aℓ ) = s + ⟨t1 , t2 , . . . , tw(aℓ ) ⟩ .

We can then loop through the affine space, produce the state values aℓ and check the weight
w( aℓ ). If the weight is low enough for the extended trail to be interesting in the search, we
can append aℓ to Q and recursively continue the search from there if necessary.
In the backward direction, we cannot use an affine space representation of a−1 as a func-
tion of a0 . As the weight is determined by the to-be-found state a−1 , we can list, for each
active row of b0 , the possible input rows and their corresponding weight in increasing or-
der. The weight of a−1 is the sum of the weights of each individual row, and we can take
advantage of this to choose input rows such that the weight stays below or equal to a given
threshold. We set each active row to the input row with lowest weight, the total weight
w( a−1 ) being equal to wrev (b0 ) for the output state b0 . Then, we can generate all other input
states a−1 by looping through the input rows, locally knowing up to which weight we can
go.

3.3.4 Linear and differential trail bounds for w ≤ 8

We have investigated the different instances of K - f [b] starting from the smallest widths
b = 25w, resulting in the lower bounds for trail weights listed in Table 3.3. Thanks to
the Matryoshka structure, a lower bound on trails for K - f [b] implies a lower bound
on symmetric trails for all larger widths. More specifically, a differential or linear trail for
K - f [25w] with weight W corresponds with a w-symmetric trail for K - f [25w′ ] with
′ w′
weight W = W w . For instance, in Table 3.3 the column DC, w = 8 expresses a lower bound
of 46 on the weight of 4-round trails in K - f [200]. This also expresses a lower bound for
4-round symmetric trails in K - f versions with larger width: 92 for 2-symmetric trails
in K - f [400], 184 for 4-symmetric trails in K - f [800] and 268 for 8-symmetric trails
in K - f [1600].
For w = 1, five rounds are sufficient to have no trails with weight below 25, the width of
the permutation. For w = 2, six rounds are sufficient to have no differential trails with weight
below the width. It can be observed that as the number of rounds grows, the difference
between the bounds for width 25 and those for width 50 grows. We expect this effect to be
similar for larger widths.
For w = 4, the search was complete up to weight 36 and 38 for 4 rounds, for differential
and linear trails respectively:

• For 4 rounds, the differential trail with minimum weight has weight 30. For the small
number of trails found up to weight 36, we checked that these trails cannot be chained
together. Hence, this guarantees that a 8-round differential trail has at least weight 36 +
37 = 73. For 5 and 6 rounds, the best trails we have found so far have weight 54 and 85,
respectively (but these do not provide bounds). For the 16 rounds of K - f [100],
we can guarantee that there are no differential trails of weight below 2 × 73 = 146.

• For 4 rounds, the linear trail with minimum weight has weight 38. For 5 and 6 rounds,
the best trails we have found so far have weight 66 and 94, respectively (but these do
not provide bounds). For the 16 rounds of K - f [100], we can guarantee that there
are no linear trails of weight below 4 × 38 = 152.

40 / 69
3. Trail propagation in K -f The K reference

Number DC LC
of rounds w=1 w=2 w=4 w=8 w=1 w=2 w=4 w=8
2 8 8 8 8 8 8 8 8
3 16 18 19 20 16 16 20 20
4 23 29 30 46 24 30 38 46
5 30 42 ≤ 54 30 40 ≤ 66
6 37 54 ≤ 85 38 52 ≤ 94

Table 3.3: Minimum weight of w-symmetric trails

For w = 8, the search was complete up to weight 49 and 48 for 4 rounds, for differential
and linear trails respectively:

• For 4 rounds, the differential trail with minimum weight has weight 46. For the small
number of trails found up to weight 49, we checked that these trails cannot be chained
together. Hence, this guarantees that a 8-round differential trail has at least weight
49 + 50 = 99. For the 18 rounds of K - f [200], we can guarantee that there are no
differential trails of weight below 2 × 99 + 8 = 206.

• For 4 rounds, the linear trail with minimum weight has weight 46. For the small num-
ber of trails found up to weight 48, we checked that these trails cannot be chained
together. Hence, this guarantees that a 8-round differential trail has at least weight
48 + 50 = 98. For the 18 rounds of K - f [200], we can guarantee that there are no
linear trails of weight below 8 × 98 + 8 = 204.

3.4 Tame trails

In this section we report on our investigations related to the search for 3-round and 4-round
differential trails with low weight and for high width. In this context, we consider differential
trails for which the intermediate states bi are in the CP-kernel and call such trails tame. Linear
trails are expected to behave in qualitatively the same way.
This is a generalization of the kernel vortices introduced in Section 2.4.5 where only pat-
terns are considered for which both χ and θ behave as the identity. In kernel vortices only
intermediate pa erns at the input of χ where considered with a single active bit per row. In
the trails considered in this section, this restriction is no longer present.

3.4.1 Construction of tame trails

Let us start with 3-round differential trails. The weight of such a trail is defined by three
states: ( a0 , a1 , a2 ). This trail is tame if b0 = λ−1 ( a1 ) and b1 = λ−1 ( a2 ) are both in the CP-
kernel. We have wri en a program that generates all values of a1 of a given Hamming weight,
such that b0 is in the CP-kernel and there is at least one three-round differential trail ( a0 , a1 , a2 )
that is tame, i.e., with b1 in the CP-kernel. The la er condition imposes that a1 must be such
that the intersection of A( a1 ) and the CP-kernel is not empty. As χ operates on rows and the
CP-kernel is determined by individual columns, this can be treated slice by slice. In other
words, every slice of a1 must be such that its linear affine variety of possible output pa erns
contains pa erns in the CP-kernel.

41 / 69
The K reference 3. Trail propagation in K -f

We call such a slice tame and a state with only tame slices also tame. For slice pa erns
with few active bits it can be easily verified whether it is tame. A slice with no active bits is
tame, a slice with a single active bit can never be tame and a slice with two active bits is tame
iff the active bits are in the same column. As the number of active bits grows, the proportion
of slice pa erns that are not tame decreases exponentially. As there are only 225 different
slice pa erns, the tameness check can be precomputed and implemented by a simple table-
lookup. Generating all states a1 of a given (small) Hamming weight, that are tame and for
which b0 is in the CP-kernel can be done efficiently.
We can now construct all valid states a1 as the combination of a number of kernel chains.
A kernel chain is set of active bits determined by a sequence of bit positions ci in a1 . A kernel
chain forms a set of active bits with the following properties:

• When moved to position b0 the kernel chain it is in the CP-kernel.

• In position a1 every slice contains exactly two bits of the kernel chain and is tame, except
the slice containing the initial bit c0 and the slice containing the final bit c2n+1 , that each
just contain a single slice and are not tame.

Actually, a kernel chain is a generalization of a kernel vortex. It follows that in b0 , the bits c2i
and c2i+1 are in the same column and in a1 , the bits c2i+1 and c2i+2 are in the same column.
This implies that the total number of kernel chains of a given length 2n starting from a given
position is only 42n−1 . Clearly, any combination of kernel chains is in the CP-kernel in b0 .
Likewise, all slices in a1 that contain only two kernel chain bits (excluding the initial and
final bits) are tame. Now we must arrange the initial and final kernel chain bits such that the
slides in a1 that contain them are tame. The first possibility is that the bits c0 and c2n−1 are in
the same column in a1 : this kernel chain forms a kernel vortex. The second possibility is to
group the initial and final bits of kernel chains in tame knots. We call a slice in a1 with more
than 2 active bits a knot. We construct states a1 by combining kernel chains such that their
initial and final bits are grouped in a set of knots. If all knots are tame, the state is tame. For
a given Hamming weight x, valid states may exist with 0 up to ⌊ x/3⌋ knots.
In our program we first fix the number of knots and their slice positions and then effi-
ciently search for all valid states. Table 3.4 lists the number of valid states a1 (modulo transla-
tion over the z-axis) for all K - f widths and up to a Hamming weight of 14. The question
marks mark the limitations of our search algorithm: we have not yet been able to compute
those values due to time constraints. It can be seen that for a given Hamming weight, overall
the number of valid states decreases with increasing width. For a given width, the number
of valid states increases with increasing Hamming weight.

