Cryptology ePrint Archive: Report 2018/004

On the Performance of Deep Learning for Side-channel Analysis

Stjepan Picek and Ioannis Petros Samiotis and Annelie Heuser and Jaehun Kim and Shivam Bhasin and Axel Legay

Abstract: Profiled side-channel attacks represent the most powerful category of side-channel attacks. There we have a number of methods promising to work well in a number of different scenarios. Still, the area is constantly improving: we started with template attack and then went into different machine learning techniques that outperformed template attack in certain settings. Recently, deep learning techniques brought promise of even better results. In this paper, we ask a question whether deep learning is actually better than machine learning, and if yes, in what situations exactly. To this end, we compare several machine learning techniques and a well-known deep learning technique -- convolutional neural networks in a number of scenarios. Our results point that convolutional neural networks indeed outperforms machine learning in several scenarios but that often there is no compelling reason to use such a complex technique. In fact, if comparing techniques without extra steps like pre-processing, we see obvious advantage for deep learning only when the level of noise is small, the number of measurements is high, and the number of features is high. All other tested situations actually show that machine learning, for a significantly lower computational cost, performs the same or even better. Finally, we conduct a small experiment that opens the question whether convolutional neural networks are actually the best choice in SCA context.

Category / Keywords: secret-key cryptography / Side-channel analysis, Machine learning, Deep learning, Convolutional Neural Networks, SCANet

Date: received 1 Jan 2018

Contact author: picek stjepan at gmail com

Available format(s): PDF | BibTeX Citation

Version: 20180102:174940 (All versions of this report)

Short URL: ia.cr/2018/004

Discussion forum: Show discussion | Start new discussion