Insecure Random Number Generator

[gonzalo-munillag]

Hello,

I would like to bring to your attention that using the random number generator from TensorFlow could lead to vulnerabilities when sampling from a distribution to fulfill differential privacy during training: https://www.tmlt.io/research/tiny-bits-matter-precision-based-attacks-on-differential-privacy

PyTorch Opacus uses a secure RNG: https://opacus.ai/api/privacy_engine.html

In contrast, TensorFlow RNG:
https://www.tensorflow.org/api_docs/python/tf/random/Generator
https://stackoverflow.com/questions/63350248/is-tf-random-normal-cryptographically-secure

Additionally, there is no documentation that states the use of floating-point vulnerability protection as in https://scholar.google.com/citations?view_op=view_citation&hl=en&user=hg3A9TgAAAAJ&citation_for_view=hg3A9TgAAAAJ:dhFuZR0502QC
and
https://research.ibm.com/publications/secure-random-sampling-in-differential-privacy

Kind regards,
Gonzalo

[kairouzp]

Thanks so much for pointing this out. We are aware of this issue (see Section 8 of https://arxiv.org/pdf/1812.06210.pdf).

…

On Mon, Nov 7, 2022 at 2:50 PM Gonzalo Munilla Garrido < ***@***.***> wrote: Hello, I would like to bring to your attention that using the random number generator from TensorFlow *could* lead to vulnerabilities when sampling from a distribution to fulfill differential privacy during training: https://www.tmlt.io/research/tiny-bits-matter-precision-based-attacks-on-differential-privacy PyTorch Opacus uses a secure RNG: https://opacus.ai/api/privacy_engine.html In contrast, TensorFlow RNG: https://www.tensorflow.org/api_docs/python/tf/random/Generator https://stackoverflow.com/questions/63350248/is-tf-random-normal-cryptographically-secure Kind regards, Gonzalo — Reply to this email directly, view it on GitHub <#349>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AB7LCU7GVRS2JMITWENY63TWHGBUJANCNFSM6AAAAAARZV5VUU> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>