Controlling the Deep Learning-Based Side-Channel Analysis: A Way to Leverage from Heuristics

摘要

Deep neural networks have become the state-of-the-art method when a profiled side-channel analysis is performed. Their popularity is mostly due to neural nets overcoming some of the drawbacks of "classical" side-channel attacks, such as the need for feature selection or waveform synchronization, in addition to their capability to bypass certain countermeasures like random delays. To design and tune a neural network for side-channel analysis systematically is a complicated task. There exist hyperparameter tuning techniques which can be used in the side-channel analysis context, like Grid Search, but they are not optimal since they usually rely on specific machine learning metrics that cannot be directly linked to e.g. the success of the attack. We propose a customized version of an existing statistical methodology called Six Sigma for optimizing the deep learning-based side-channel analysis process. We demonstrate the proposed methodology by successfully attacking a masked software implementation of AES.