On the Performance of Convolutional Neural Networks for Side-Channel Analysis

Abstract

In this paper, we ask a question whether convolutional neural networks are more suitable for SCA scenarios than some other machine learning techniques, and if yes, in what situations. Our results point that convolutional neural networks indeed outperforms machine learning in several scenarios when considering accuracy. Still, often there is no compelling reason to use such a complex technique. In fact, if comparing techniques without extra steps like preprocessing, we see an obvious advantage for convolutional neural networks only when the level of noise is small, and the number of measurements and features is high. The other tested settings show that simpler machine learning techniques, for a significantly lower computational cost, perform similar or even better. The experiments with the guessing entropy metric indicate that simpler methods like Random forest or XGBoost perform better than convolutional neural networks for the datasets we investigated. Finally, we conduct a small experiment that opens the question whether convolutional neural networks are actually the best choice in side-channel analysis context since there seems to be no advantage in preserving the topology of measurements.

Recent advances in deep learning‐based side‐channel analysis

“…After the CNN‐based DLSCA technique was proposed, studies were conducted to compare the performance of CNN‐based SCA and other profiling SCA . The ASCAD public dataset was proposed to compare the performances of DLSCA and other profiling SCA methods objectively .…”

A Comprehensive Tutorial and Survey of Applications of Deep Learning for Cyber Security

“…The second approach is to efficiently correct the model predictions based on confusion matrix. In [309], the authors studies several ML and DL models for side-channel attacks. They have found that CNN performs better when the noise level is low and number of features are high.…”

A Systematic Evaluation of Profiling Through Focused Feature Selection

“…Such techniques either completely transform the features or use them in a manner too complicated to be understood by human experts. Moreover, deep learning could often have no performance advantage against "standard" ML if the number of measurements is not very large [6]. Note that, in this article, we do not consider comparisons with deep learning techniques, but we refer interested readers to [7]- [9].…”