3.4.2 Bounds for three-round tame trails

Starting from the valid pa erns a1 we have for each one searched for the 3-round tame trail
with the smallest (restriction) weight. This was done in the following way. Given a1 we
compute the minimum weight of a0 by taking wrev (λ−1 ( a1 )). In the forward direction, we
iterate over all states a2 for which b1 is in the CP-kernel. The results are given in Table 3.5.
Cases containing a dash indicate that there are no tame 3-round trails for the given width
and Hamming weight. Entries of the form “≤ n” indicate that not all the pa erns were
investigated. It can be seen that increasing the width results only in a limited growth of
the (restriction) weight. Clearly, the limited diffusion due to the existence of the CP-kernel
results in low-weight trails.

42 / 69
3. Trail propagation in K -f The K reference

Width Hamming weight

4 6 8 10 12 14
25 825 12100 95600 465690 1456725 ?
50 150 13835 905135 22392676 ? ?
100 48 2712 137078 6953033 ? ?
200 10 481 24037 1143550 56824109 ?
400 4 83 4006 164806 7290847 ?
800 0 28 918 30771 1154855 44788752
1600 0 10 304 8231 259567 8399589

Table 3.4: The number of valid difference pa erns a1 per K - f width and Hamming
weight

Width Hamming weight

4 6 8 10 12 14
25 18 20 22 ≤ 22 ? ?
50 18 22 25 28 ? ?
100 19 24 29 ≤ 36 ? ?
200 20 29 33 38 ? ?
400 24 30 35 40 47 ?
800 - 35 41 47 53 58
1600 - 35 41 48 56 62

Table 3.5: Minimum differential trail weight values of tame 3-round trails per K -f
width and Hamming weight of a1 .

43 / 69
The K reference 3. Trail propagation in K -f

Width Hamming weight

4 6 8 10 12 14
25 25 28 29 ≤ 31 ? ?
50 30 31 34 38 ? ?
100 30 36 41 ≤ 48 ? ?
200 - 56 56 61 ? ?
400 - - - - 90 ?
800 - - - - - -
1600 - - - - - -

Table 3.6: Minimum differential trail weight values of tame 4-round trails per K -f
width and Hamming weight of pa ern a er first χ

3.4.3 Bounds for four-round tame trails

We have conducted the a similar search for the 4-round tame trails with the smallest (re-
striction) weight. This was done in the same way as for 3-round trails with this difference:
rather than stopping at a2 , we iterate over all states a3 ∈ A( a2 ) in the CP-kernel and compute
its restriction weight. To speed up the search we apply pruning if the restriction weight of
( a0 , a1 , a2 ) is such that a3 cannot result in a trail with a lower restriction weight than one found
for the given category. The results are given in Table 3.6. While for small widths the increase
in width does not substantially increase the restriction weights, for higher widths it can be
observed that the minimum Hamming weight of a1 for which there exist tame 4-round dif-
ferential trails increases dramatically. For width 400 we have to go up to a Hamming weight
of 12 to find a tame 4-round trail and for 800 and 1600 no tame 4-round trails exist for a1
pa erns with Hamming weight below 16. As for the minimum restriction weights of the
4-round tame trails found, the increase in 30 to 56 from width 100 to 200 and to 90 for width
400 suggests that the CP-kernel plays a much smaller role than for 3-round trails.

44 / 69
Chapter 4

Analysis of K -f

In this chapter we report on analysis and experiments performed on reduced-round versions

of K - f , either by the designers or third parties. In Section 4.1 we describe our experi-
ments based on the algebraic normal form representations. In Section 4.2 we report on the
outcome of a empts to solve CICO problem instances. In Section 4.3 we describe statisti-
cal properties of reduced-round versions of K - f [25]. Finally, in Section 4.4 we discuss
distinguishers that exploit the low algebraic degree of the round function and its inverse.

4.1 Algebraic normal form

In this section, we explain how the algebraic normal form can be used to evaluate the pseudo-
randomness of K - f in different aspects.

4.1.1 Statistical tests

There are several ways to describe K - f algebraically. One could compute the algebraic
normal form (ANF) [8, Section “Algebraic expressions”] with elements in GF(2), GF(25 ),
GF(225 ) or GF(2w ), but given the bit-oriented structure and matching θ, ρ, π, ι and χ as
operations in GF(2), the ANF in GF(2) seems like a natural way to represent the K -f
permutation. For instance, one could take the rows as variables in GF(25 ). This way, the
χ operation applies independently per variable. However, the other operations will have a
complex expression.
We performed several statistical tests based on the ANF of K - f [b, nr = n], from b =
25 to b = 1600 and their inverses in GF(2). The number of rounds n is also varied from 1
to its nominal value, although in practice we can limit ourselves to a reasonable number of
rounds a er which no significant statistical deviation can be found.
In general, the test consists in varying 25 bits of input and counting the number of mono-
mials of degree d of all b output bits. The statistical test is performed per degree indepen-
dently. The number of monomials of degree d should be present in a ratio of about one half.
The test fails when the observed number of monomials is more than two standard deviations
away from the theoretical average. We look for the highest degree that passes the test.
The different tests determine which input bits are varied and/or which variant of the
K - f permutation is used.

• Bits in slice (BIS) In this test, the 25 bits of the slice z = 0 are varied. For b > 25, the
remaining b − 25 bits are set to zero.

45 / 69
The K reference 4. Analysis of K -f

Rounds Maximum degree Monomials exist

to pass test up to degree
f f −1 f f −1
1 (none) (none) 2 3
2 (none) 3 3 9
3 (none) 10 5 17
4 5 25 9 25
5 16 25 17 25
6-24 25 25 25 25

Table 4.1: The BIS ANF statistical test on K - f [1600] and its inverse

• Bits in lane (BIL) In this test, the first 25 bits of lane x = y = 0 are varied. This test
applies only to b ≥ 800. The remaining b − 25 bits are set to zero.

• Bits in two lanes, kernel In this test, the first 25 bits of lane x = y = 0 and of lane
( x, y) = (0, 1) are varied simultaneously. The remaining b − 50 bits are set to zero. The
idea behind this test is that θ will behave like the identity in the first round since the
input state is always in the column-parity kernel (see 2.4.3).

• BIS, symmetry in lanes In this test, the 25 bits of the slice z = 0 are varied. The b − 25
other bits are set as a[ x ][y][z] = a[ x ][y][0] so that all lanes contain either all zeroes or all
ones. The output bits a[ x ][y][0] are xored into a[ x ][y][z] for all z > 0. Only the b − 25
output bits with z > 0 are considered for the test—this test applies only when b > 25.
The purpose of this test is discussed in Section 4.1.2.

• BIS, slide This test is like the first BIS test, except that the function tested is differ-
ent. Instead of testing it against K - f [b, nr = n] itself, we test it against the func-
tion slide[b, n] defined in Equation (4.1). The purpose of this test is discussed in Sec-
tion 4.1.3.

The results for K - f [1600] are summarized in Tables 4.1 and 4.2. All results of the
tests can be found in the file ANF-Keccak-f.ods. It is interesting to observe the fast increase of
degree of the monomials that are densely present in the algebraic description of the K -f
permutations. Note that the round function has only degree two and thus no monomial of
degree higher than 2i can appear a er a er i rounds. The degree of the inverse of the round
function is three and thus no monomial of degree higher than 3i can appear a er a er i
inverse rounds.
Taking the worst case among all these tests, the K - f permutations and their inverses
pass the test for the maximum degree tested here (i.e., max(25, b − 1)) a er 6 to 8 rounds,
depending on the width.

