Paper 2020/1293

I Choose You: Automated Hyperparameter Tuning for Deep Learning-based Side-channel Analysis

Lichao Wu, Guilherme Perin, and Stjepan Picek

Abstract

Deep learning-based SCA represents a powerful option for profiling side-channel analysis. Numerous results in the last few years indicate neural networks can break targets protected with countermeasures even with a relatively small number of attack traces. Intuitively, the more powerful neural network architecture we require, the more effort we need to spend in its hyperparameter tuning. Current results commonly use random search and reach good performance. Yet, we remain with the question of how good are such architectures if compared with the architectures that are carefully designed by following a specific methodology. Unfortunately, the works considering methodologies are sparse and difficult to ease without prior knowledge about the target. This work proposes an automated way for deep learning hyperparameter tuning that is based on Bayesian Optimization. We build a custom framework denoted as AutoSCA that supports both machine learning and side-channel metrics. Our experimental analysis shows that Bayesian optimization performs well regardless of the dataset, leakage model, or neural network type. What is more, we find a number of neural network architectures outperforming state-of-the-art attacks. Finally, we note that random search, despite being considered not particularly powerful, manages to reach top performance for a number of considered settings. We postulate this happens since the datasets are relatively easy to break, and there are many neural network architectures reaching top performance.