Paper 2020/058

Learning when to stop: a mutual information approach to fight overfitting in profiled side-channel analysis

Guilherme Perin, Ileana Buhan, and Stjepan Picek

Abstract

Today, deep neural networks are a common choice for conducting the profiled side-channel analysis. Such techniques commonly do not require pre-processing, and yet, they can break targets protected with countermeasures. Unfortunately, it is not trivial to find neural network hyper-parameters that would result in such top-performing attacks. The hyper-parameter leading the training process is the number of epochs during which the training happens. If the training is too short, the network does not reach its full capacity, while if the training is too long, the network overfits, and is not able to generalize to unseen examples. Finding the right moment to stop the training process is particularly difficult for side-channel analysis as there are no clear connections between machine learning and side-channel metrics that govern the training and attack phases, respectively. In this paper, we tackle the problem of determining the correct epoch to stop the training in deep learning-based side-channel analysis. We explore how information is propagated through the hidden layers of a neural network, which allows us to monitor how training is evolving. We demonstrate that the amount of information, or, more precisely, mutual information transferred to the output layer, can be measured and used as a reference metric to determine the epoch at which the network offers optimal generalization. To validate the proposed methodology, we provide extensive experimental results that confirm the effectiveness of our metric for avoiding overfitting in the profiled side-channel analysis.