4.1.2 Symmetric trails

The design of K - f has a high level of symmetry. Due to this, the weight of symmetric
trails may no longer be relevant for the security. (See Section 2.4.2 for more details.) We
investigate how an a acker may be able to exploit the symmetry in his advantage.
The weight and existence of trails (both differential and linear) is independent of ι. The
fact that all other step mappings of the round function are translation-invariant in the direc-
tion of the z axis, makes that a trail Q implies w − 1 other trails: those obtained by translating

46 / 69
4. Analysis of K -f The K reference

Rounds Maximum degree Monomials exist

to pass test up to degree
f f −1 f f −1
1 (none) (none) 2 1
2 (none) 3 4 3
3 (none) 8 8 9
4 8 25 15 25
5 25 25 25 25
6 24 25 25 25
7-24 25 25 25 25

Table 4.2: The BIL ANF statistical test on K - f [1600] and its inverse

the pa erns of Q over any non-zero offset in the z direction. If all pa erns in a trail have a
z-period below or equal to d, this implies only d − 1 other trails.
Moreover, a trail for a given width b implies a trail for all larger widths b′ . The pa erns
are just defined by their z-reduced representations and the weight must be multiplied by
b′ /b. Note that this is not true for the cardinality of differential trails and the sign of the
correlation contribution of linear trails, as these do depend on the round constants.
To find a pair in a differential trail of weight W requires the a acker to fulfill W conditions
on the absolute values when following that trail. In the case of a b′ -symmetric case, the
conditions are repeated b′ /b times on translated sets of bits. The question is to determine
whether the symmetry induced by this duplication can be exploited by the a acker, even
with the asymmetry introduced by ι.
In the absence of ι, K - f [b] presented with a symmetric input behaves as b parallel
identical instances of K - f [25]. In such a modified permutation, a symmetric trail with
weight eb would only impose e conditions (on the symmetric absolute values) rather than eb.
To determine how much asymmetry ι introduces on the absolute values, we express the
K - f [b] permutation on a different set of variables and compute the ANF on it. The
change of variables is defined as:

a′ [ x ][y][0] = a[ x ][y][0],
a′ [ x ][y][z] = a[ x ][y][z] ⊕ a[ x ][y][0], z > 0.

In the absence of asymmetry (i.e., without ι), a′ [ x ][y][z], z > 0 remains zero if the input
of the permutation is 1-symmetric, i.e., if a[ x ][y][z] depends only on x and y.
The a acker can try to introduce a symmetric difference and to keep it symmetric through
the rounds. In these new variables, it is equivalent to keeping a′ [ x ][y][z] = 0 for z > 0. By
analyzing the ANF in these new variables, it gives the degree of the equations to solve to
keep the symmetry in the state at a given round by adjusting the input.
The results of these tests can be found in the file ANF-Keccak-f.ods. The maximum degree
to pass test is indicated in the BIS, sym lane columns. To impose dense non-linear equations
to the a acker, K - f [50] to K - f [400] need at least 3 rounds, while K - f [800]
and K - f [1600] need at least 4 rounds. A dense number of monomials with maximum
degree tested here (i.e., max(25, b − 1)) is reached a er 6 rounds for all widths. Similar con-
clusions apply to the inverse of K - f : The maximum degree tested is reached a er 6
rounds for the inverse of K - f [50], 5 rounds for the inverse of K - f [100] and of
K - f [200] and 4 rounds for the other inverses. According to this test, the a acker should
have a very difficult time to keep the differences symmetric a er such a number of rounds.

47 / 69
The K reference 4. Analysis of K -f

4.1.3 Slide a acks

Slide a acks [12, 27] are a acks that exploit symmetry in a primitive that consists of the
iteration of a number of identical rounds. We investigate how much asymmetry must be
applied for these a acks to be non-effective.
In the absence of ι, all rounds are identical. In such a case, this allows for distinguishing
properties of K - f : the distribution of the cycle lengths will be significantly different
from that of a typical randomly-chosen permutation.
To evaluate the amount of inter-round asymmetry brought by ι, we performed a specific
test based on ANF. We compute the ANF of the function slide[b, n], which is obtained by
XORing together the output of the first n rounds of K - f [b] and the output of n rounds
starting from the second one. Alternatively, it is defined as

slide[b, n] = (roundn−1 ◦ · · · ◦ round0 ) ⊕ (roundn ◦ · · · ◦ round1 ), (4.1)

where roundi is the round permutation number i. In the absence of ι, the slide[b, n] function
would constantly return zeroes. With ι, it returns the difference between two sets of n rounds
slid by one round.
The results of these tests can be found in the file ANF-Keccak-f.ods. The maximum degree
to pass test is indicated in the BIS, slide columns. The degree increases more slowly than for
other tests such as bits in slice. However, the maximum degree tested here (i.e., max(25, b −
1)) is reached a er 8 rounds for K - f [25] and a er 6 rounds for all other widths. For the
inverse of K - f [25], the maximum degree tested is reached a er 7 rounds and, for the
other inverses, a er 4 rounds.

4.2 Solving CICO problems algebraically

4.2.1 The goal
As explained in [8, Section “Conducting primary a acks using structural distinguishers”]
the security of a sponge function relies critically on the infeasibility of solving non-trivial
CICO problems for its underlying permutation. The resistance of K - f against solving
non-trivial CICO problems is hence critical for the security of K . We have developed
so ware to generate the sets of equations in a form that they can be fed to mathematics
so ware such as MAGMA or SAGE, and subsequently using this so ware to solve instances
of the CICO problem for different sets of parameters.

4.2.2 The supporting so ware

K T [9] supports the generation of round equations of all supported width values,
compatible with SAGE or other computer algebra tools. It can generate the equations for all
operations within a round, θ, ρ, π, χ and ι, or a sequence of them. The functions that generate
the equations are based on the same code that actually evaluates the K - f permutations.
Using C++ template classes, the evaluation of the operations is made symbolically, hence
producing equations. This way of working reduces the chances of making mistakes, as it
avoids to duplicate the description of K - f in C++.
The format of the generated equations has been chosen to match the syntax of SAGE [38].
Also, the bits inside the state are named in such a way that their alphabetical order matches
the bit numbering defined at the level of the sponge construction. This feature comes in
handy when defining a concrete problem to solve, e.g., finding a pre-image or a collision.
One needs to separate bits that are known and fixed from those that are the unknowns of the

48 / 69
4. Analysis of K -f The K reference

problem instance. This separation is defined at the sponge level and can be done alphabeti-
cally on the bit names.
We have installed a SAGE server version 1.4 [38] and automated the tests using Python
scripts [18] interpreted by the SAGE server.

4.2.3 The experiments

In our experiments we have used SAGE to solve a wide range of CICO problems applied
to the K - f permutations. In all problems a subset of the input bits and output bits
are fixed to randomly generated values. More particularly, the range of CICO problems we
investigated can be characterized with the following parameters:

• Number of rounds nr : the number of rounds of K -f

• Width b: the width of K -f

• Rate r: the number of unknown input bits

• Output length n: the number of known output bits

There is a variable for every bit at the input of each round and for the output bits, totalling to
b(nr + 1) variables. There is a round equation for each output bit of every round, expressing
the output bit as a Boolean expression of the input bits. Additionally there is an equation for
each input or output bit that is fixed, simply expressing the equality of the corresponding
variable with a binary constant. Hence in total the number of equations is bnr + (b − r ) + n =
b(nr + 1) + n − r.
We use SAGE to solve the CICO problems using Ideals and Gröbner bases [19]. We pro-
vide a short intuitive explanation here and refer to [19] for thorough treatment of Ideals and
Gröbner bases. To solve a CICO problem we do the following:

• Define a ring R of Boolean polynomials, providing the b(nr + 1) variable names and
specifying the so-called term ordering [19].

• Define an ideal over the ring, providing the b(nr + 1) + n − r equations as generator
polynomials.

• Compute a Gröbner basis for this ideal.

