Revisiting a Methodology for Efficient CNN Architectures in Profiling Attacks

Lennert Wouters,Benedikt Gierlichs,Bart Preneel,Victor Arribas

doi:10.46586/tches.v2020.i3.147-168

Lennert Wouters, Benedikt Gierlichs + Show 2 more

Open Access

https://doi.org/10.46586/tches.v2020.i3.147-168

Copy DOI

Abstract

This work provides a critical review of the paper by Zaid et al. titled “Methodology for Efficient CNN Architectures in Profiling attacks”, which was published in TCHES Volume 2020, Issue 1. This work studies the design of CNN networks to perform side-channel analysis of multiple implementations of the AES for embedded devices. Based on the authors’ code and public data sets, we were able to cross-check their results and perform a thorough analysis. We correct multiple misconceptions by carefully inspecting different elements of the model architectures proposed by Zaid et al. First, by providing a better understanding on the internal workings of these models, we can trivially reduce their number of parameters on average by 52%, while maintaining a similar performance. Second, we demonstrate that the convolutional filter’s size is not strictly related to the amount of misalignment in the traces. Third, we show that increasing the filter size and the number of convolutions actually improves the performance of a network. Our work demonstrates once again that reproducibility and review are important pillars of academic research. Therefore, we provide the reader with an online Python notebook which allows to reproduce some of our experiments1 and additional example code is made available on Github.2

Highlights

Machine learning techniques gained a substantial interest in the recent literature on side-channel analysis [PSB+18, CCC+19, Tim19, ZBHV20, PHJ+19]
For an introduction to the topic of profiled side-channel attacks using neural networks based methods, we point to the work by Benadjila et al [PSB+18], who conduct a broad investigation on deep learning algorithms, discussing the relationship with classical template attacks
We explore the effects of commonly used preprocessing techniques in the context of deep learning based profiled side-channel attacks

Summary

Introduction

Machine learning techniques gained a substantial interest in the recent literature on side-channel analysis [PSB+18, CCC+19, Tim, ZBHV20, PHJ+19] This rapid increase in research volume in an emerging topic creates the need for a critical review of these results. Profiled side-channel attacks correspond to Time Series Classification (TSC) problem Several works regarding this topic are published and provide valuable information, from which we highlight the review manuscripts by Bagnall et al [BLB+17] and Fawaz et al [FFW+19]. They provide a thorough inspection of the most recent advances within the TSC domain, defining a solid base for successful machine learning based side-channel attacks. As our work is a thorough analysis of the work by Zaid et al we recommend the reader to read their manuscript [ZBHV20]

Objectives

Methods

Findings

Discussion

Conclusion