Paper 2021/717

Explain Some Noise: Ablation Analysis for Deep Learning-based Physical Side-channel Analysis

Lichao Wu and Yoo-Seung Won and Dirmanto Jap and Guilherme Perin and Shivam Bhasin and Stjepan Picek

Abstract

Deep learning-based side-channel analysis represents a powerful option for profiling attacks on power and electromagnetic leakages as it breaks targets protected with countermeasures. While most of the papers report successful results, it is not difficult to find cases where deep learning works better or worse, especially concerning various countermeasures. Current approaches concentrate on various data augmentations or hyperparameter tuning options to make the attacks more powerful. At the same time, understanding what makes an attack difficult has received very little attention. This paper proposes a side-channel analysis methodology based on the ablation paradigm to explain how neural networks process countermeasures. Our results show that an ablation is a powerful tool as it allows to understand 1) in which layers various countermeasures are processed, 2) whether it is possible to use smaller neural network architectures without performance penalties, and 3) how to redesign neural networks to improve the attack performance when the results indicate that the target cannot be broken. By using the ablation-based approach, we manage to mount more powerful attacks or use simpler neural networks without any attack performance penalties. We hope this is just the first of the works in the direction of countermeasure explainability for deep learning-based side-channel analysis.