The generated Gröbner basis consists of a sequence of polynomials (equivalent to equations)

that allows to easily generate all solutions of the set of equations generating the ideal. If there
is no solution, it simply consists of the polynomial 1 (implying the equation 1 = 0 that has no
solution). If there is one solution, it simply consists of a sequence of equations of type x + 1
(implying x = 1) or x (implying x = 0). The order of the variable names and term ordering
provided in the definition of the ring R have an impact on the Gröbner basis. The sequence
of polynomials in the Gröbner basis satisfy an ordering that is determined by the order of
variable names and the term ordering. This is not just the way the basis is presented but has
an impact on the way it is generated and hence may impact its computational and memory
complexity.
For executing these functions, SAGE calls an underlying library dedicated to Boolean
polynomials and monomials. This library is called PolyBoRi [17] and is part of the SAGE
distribution. As opposed to the generic polynomial ideal oriented functions in SAGE, it
heavily exploits the fact that Boolean polynomials can be modelled in a very simple way,
with both coefficients and degree per variable in {0, 1}. The ring of Boolean polynomials

49 / 69
The K reference 4. Analysis of K -f

is not a polynomial ring, but rather the quotient ring of the polynomial ring over the field
with two elements modulo the field equations x2 = x for each variable x. Therefore, the
usual polynomial data structures seem not to be appropriate for fast Gröbner basis com-
putations. The PolyBoRi authors state that they introduces a specialised data structure for
Boolean polynomials, capable of handling these polynomials more efficiently with respect
to memory consumption and also computational speed.
If the number of equation is smaller than the number of variables, we expect there to be
multiple solutions. If the number of equations is larger than the number of variables, chances
are that there is no solution at all. A priori, the expected number of solutions of a CICO
problem is 2r−n . In our experiments we have focused on the case r = n where we expect to
have one solution on the average and across experiments we expect the number of solutions
to have a Poisson distribution with λ = 1. This is an interesting case as it corresponds with
the CICO problems encountered when searching a (second) pre-image.
We investigated the case n = r for values of r up to 12, values of nr up to 8 and widths
from 25 to 400. We used lexicographical order with the output variables declared first and
the input variables declared last, as from preliminary experiments this turned out to be the
most efficient choice. For each set of parameters, a number of CICO problems was solved
with at least 5 problems resulting in a solution and 5 problems resulting in the absence of a
solution. We arranged the output of SAGE for analysis in the file Keccak-CICO-results.ods.
Analysing this data, we can make the following observations:

• For small values of r PolyBoRi is efficient in computing the Gröbner basis, and this for
all values of nr and width b.

• For certain parameter choices, solving CICO problems that have no solution takes sig-
nificantly less time than CICO problems that have a solution. The difference is espe-
cially large for large widths and small rate values. When the rate increases, the com-
putation times for the two cases (solution and no solution) converge.

• Doubling the width, keeping all other parameters constant also roughly doubles the
computation time. Hence the computation time appears to grows roughly linearly in
the width.

• Increasing the number of rounds from nr to nr + 1, keeping all other parameters con-
stant, results in an increase roughly independent from the value of nr . Hence the com-
putation time appears to grow linearly in the number of rounds.

• The effect of increasing bitrate by 1 from r to r + 1, keeping all other parameters con-
stant, appears to increase the computation time by a factor that weakly increases with
r and the number of rounds. At r = 12 its value is around 3. If we may extrapolate this
behaviour, solving a CICO problem quickly becomes infeasible with this method as r
grows.

Hence for this class of CICO problems the case of a small rate and small output length can be
relatively easily solved. Although surprising at first sight, this poses no threat to the security
of K - f as such a CICO problem can be efficiently solved by exhaustive search for any
permutation. It suffices to try all 2r possible values of the r unknown input bits, apply the
permutation and verify whether the generated output has the correct value in the known bit
positions.

50 / 69
4. Analysis of K -f The K reference

Rounds Maximum degree Monomials exist

to pass test up to degree
f f −1 f f −1
1 (none) (none) 2 3
2 1 2 4 9
3 6 10 8 17
4 14 18 16 21
5 22 23 22 23
6 23 23 24 24
7-12 24 24 24 24

Table 4.3: The ANF statistical test on K - f [25] and its inverse

4.2.4 Third-party analysis

Aumasson and Khovratovich report in [1] on a empts of solving instances of the CICO prob-
lem for different widths of K - f using a so-called triangulation tool. Solutions were
found for K - f [1600] reduced to three rounds. For more rounds the fast backwards
diffusion in θ apparently prevented solving the CICO problems.
Morawiecki and Srebrny report in [36] on a empts of solving instances of the CICO
problem for different widths of K - f by first expressing the round equations in con-
junctive normal form (CNF) and then run SAT solvers on the resulting set of equations. The
SAT solver performed be er than exhaustive search for some CICO problem instances of
K - f [1600], K - f [200] and K - f [50] reduced to three rounds. For more rounds
the SAT solvers were less efficient than exhaustive search.

4.3 Properties of K - f [25]

The K - f permutations should have no propagation properties significantly different

from that of a random permutation. For the smallest K - f version, K - f [25], it is
possible to experimentally verify certain properties.
First, we report on the algebraic normal form investigations, applied to K - f [25].
Second, we have reconstructed significant parts of the distribution of differential probabil-
ities and input-output correlation of K - f [25] and its reduced-round versions. Third,
we have determined the cycle structure of K - f [25] and all its reduced-round versions.
As a reference for the distributions, we have generated a pseudorandom permutation
operating on 25 bits using a simple algorithm from [32] taking input from a pseudorandom
bit generator based on a cipher that is remote from K and its inventors: RC6 [37]. We
denote this permutation by the term Perm-R.

4.3.1 Algebraic normal statistics

The results of the ANF analysis of K - f [25] are displayed in Table 4.3. Starting from 7
rounds, all monomials up to order 24 exist and appear with a fraction close to one half. Since
K - f [25] is a permutation, the monomial of order 25 does not appear.

51 / 69
The K reference 4. Analysis of K -f

1.E+13
1.E+12
1.E+11
1.E+10
1.E+09
1.E+08
1.E+07
1.E+06
1.E+05
1.E+04
1.E+03
1.E+02
1.E+01
1.E+00
1.E-01

0 2 4 6 8 10 12 14

Figure 4.1: Cardinality histogram of sampling of Perm-R

4.3.2 Differential probability distributions

We have investigated the distribution of the cardinality of differentials for K - f [25],
several reduced-round versions of K - f [25] and of Perm-R. For these permutations, we
have computed the cardinalities of 241 differentials of type ( a′ , b′ ) where a′ ranges over 216
different non-zero input pa erns and b′ over all 225 pa erns. For Perm-R we just tested
the first (when considered as an integer) 216 non-zero input pa erns. For the K - f [25]
16
variants we tested as input pa erns the first 2 non-zero entries in the lookup table of Perm-
R.
In a random permutation the cardinality of differentials has a Poisson distribution with
λ = 1/2. This is studied and described among others in [25]. Moreover, [25] also determines
the distribution of the maximum cardinality of a large number of differentials over a random
permutation. According to [25, Section 5.2], the expected value of the maximum cardinality
over the 241 samples is 12 and the expected value of the maximum cardinality over all 250 − 1
non-trivial differentials ( a′ , b′ ) is 14.
We provide in a sequence of diagrams the histograms obtained from these samplings,
indicating the envelope of the theoretical Poisson distribution for a random permutation
as a continuous line and the results of the measurements as diamond-shaped dots. We have
adopted a logarithmic scale in the y axis to make the deviations stand out as much as possible.
Figure 4.1 shows that Perm-R exhibits a distribution that follows quite closely the theo-
retically predicted one. The maximum observed cardinality is 11.
Figure 4.2 shows the distribution for the two-round version of K - f [25]: the distri-
bution deviates significantly from the theoretical Poisson distribution. Note that here also
the x axis has a logarithmic scale. The largest cardinality encountered is 32768. It turns out
that the pairs of this differential are all in a single trail with weight 9. The number of pairs
is equal to the number of pairs predicted by the weight: 224−9 = 215 . Note that there are
2-round trails with weight 8 (see Table 3.3) but apparently no such trail was encountered in
our sampling.
Figure 4.3 shows the distribution for the three-round version of K - f [25]. The devia-
tion from the theoretical Poisson distribution is smaller. The largest cardinality encountered
is now 146. The pairs of this differential are all in a single trail with weight 17. The number

52 / 69
4. Analysis of K -f The K reference

1.E+12
1.E+11
1.E+10
1.E+09
1.E+08
1.E+07
1.E+06
1.E+05
1.E+04
1.E+03
1.E+02
1.E+01
1.E+00
1.E-01

1 10 100 1000 10000 100000

Figure 4.2: Cardinality histogram of sampling of 2-round version of K - f [25]

of pairs is slightly higher than the number of pairs predicted by the weight: 224−17 = 27 . The
3-round trails with weight 16 (see Table 3.3) were not encountered in our sampling.
Figure 4.4 shows the distribution for the four-round version of K - f [25]. The sam-
pling does no longer allow to distinguish the distribution from that of a random permutation.
The largest cardinality encountered is now 12. The pairs of this differential are in 12 different
trails with weight ranging from 56 to 64. For the 4-round trails with weight 23 (see Table 3.3)
it is not clear whether they were encountered in our sampling: the expected number of pairs
is only 2 and this may have gone unnoticed.
Finally, Figure 4.5 shows the distribution for the 12-round version of K - f [25]. As
expected, the distribution is typical of a random permutation. The maximum cardinality
observed is 12.

