Paper 2021/072

Toward Practical Autoencoder-based Side-Channel Analysis Evaluations

Servio Paguada, Lejla Batina, and Igor Armendariz

Abstract

This paper introduces a practical evaluation procedure based on autoencoders for profiled side-channel analysis evaluations. An autoencoder is a learning model able to pre-process leakage traces improving in this way the guessing entropy. Nevertheless, this learning model's design should aim to code the leakage distribution to avoid relevant information being removed. For this reason, we propose an autoencoder built upon dilated convolutions. When using these learning models, the evaluation produces new assets, e.g., new versions of the dataset and new models based on learning algorithms. Our procedure comprises meaningful metrics and visualization techniques, namely signal-to-noise ratio and weight visualization, to evaluate those assets' effectiveness. After applying our procedure and our new autoencoder architecture to the ASCAD random key database, our results outperform state-of-the-art.