Learning to localize leakage of cryptographic keys through power consumption

Jimmy Gammell, Anand Raghunathan, Abolfazl Hashemi, Kaushik Roy

Keywords: deep learning, adversarial, mutual information, side channel attack, cryptography

TL;DR: A principled deep learning framework for defending against power side-channel attacks on AES hardware.

Abstract: While cryptographic algorithms such as the ubiquitous Advanced Encryption Standard (AES) are secure, *physical implementations* of these algorithms in hardware inevitably 'leak' sensitive information such as cryptographic keys. A particularly insidious form of leakage arises from the fact that hardware's power consumption over time is statistically associated with the data it processes and the instructions it executes. Supervised deep learning has emerged as a state-of-the-art tool for carrying out *power side-channel attacks*, which exploit this leakage to break cryptographic implementations by learning to map power consumption measurements recorded during encryption to the secret key used for that encryption. In this work, we seek instead to develop a principled deep learning framework for *defense* against such attacks by understanding the relative leakage due to power measurements recorded at different points in time. This information is invaluable to cryptographic hardware designers for understanding *why* their hardware leaks and how they can mitigate the leakage (e.g. by indicating that a particular section of code or electronic component is responsible for leakage and should be revised). Towards this end, we propose a novel deep learning algorithm by formulating an adversarial game played between a classifier trained to estimate the conditional distribution of a key given power measurements, and an 'obfuscator' which probabilistically erases individual power measurements and is trained to minimize the classifier-estimated log-likelihood of the correct key, subject to a penalty on erasure probability. We theoretically characterize the ideal output of our algorithm in terms of conditional mutual information quantities involving the key and individual power measurements. We then demonstrate the efficacy of our algorithm on real and synthetic datasets of power measurements from implementations of the AES cryptographic standard. Our code can be found (redacted).

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 12073