4.3.3 Correlation distributions

We have investigated the distribution of the correlations for K - f [25], several reduced-
round versions of K - f [25] and Perm-R. For these permutations, we have computed the
correlations of 239 couples (v, u) where u ranges over 214 different non-zero output masks
and v over all 225 pa erns. For Perm-R we just tested the first (when considered as an integer)
214 non-zero output masks. For the K - f [25] variants we tested as output masks the first
14
2 non-zero entries in the lookup table of Perm-R.
In a random permutation with width b the input-output correlations have a discrete dis-
tribution enveloped by a normal distribution with σ2 = 2−b . This is studied and described
in [25]. Moreover, [25] also determines the distribution of the maximum correlation mag-
nitude of a large number of couples (v, u) over a random permutation. According to [25,
Section 5.4], the expected value of the maximum correlation magnitude over the 239 samples
is 0.00123 and the expected value of the maximum correlation magnitude over all 250 − 1
non-trivial correlations (v, u) is 0.0017.
We provide in a sequence of diagrams the histograms obtained from these samplings,
indicating the envelope of the theoretical normal distribution for a random permutation as
a continuous line and the results of the measurements as diamond-shaped dots. We have

53 / 69
The K reference 4. Analysis of K -f

1.E+12
1.E+11
1.E+10
1.E+09
1.E+08
1.E+07
1.E+06
1.E+05
1.E+04
1.E+03
1.E+02
1.E+01
1.E+00
1.E-01

0 20 40 60 80 100 120 140 160

Figure 4.3: Cardinality histogram of sampling of 3-round version of K - f [25]

1.E+13
1.E+12
1.E+11
1.E+10
1.E+09
1.E+08
1.E+07
1.E+06
1.E+05
1.E+04
1.E+03
1.E+02
1.E+01
1.E+00
1.E-01

0 2 4 6 8 10 12 14

Figure 4.4: Cardinality histogram of sampling of 4-round version of K - f [25]

54 / 69
4. Analysis of K -f The K reference

1.E+13
1.E+12
1.E+11
1.E+10
1.E+09
1.E+08
1.E+07
1.E+06
1.E+05
1.E+04
1.E+03
1.E+02
1.E+01
1.E+00
1.E-01

0 2 4 6 8 10 12 14

Figure 4.5: Cardinality histogram of sampling of K - f [25]

adopted a logarithmic scale in the y axis to make the deviations stand out as much as possible.
Figure 4.6 shows that Perm-R exhibits a distribution that follows quite closely the normal
envelope. At its tails the experimental distribution exhibits its discrete nature. Because it is
a permutation, the correlation can only be non-zero in values that are a multiple of 22−b . For
a given correlation value c that is a multiple of 22−b , the a priori distribution of the corre-
sponding value in the histogram is a Poisson distribution with λ given by the value of the
normal envelope in that point. The largest correlation magnitude observed is 0.001226, quite
close to the theoretically predicted value.
Figure 4.7 shows the distribution for the two-round version of K - f [25]: the distri-
bution deviates significantly from the theoretical normal envelope. Additionally, it is zero
for all values that are not a multiple of 2−15 (rather than 2−23 ). This is due to the fact that the
Boolean component functions of K - f [25] have only reached degree 4 a er two rounds,
rather than full degree 24. The largest correlation magnitude encountered is 0.03125 (outside
the scale of the figure). This is the correlation magnitude 2−5 one would obtain by a single
linear trail with weight 10. By measuring the correlation of the same pair of masks for vari-
ants of the two-round version of K - f [25] where different constant vectors are XORed
in between the two rounds, it turns out that the correlation value is either 25 or −2−5 . This
implies that the correlation is the result of a single trail. The 2-round linear trails with weight
8 (see Table 3.3) were apparently not encountered in our sampling.
Figure 4.8 shows the distribution for the three-round version of K - f [25]: the devi-
ation from the theoretical normal envelope becomes smaller. This distribution is zero for all
values that are not a multiple of 2−18 due to the fact that the Boolean component functions
of K - f [25] have only reached degree 8 a er three rounds. The largest correlation mag-
nitude encountered is 0.003479. This is a correlation magnitude that cannot be obtained by
a single linear trail. 3-round linear trails with weight 16 would give correlation magnitude
2−8 ≈ 0.0039. It is quite possible that the observed correlation value is the sum of the (signed)
correlation contributions of some trails, including one with weight 16 and some with higher
weight. By measuring the correlation of this pair of masks in variants of the three-round
version of K - f [25] where different constant vectors are XORed in between the rounds,
we obtain 491 different values. This implies that this correlation has contributions from at

55 / 69
The K reference 4. Analysis of K -f

1.E+09

1.E+08

1.E+07

1.E+06

1.E+05

1.E+04

1.E+03

1.E+02

1.E+01

1.E+00

1.E-01
-0.0014 -0.0012 -0.001 -0.0008 -0.0006 -0.0004 -0.0002 0 0.0002 0.0004 0.0006 0.0008 0.001 0.0012 0.0014

Figure 4.6: Correlation histogram of sampling of Perm-R

1.E+11

1.E+10

1.E+09

1.E+08

1.E+07

1.E+06

1.E+05

1.E+04

1.E+03

1.E+02

1.E+01

1.E+00

1.E-01
-0.01 -0.008 -0.006 -0.004 -0.002 0 0.002 0.004 0.006 0.008 0.01

Figure 4.7: Correlation histogram of sampling of 2-round version of K - f [25]

56 / 69
4. Analysis of K -f The K reference

1.E+10

1.E+09

1.E+08

1.E+07

1.E+06

1.E+05

1.E+04

1.E+03

1.E+02

1.E+01

1.E+00

1.E-01
-0.004 -0.003 -0.002 -0.001 0 0.001 0.002 0.003 0.004

Figure 4.8: Correlation histogram of sampling of 3-round version of K - f [25]

least 9 trails.
Figure 4.9 shows the distribution for the four-round version of K - f [25]. The shape
of the distribution and the maximum values do no longer allow to distinguish the distribu-
tion from that of a random permutation. The largest correlation magnitude encountered is
0.001196. However, this distribution differs from that of a random permutation because it is
zero for all values that are not a multiple of 2−20 due to the fact that the Boolean component
functions of K - f [25] have only reached degree 16 a er four rounds. By measuring the
correlation of this pair of masks in variants of the four-round version of K - f [25] where
different constant vectors are XORed in between the rounds, we obtain many different val-
ues implying that this correlation is the result of a large amount of trails. Moreover, the value
of the correlation exhibits a normal distribution.
A er 5 rounds the distribution is zero for values that are not a multiple of 2−22 and only
a er 6 rounds this becomes 2−23 .
Finally, Figure 4.10 shows the distribution for the 12-round version of K - f [25]. As
expected, the distribution is typical of a random permutation. The maximum correlation
magnitude observed is 0.001226.

