Abstract
Recently, the potential of machine and consequently deep learning on side-channel analysis was discovered and confirmed even on protected cryptographic implementations. The success of those experiments has led to deep learning techniques becoming a mainstream component in side-channel leakage evaluations.
Conversely, recent work has shown that neural networks can be reversed engineered by the side-channel attacker, i.e., the adversary using physical leakage such as timing and EM. This makes neural nets an interesting target as in some applications such as security evaluation, HD maps for autonomous vehicles etc., as optimized network architectures are considered an IP.
This talk will survey some recent results on this two-fold connection between deep learning and side-channel analysis and discuss directions for future work.
