The deep learning-based side-channel analysis gave some of the most prominent side-channel attacks against protected targets in the past few years. To this end, the research community’s focus has been on creating the following: (1) powerful multilayer perceptron or convolutional neural network architectures and (2) (if possible) minimal multilayer perceptron or convolutional neural network architectures. Currently, we see that, computationally intensive hyperparameter tuning methods (e.g., Bayesian optimization or reinforcement learning) provide the best results. However, as targets with more complex countermeasures become available, these minimal architectures may be insufficient, and we will require novel deep learning approaches.This work explores how residual neural networks (ResNets) perform in side-channel analysis and how to construct deeper ResNets capable of working with larger input sizes and requiring minimal tuning. The resulting architectures, obtained by following our guidelines, are significantly deeper than commonly seen in side-channel analysis, require minimal hyperparameter tuning for specific datasets, and offer competitive performance with state-of-the-art methods across several datasets. Additionally, the results indicate that ResNets work especially well when the number of profiling traces and features in a trace is large.

The deep learning-based profiling attacks have received significant attention for their potential against masking-protected devices. Currently, additional capabilities like exploiting only a segment of the side-channel traces or having knowledge of the specific countermeasure scheme have been granted to attackers during the profiling phase. In case either capability is removed, a practical profiling attack faces great difficulty and complexity. To address this challenge, we propose an efficient and scheme-agnostic Leakage Distillation-based Profiling Attack (LD-PA). By distilling univariate leakage from a reference, we can train an encoder that extracts multivariate leakage from raw traces and transforms it into an effective representation (transitional leakage). An indirect connection between multivariate leakage and the target variable is established by bridging through the transitional leakage, thereby facilitating the inference of leaked values. Remarkably, LD-PA achieves successful attacks on multiple public datasets using a simple multilayer perceptron (MLP) without necessitating an exhaustive hyperparameter search, while its performance is competitive with state-of-the-art methods. Simultaneously, we delve into the nature of transitional leakage, confirming the existence of combined leakage. This, in turn, validates that the guidance from univariate leakage references aids in the combination of multivariate leakage. Besides that, each component of the multivariate leakage is extracted and stacked in a highly aligned manner. Moreover, we explored several factors impacting LD-PA performance, covering scenarios with limited profiling traces, noisy references, alternative references, and hyperparameter tuning.

Deep neural networks have proven effective for second-order profiling side-channel attacks, even in a black-box setting with no prior knowledge of masks and implementation details. While such attacks have been successful, no explanations were provided for understanding why a variety of deep neural networks can (or cannot) learn high-order leakages and what the limitations are. In other words, we lack the explainability on neural network layers combining (or not) unknown and random secret shares, which is a necessary step to defeat, e.g., Boolean masking countermeasures. In this paper, we use information-theoretic metrics to explain the internal activities of deep neural network layers. We propose a novel methodology for the explainability of deep learning-based profiling side-channel analysis (denoted ExDL-SCA) to understand the processing of secret masks. Inspired by the Information Bottleneck theory, our explainability methodology uses perceived information to explain and detect the different phenomena that occur in deep neural networks, such as fitting, compression, and generalization. We provide experimental results on masked AES datasets showing what relevant features deep neural networks use, and where in the networks relevant features are learned and irrelevant features are compressed. Using our method, evaluators can determine what secret masks are being exploited by the network, which allows for more detailed feedback on the implementations. This paper opens new perspectives for understanding the role of different neural network layers in profiling side-channel attacks