4.3.4 Cycle distributions

We have determined the cycle structure of K - f [25] and all its reduced-round versions.
Table 4.4 lists all cycles for K - f [25] and Table 4.5 the number of cycles for all reduced-
round versions. For a random permutation, the expected value of the number of cycles is
ln(225 ) = 25 ln 2 ≈ 17.3. The average of Table 4.5 is 16.3.
It can be observed that K - f [25] and all its reduced-round versions have an even
number of cycles. For a permutation operating on a domain with an even number of ele-
ments, an even number of cycles implies that it is an even permutation [40], hence they are
all even permutations. Actually, it is easy to demonstrate that all members of the K -f
family are even permutations. We do however not think this property can be exploited in an
a ack or to build a usable distinguisher.
The members of the K - f family are even permutations because the composition of

57 / 69
The K reference 4. Analysis of K -f

1.E+10

1.E+09

1.E+08

1.E+07

1.E+06

1.E+05

1.E+04

1.E+03

1.E+02

1.E+01

1.E+00

1.E-01
-0.0014 -0.0012 -0.001 -0.0008 -0.0006 -0.0004 -0.0002 0 0.0002 0.0004 0.0006 0.0008 0.001 0.0012 0.0014

Figure 4.9: Correlation histogram of sampling of 4-round version of K - f [25]

1.E+09

1.E+08

1.E+07

1.E+06

1.E+05

1.E+04

1.E+03

1.E+02

1.E+01

1.E+00

1.E-01
-0.0014 -0.0012 -0.001 -0.0008 -0.0006 -0.0004 -0.0002 0 0.0002 0.0004 0.0006 0.0008 0.001 0.0012 0.0014

Figure 4.10: Correlation histogram of sampling of K - f [25]

58 / 69
4. Analysis of K -f The K reference

18447749 147821 168 12

13104259 40365 27 3
1811878 2134 14 2

Table 4.4: Cycle lengths in K - f [25]

rounds cycles rounds cycles rounds cycles

1 14 5 18 9 14
2 12 6 20 10 20
3 16 7 18 11 18
4 16 8 18 12 12

Table 4.5: Number of cycles in reduced-round versions of K - f [25]

two even permutation is an even permutation and that all step mappings of K - f are
even permutations. We cite here a number of arguments we found in [40, Lemma 2]:

• The mappings θ, π and ρ are linear. In fact all invertible linear transformations over Z2b
with b > 2 are even permutations. This follows from the fact that each invertible binary
matrix can be obtained from the identity matrix by elementary row transformations
(binary addition of one row to another row), and that these elementary row transfor-
mations (considered as linear mappings) are permutations with 2b−1 fixed points and
2b−2 cycles of length 2.

• The mapping ι consists of the addition of a constant. Addition of a constant in Z2b is

the identity mapping if the constant is all-zero and has 2b−1 cycles of length 2 if the
constant is not all-zero.

• The mapping χ is an even permutation because it can be represented as a composition

of 5w permutations that carry out the χ mapping for one row and leave the other 5w − 1
rows fixed. The cycle representation of each such permutation contains a number of
cycles that is a multiple of 225w−5 and hence even.

4.4 Distinguishers exploiting low algebraic degree

The K - f round function and its inverse have low algebraic degrees, 2 and 3 respectively.
This has been exploited in third-party cryptanalysis to construct a number of distinguishers.
Aumasson and Khovratovich report in [1] on observations showing that K - f [1600]
reduced to 3 and 4 rounds does not have an ideal algebraic structure and conjectures that
this can be observed up to reduced-round versions of ten rounds, for which the algebraic
degree is at most 1024 and hence not maximal.
In [34] Lathrop tested the resistance of K against cube a acks by executing efficient
cube a ack against the 224-bit output-length version of K configured as a MAC and
calling a version of K - f [1600] reduced to 4 rounds or less. Based on his analysis, the au-
thor suggests that a cube a ack against such a configuration would only be practical against
reduced-round versions up to 7 rounds.

59 / 69
The K reference 4. Analysis of K -f

4.4.1 Zero-sum distinguishers

In [2] Aumasson and Meier introduced very simple and elegant distinguishers against reduced-
round versions of K - f [1600]. They are based on the simple observation that the degree
of n K - f rounds is at most 2n and that the degree of n inverse rounds is at most 3n . We
cite [2]:
Suppose one fixes 1600 − 513 = 1087 bits of the initial state to some arbitrary value, and consider
the 2513 states obtained by varying the 513 bits le . Our main observation is that applying the 9-round
K - f to each of those states and xoring the 2513 1600-bit final states obtained yields the zero state.
This is because, for each of the 1600 Boolean components, the value obtained is the order-513 derivative
of a degree-512 polynomial, which by definition is null.
If these states are chosen in some intermediate round and one computes the correspond-
ing inputs and outputs, this is also the case for those inputs as long as the polynomials ex-
pressing the inputs in terms of the intermediate state bits have small enough degree. Hence
one has a systematic way to construct a set of inputs that XOR to zero and for which the cor-
responding outputs XOR to zero, which is qualitatively different from the generic method
[7]. Additionally, if the positions of the free bits in the intermediate round are chosen care-
fully, one may even reduce the degree of the forward and/or backward polynomials. Using
these simple elements, Aumasson and Meier constructs distinguishers for up to 16 rounds of
K - f [1600] [2]. This last one makes use of 6 backward rounds and 10 forward rounds.
Given an upper bound of N free bits one can construct a zero-sum distinguisher with
log2 N forward rounds and log3 N backward rounds. While in a distinguisher using only
the forward direction maximum degree cannot be reached before 11 rounds, the start-in-
the-middle technique allows going up to 16 rounds.
This was further refined by Boura and Canteaut in two papers [15, 14] extending the
distinguishers to 18 rounds for K - f [1600] in [15] and up to 20 rounds in [14]. The new
ideas are the following. First, the Walsh spectrum of χ allows one to bound the degree of
the inverse rounds more tightly. In particular, while 7 inverse rounds are expected to be
of maximal degree (as 37 > 1599), they show that it cannot be higher than 1369. Second,
by aligning the bits that are varied and those that are fixed on row boundaries, χ works
independently and bijectively row per row. This can be generalized to cover several rounds.
Boura and Canteaut were joined by De Cannière in [16] and they extended the distin-
guisher to the full 24 rounds of K - f [1600]. The novel idea that allowed this is the dis-
covery of a saturation effect imposing an upper bound on the degree of iterated permuta-
tions. They showed this effect for round functions in which the nonlinear step consists of
the parallel application of S-boxes with a very elegant and simple argument. The effect is
strongest for small S-boxes but can also be observed for large S-boxes. While initially the
algebraic degree D increases exponentially by a factor d (with d the algebraic degree of the
round function) per round, above some degree it only converges exponentially to b.
Finally, Duan and Lai further considered the evolution of the degree of the inverse per-
mutation [26]. They noticed that the product of two output bits of χ−1 yields degree-3 poly-
nomials, and this leads to be er bounds on the degree of the inverse rounds. In Table 4.6, we
list the bounds on the algebraic degrees of reduced-round versions of K - f [1600] and
its inverse, based on this property and our own analysis.
Triggered by the initial publication on zero-sums by Aumasson and Meier in [2], we
decided to increase the number of rounds in K - f from 12 + ℓ to 12 + 2ℓ. The motivation
behind this was the following. The applicability of the zero-sum distinguishers is limited by
the maximum number of free bits N, namely the width. Doubling the width allows to add a
forward round in the distinguisher and possibly a backward round. So when doubling the
width, roughly two additional rounds are required to provide resistance against zero-zum

60 / 69
4. Analysis of K -f The K reference

# Rounds Degree f 1600− Degree f Degree f −1 1600− Degree f −1

1 2 3
2 4 9
3 8 27
4 16 81
5 32 243
6 64 729 871
7 128 1164 432
8 256 1382 218
9 512 1491 109
10 1024 576 1545 55
11 1408 192 1572 28
12 1536 64 1586 14
13 1578 22 1593 7
14 1592 8 1596 4
15 1597 3 1598 2
16 1599 1 1599 1

Table 4.6: Bounds on the algebraic degrees of reduced-round versions of K - f [1600]

and its inverse.

distinguishers. This may be the case for other distinguishers too and so we felt it would be
more appropriate to increase the number of rounds by 2 when doubling the width.
The zero-sum partitions distinguish the 24 rounds of K - f [1600] from a randomly-
chosen permutation, although without implying a distinguisher on K itself [16, 26].
Not increasing the number of rounds, strictly speaking, contradicts the hermetic sponge
strategy. Still, we decided to stick the number of rounds specified by 12 + 2ℓ due to the fact
that these distinguishers can in no way be used to a ack K for many reasons, of which
the following two are the most important (see also [8, Section “The usability of structural
distinguishers”]).
First, the advantage of a zero-sum distinguisher (or partition) is 0 for any number of
queries N to f or f −1 with N smaller than the size of the partition. This means that using
a permutation f that has zero-sum distinguishers has no impact on the RO differentiating
advantage for N below the size of the partition. In the flat sponge claim, no resistance is
claimed for a acks requiring a workload of more than 2c/2 queries to f and the maximum
value for the capacity is c = b − 1. So distinguishers that have zero advantage for N below
2b/2 do not compromise the flat sponge claim. The zero-sum distinguishers for the full 24
rounds of K - f [1600] have zero advantage below something like 21575 queries. Only
structural distinguishers on f that have non-zero advantage below 2800 queries can possibly
qualify as a threat for the security of a sponge function that uses it.
Second, exploiting zero-sum distinguishers would require the adversary to apply inputs
to the sponge function such that the input to f has the values of her choice, for many different
inputs. In the sponge construction, an a acker can not choose the value of bits in the inner
part of the input to f directly, but only influence their value in an indirect way by injecting
bits in the previous iteration of f .
Further discussions about the applicability of the zero-sum distinguishers can be found
in [7].

