Generating real random bytes requires to get different results from the execution of a same operation. It is a problem for the computer, whose operations are normally deterministic (like 1 + 1 always results 2), to generate real random bytes.

We observe that the time costs can be different for the computer to run the same code. For example, we run the below code for twice:

from time import time
print([-(time() - time()) for _ in range(8)])

and it reports:

[-0.0, 1.1920928955078125e-06, -0.0, -0.0, 9.5367431640625e-07, -0.0, -0.0, -0.0] and [-0.0, -0.0, -0.0, -0.0, -0.0, -0.0, 9.5367431640625e-07, -0.0] respectively, showing the differences.

Based on this observation, we may record the time costs during the execution of codes, and generate real random bytes based on the records.

We implement this idea, (T)ime differences between (O)perations for real (Rand)om bytes generation, in python as this library.

Install with pip install torand.

torandom(k, fast_mode=True, use_cache=True)

Generate k random bytes. Where fast_mode (default to True) and use_cache (default to True) indicate the use of the hash-based fast generation and a small random bytes' cache respectively. Setting use_cache to True is beneficial for the performance when the function will be called to generate a small number (< 32) of random bytes for many times, but insecure as the cache can be manually assigned by the other codes that imports this library.

from torand import torandom

k = 127
rs = torandom(k, fast_mode=True, use_cache=True)
print(rs)
print(len(rs) == k)

Execute either

python -m torand (f/s) k

or

torand (f/s) k

to generate k real random bytes. Where f and s are optional arguments for the hash-based fast generation mode and the standard generation mode respectively, default is f.

Test with CPython 3.10.5, using a single Intel Core M3-7Y30 CPU thread. Despite that this library only generates ~87.57 k random bytes per second in the standard mode, which is much slower than the os.urandom (~12.52 MB/s). We still like to share the idea, and the library is also useful in case we only generate some real random bytes with the idea for fast random byte generation algorithms or hash functions (this library can also achieve a generation speed of ~12.52 MB/s in the fast mode).