Thanks to visit codestin.com
Credit goes to www.scribd.com

Scribd
Upload
66 views2 pages

Deepxplore

DeepXplore is an automated white-box testing technique for deep learning systems. It formulates finding test inputs as an optimization problem to maximize neuron coverage and differential behavior across neural networks. Starting with a seed input, it uses gradient ascent to find an input region where all but one network classify inputs in the same class, displaying differential behavior. This helps uncover errors missed by normal testing. DeepXplore applies constraints to keep generated inputs realistic while modifying seeds. It found thousands of corner cases for standard datasets and improved DNN accuracy.

Uploaded by

Ankur Tomar
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
66 views2 pages

Deepxplore

DeepXplore is an automated white-box testing technique for deep learning systems. It formulates finding test inputs as an optimization problem to maximize neuron coverage and differential behavior across neural networks. Starting with a seed input, it uses gradient ascent to find an input region where all but one network classify inputs in the same class, displaying differential behavior. This helps uncover errors missed by normal testing. DeepXplore applies constraints to keep generated inputs realistic while modifying seeds. It found thousands of corner cases for standard datasets and improved DNN accuracy.

Uploaded by

Ankur Tomar
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 2

DeepXplore: Automated White-box Testing of Deep Learning Systems

Summary

The paper proposes automated whitebox testing based on an optimization problem of maxi-
mizing neuron coverage and differential behavior for a set of DNNs. It justifies finding inputs
to maximize neuron coverage and differential behavior as optimization problem that can be
solved using gradient ascent technique. Starting with a seed input, DeepXplore finds the input
region in which all but one DNN classify the inputs as one class, displaying differential behavior.
Neuron coverage measures span of test inputs over DNN and maximizes it to uncover erroneous
behaviors missed in normal testing. While modifying seeds, DeepXplore imposes domain con-
straints to keep inputs realistic; constraints are based on type of data the DNN is trained on.
DeepXplore finds thousands of corner cases within short span for standard datasets; further-
more these cases are used to improve accuracy of DNNs. Performance of neuron coverage is
compared with code coverage as in software analysis and of DeepXplore with random sampling
and adversarial testing.

Critical review and improvements

• Neuron coverage is an apt metric for DNN testing. However, it is unclear as to why it is
compared with code coverage because, DNNs are models which learn and embed rules by
observations and there is no concept of code of DNN.

• DeepXplore identifies and maximizes differential behavior of a set of DNNs. This implies
whenever all but one DNN classify input in one class it is assumed to be correct based on
intuition that probability of happening otherwise is low. This needs to be evaluated by
using the training datasets of different DNNs that belong to same set of classes but are
disjoint.

• The paper assumes that a behavior is differential if all but one DNN classifies differently.
Hence, it should be investigated that how much the definition of differential
P behavior can
be relaxed? In P other words current
P obj1 (x) should be changed from k6=j Fk (x)[c] −
λ1 .Fj (x)[c] to k∈K Fk (x)[c] − λ1 . j∈J Fj (x)[c] such that K ∩ J = ∅ and K ∪ J = N ; N
being the set of DNNs under test.

• Neuron coverage is reported for non-fully-connected layers as for these layers activating
neurons is daunting. To tackle this scenario, one can explore the possible space of inputs
(not as optimization) by changing feature values but will require applying appropriate
domain constraints.

• Activation of neurons for different sample classes is argued to be better than that for same
classes. However, the reported value of average number of activated neurons is more for
the latter. This implies even if there is high redundancy (overlap), utility of latter is same
which needs to be explained better.

• Performance of DeepXplore with random sampling and adversarial testing is compared


based on a metric that forms the basis of it’s optimization function which seems redundant
due to obviousness.

• It is clear the DeepXplore generates test inputs for DNNs which are then applied to many
DNNs before deployment. If number of products, say Tesla vehicle, are high this can

1
result in high test time. So to reduce it, possibility of reducing test inputs by combination
while keeping neuron coverage high should be explored.

You might also like