61 / 69
The K reference 4. Analysis of K -f

4.4.2 Pre-image a acks

The first exploitation of the low algebraic degree of the K - f round function in an actual
a ack on K was published by Bernstein in [3]. It is a pre-image a ack that, making
abstraction of the cost of memory access, would be able to break the claimed security level
for reduced-round versions of K - f [1600] up to 8 rounds. When counting the cost of
memory access, these a acks are much less efficient than generic pre-image a acks.
The a ack exploits the fact that K - f reduced to 6, 7 and 8 rounds has an algebraic
degree of only 64, 128 and 256 respectively. This allows evaluating a subset of the output bits
for large sets of inputs more efficiently than computing K - f [1600] for the same sets of
inputs. This can be used as a filter for second pre-image candidates, reducing the number of
candidates for which K - f [1600] must be computed to only a fraction.
We give here short summary of what the a ack can do.
Consider a hash function with an m-bit digest and consider messages of a certain fixed
length L. Let there be:

• Subset A (with | A| = n) of the message bit positions and

• Subset B (with | B| = m) of the digest bit positions, such that for any fixed value of the
message bits not in A, the bits in B have algebraic degree d < n as functions of bits in
A.

If this degree d is sufficiently small, the computation effort of finding (second) pre-images
can be reduced.
This a ack can be applied to a hash function (possibly round-reduced) where a subset
B of the digest bits have a low algebraic degree in a subset A of the message bits of the last
block. The bits in B are only separated from the bits in A by a single compression function
(or permutation) call.
The downside of the a ack is that this workload reduction comes at the cost of memory.
For it to be a reduction, one considers the computational effort not including the cost of mem-
ory access. The memory required for the a ack is m2n bits. This may be reduced somewhat
by some clever programming techniques at the cost of increasing the computational effort.
For K with K - f [1600] reduced to 6, 7 or 8 rounds, the length of the message L
plays no role. The active bits must all be in the last block, so n is limited by the bitrate (minus
the padding). One can tune m and n resulting in different memory-time trade-offs. The ones
reported in [3] are:

• 6 rounds: 2176 bits of memory give a workload reduction by a factor 50 ( ≈ 6 bits)

• 7 rounds: 2320 bits of memory give a workload reduction by a factor 37 ( ≈ 5 bits)

• 8 rounds: 2508 bits of memory give a workload reduction by a factor 1.4 (half a bit)

62 / 69
Chapter 5

Design rationale summary

The purpose of this chapter is to list the design choices and to briefly motivate them, although
further analysis is provided in the subsequent chapters.

5.1 Choosing the sponge construction

The K hash function makes use of the sponge construction, following the definition of
[8]. This results in the following property:
Provability It has a proven upper bound for the success probability, and hence also a lower
bound for the expected workload, of generic a acks. We refer to [8, Chapter “Security
proofs”] for a more in-depth discussion.
The design philosophy underlying K is the hermetic sponge strategy. This consists
of using the sponge construction for having provable security against all generic a acks and
calling a permutation (or transformation) that should not have structural properties with the
exception of a compact description [8, Section “The philosophy”].
Additionally, the sponge construction has the following advantages over constructions
that make use of a compression function:
Simplicity Compared to the other constructions for which upper bounds have been proven
for the success of generic a acks, the sponge construction is very simple, and it also
provides a bound that can be expressed in a simple way.

Variable-length output It can generate outputs of any length and hence a single function
can be used for different output lengths.

Flexibility Security level can be incremented at the cost of speed by trading in bitrate for
capacity, using the same permutation (or transformation).

Functionality Thanks to its long outputs and proven security bounds with respect to generic
a acks, a sponge function can be used in a straightforward way as a MAC function,
stream cipher and a mask generating function. Thanks to the duplex construction,
a sponge function can be used as a reseedable pseudorandom bit generator and for
efficient authenticated encryption (see [8, Section “Authenticated encryption”]).
To support arbitrary bit strings as input, the sponge construction requires a padding func-
tion. We have chosen the simplest padding rule that allows to use the same permutation in
combination with different bitrate values [8, Section “Optimum security of multi-rate sponge
functions”] without loss of security.

63 / 69
The K reference 5. Design rationale summary

5.2 Choosing an iterated permutation

The sponge construction requires an underlying function f , either a transformation or a per-
mutation. Informally speaking, f should be such that it does not have specific properties that can
be exploited in a acks. We have chosen a permutation, constructed as a sequence of (almost)
identical rounds because of the following advantages:
Block cipher experience An iterated permutation is an iterated block cipher with a fixed
key. In its design one can build on knowledge obtained from block cipher design and
cryptanalysis (see Chapter 2).
Memory efficiency O en a transformation is built by taking a permutation and adding a
feedforward loop. This implies that (at least part of) the input must be kept during the
complete computation. This is not the case for a permutation, leading to a relatively
small RAM footprint.
Compactness Iteration of a single round leads to a compact specification and potentially
compact code and hardware circuits.

5.3 Designing the K - f permutations

The design criterion for the K - f permutations is to have no specific properties that can
be exploited in an a ack when being used in the sponge construction. It is constructed as
an iterated block cipher similar to N [23] and R [24], with the key schedule
replaced by some simple round constants. Here we give a rationale for its features:
Bit-oriented structure A acks where the bits are grouped (e.g., in bytes), such as integral
cryptanalysis and truncated trails or differentials, are unsuitable against the K -f
structure.
Bitwise logical operations and fixed rotations Dependence on CPU word length is only due
to rotations, leading to an efficient use of CPU resources on a wide range of processors.
Implementation requires no large tables, removing the risk of table-lookup based cache
miss a acks. They can be programmed as a fixed sequence of instructions, providing
protection against timing a acks.
Symmetry This allows to have very compact code in so ware and a very compact co-processor
circuit suitable for constrained environments [10] .
Parallelism Thanks to its symmetry and the chosen operations, the design is well-suited for
ultra-fast hardware implementations and the exploitation of SIMD instructions and
pipelining in CPUs.
Round degree 2 This makes the analysis with respect to differential and linear cryptanal-
ysis easier, leads to relatively simple (albeit large) systems of algebraic equations and
allows the usage of very powerful protection measures against differential power anal-
ysis (DPA) both in so ware and hardware that are not suited for most other nonlinear
functions [6].
Matryoshka structure The analysis of small versions is relevant for larger versions (see Sec-
tion 2.2).
Number of rounds nr = 12 + 2ℓ: The value of nr has been chosen to have a good safety
margin and still have good performance. See Section 5.4.

64 / 69
5. Design rationale summary The K reference

A ack Number of rounds

K - f distinguisher below 2b/2 9 + 2ℓ
K distinguisher 7+ℓ
Inner collision 5+ℓ
State recovery 5+ℓ

Table 5.1: Sufficient number of rounds of K - f to resist to different types of a ack or

distinguisher

5.4 Strength estimation

We here provide our estimates of how many rounds in K - f are sufficient to provide
resistance against two types of distinguisher and two types of a ack. These estimations cover
all instances of K [r, c].
K - f distinguisher below 2b/2 A structural distinguisher for K - f [b] requiring less
than 2b/2 queries, for any supported width.

K distinguisher A structural distinguisher for K [r, c].

Inner collision The generation of an inner collision with success probability above N 2 2−(c+1) ,
with a workload equivalent to N calls to K - f [r + c ].

State recovery State recovery from a keyed sponge with success probability above N M2−c
with a workload equivalent to N calls to K - f [r + c] and a sequence of queries to
the keyed sponge requiring it to make in total M calls to the permutation.

We list the number of rounds that we estimate to be sufficient to resist these a acks in
Table 5.1. These estimates are based on the results of our preliminary analysis that is treated
in this document and the third-party analysis in [1, 34, 2, 15, 14, 36, 16, 3].

65 / 69
The K reference 5. Design rationale summary

66 / 69
Bibliography

[1] J.-P. Aumasson and D. Khovratovich, First analysis of Keccak, Available online, 2009,
http://131002.net/data/papers/AK09.pdf.

[2] J.-P. Aumasson and W. Meier, Zero-sum distinguishers for reduced Keccak-f and for the core
functions of Luffa and Hamsi, Available online, 2009, http://131002.net/data/papers/
AM09.pdf.

[3] D. J. Bernstein, Second preimages for 6 (7? (8??)) rounds of keccak?, 2010,
h p://cr.yp.to/hash/keccak-20101127.txt.

[4] G. Bertoni, J. Daemen, M. Peeters, and G. Van Assche, R G , a belt-and-mill hash

function, Second Cryptographic Hash Workshop, Santa Barbara, August 2006, http://
radiogatun.noekeon.org/.

[5] ,K specifications, version 2, NIST SHA-3 Submission, September 2009, http:

//keccak.noekeon.org/.

[6] , Note on side-channel a acks and their countermeasures, Comment on

the NIST Hash Competition Forum, May 2009, http://keccak.noekeon.org/
NoteSideChannelAttacks.pdf.

[7] , Note on zero-sum distinguishers of K - f , Comment on the NIST Hash Com-

petition Forum, January 2010, http://keccak.noekeon.org/NoteZeroSum.pdf.

[8] , Cryptographic sponge functions, January 2011, http://sponge.noekeon.org/.

[9] ,K T so ware, January 2011, http://keccak.noekeon.org/.

[10] G. Bertoni, J. Daemen, M. Peeters, G. Van Assche, and R. Van Keer, K implementa-
tion overview, January 2011, http://keccak.noekeon.org/.

[11] E. Biham, O. Dunkelman, and N. Keller, The rectangle a ack - rectangling the serpent, Ad-
vances in Cryptology – Eurocrypt 2001 (B. Pfitzmann, ed.), Lecture Notes in Computer
Science, vol. 2045, Springer, 2001, pp. 340–357.

[12] A. Biryukov and D. Wagner, Slide a acks, in Knudsen [31], pp. 245–259.

[13] C. Bouillaguet and P.-A. Fouque, Analysis of the collision resistance of RadioGatún using al-
gebraic techniques, Selected Areas in Cryptography, Lecture Notes in Computer Science,
vol. 4876, Springer, 2008.

[14] C. Boura and A. Canteaut, Zero-sum distinguishers on the Keccak-f permutation with 20
rounds (working dra ), private communication, 2010.

67 / 69
The K reference BIBLIOGRAPHY

[15] , A zero-sum property for the Keccak-f permutation with 18 rounds, Available online,
2010, http://www-roc.inria.fr/secret/Anne.Canteaut/Publications/zero_sum.pdf.

[16] C. Boura, A. Canteaut, and C. De Cannière, Higher-order differential properties of Keccak

and Luffa, Fast So ware Encryption 2011, 2011, to appear, dra available from Cryptol-
ogy ePrint Archive, Report 2010/589.

[17] M. Brickenstein and A. Dreyer, PolyBoRi: A framework for Gröbner-basis computations with
Boolean polynomials, Journal of Symbolic Computation 44 (2009), no. 9, 1326–1345, Effec-
tive Methods in Algebraic Geometry.

[18] The Python community, Python Programming Language, Python So ware Foundation,
2009, http://www.python.org/.

[19] D. A. Cox, J. B. Li le, and D. O’Shea, Ideals, varieties, and algorithms, third ed., Springer,
2007.

[20] J. Daemen, Cipher and hash function design strategies based on linear and differential crypt-
analysis, PhD thesis, K.U.Leuven, 1995.

[21] J. Daemen and C. S. K. Clapp, Fast hashing and stream encryption with PANAMA, Fast
So ware Encryption 1998 (S. Vaudenay, ed.), LNCS, no. 1372, Springer-Verlag, 1998,
pp. 60–74.

[22] J. Daemen, L. R. Knudsen, and V. Rijmen, The block cipher Square, Fast So ware Encryp-
tion 1997 (E. Biham, ed.), Lecture Notes in Computer Science, vol. 1267, Springer, 1997,
pp. 149–165.

[23] J. Daemen, M. Peeters, G. Van Assche, and V. Rijmen, Nessie proposal: the block cipher
, Nessie submission, 2000, http://gro.noekeon.org/.

[24] J. Daemen and V. Rijmen, The design of Rijndael — AES, the advanced encryption standard,
Springer-Verlag, 2002.

[25] , Probability distributions of correlation and differentials in block ciphers, Journal of

Mathematical Cryptology 1 (2007), no. 3, 221–242.

[26] M. Duan and X. Lai, Improved zero-sum distinguisher for full round keccak-f permutation,
Cryptology ePrint Archive, Report 2011/023, 2011, http://eprint.iacr.org/.

[27] M. Gorski, S. Lucks, and T. Peyrin, Slide a acks on a class of hash functions, Asi-
acrypt (J. Pieprzyk, ed.), Lecture Notes in Computer Science, vol. 5350, Springer, 2008,
pp. 143–160.

[28] J. Kelsey, T. Kohno, and B. Schneier, Amplified boomerang a acks against reduced-round
mars and serpent, Fast So ware Encryption 2000 (B. Schneier, ed.), Lecture Notes in Com-
puter Science, vol. 1978, Springer, 2000, pp. 75–93.

[29] D. Khovratovich, Two a acks on RadioGatún, 9th International Conference on Cryptology

in India, 2008.

[30] L. R. Knudsen, Truncated and higher order differentials, Fast So ware Encryption 1994
(B. Preneel, ed.), Lecture Notes in Computer Science, vol. 1008, Springer, 1994,
pp. 196–211.

68 / 69
BIBLIOGRAPHY The K reference

[31] L. R. Knudsen (ed.), Fast so ware encryption, 6th international workshop, fse ’99, rome, italy,
march 24-26, 1999, proceedings, Lecture Notes in Computer Science, vol. 1636, Springer,
1999.

[32] D. E. Knuth, The art of computer programming, vol. 2, third edition, Addison-Wesley Pub-
lishing Company, 1998.

[33] S. K. Langford and M. E. Hellman, Differential-linear cryptanalysis, Advances in Cryp-

tology – Crypto ’94 (Y. Desmedt, ed.), Lecture Notes in Computer Science, vol. 839,
Springer, 1994, pp. 17–25.

[34] J. Lathrop, Cube a acks on cryptographic hash functions, Master’s thesis, Available online,
2009, http://www.cs.rit.edu/~jal6806/thesis/.

[35] A. J. Menezes, P. C. van Oorschot, and S. A. Vanstone, Handbook of applied cryptography,

CRC Press, 1997.

[36] P. Morawiecki and M. Srebrny, A sat-based preimage analysis of reduced KECCAK hash
functions, Cryptology ePrint Archive, Report 2010/285, 2010, http://eprint.iacr.org/.

[37] R. L. Rivest, M. J. B. Robshaw, R. Sidney, and Y. L. Yin, The RC6 block cipher, AES pro-
posal, August 1998.

[38] W. A. Stein et al., Sage Mathematics So ware, The Sage Development Team, 2009, http:
//www.sagemath.org/.

[39] D. Wagner, The boomerang a ack, in Knudsen [31], pp. 156–170.

[40] R. Wernsdorf, The round functions of Rijndael generate the alternating group, Fast So ware
Encryption 2002 (J. Daemen and V. Rijmen, eds.), Lecture Notes in Computer Science,
vol. 2365, Springer, 2002, pp. 143–148.

[41] Wikipedia, Impossible differential cryptanalysis, 2008, http://en.wikipedia.org/wiki/

Miss_in_the_middle_attack.

[42] M. R. Z’aba, H. Raddum, M. Henricksen, and E. Dawson, Bit-pa ern based integral a ack,
Fast So ware Encryption 2008 (K. Nyberg, ed.), Lecture Notes in Computer Science,
vol. 5086, Springer, 2008, pp. 363–381.

69 / 69