Cryptology ePrint Archive: Report 2018/1196

Gradient Visualization for General Characterization in Profiling Attacks

Lo´c Masure and CÚcile Dumas and Emmanuel Prouff

Abstract: Past few years have seen the emergence of Machine Learning and Deep Learning algorithms as promising tools for profiling attacks, especially Convolutional Neural Networks (CNN). The latters have indeed been shown to overcome countermeasures such as de-synchronization or masking. However, CNNs are not widely used yet and Gaussian Templates are usually preferred. Though their efficiency is highly impacted by the countermeasures previously mentioned, their relevance relies on theoretical and physical justifications fairly recognized among the Side Channel community. Instead, the efficiency of CNNs still raises a certain scepticism as they act as a black-box tool. This scepticism is not specific to the Side Channel Analysis context: understanding to what extent CNNs would be so powerful and how they learn to recognize discriminative features for classification problems is still an open problem. Some methods have been proposed by the computer vision community, without satisfying performance in this field. However, methods based on Sensitivity Analysis particularly fit our problem. We propose to apply one of them called Gradient Visualization that uses the derivatives of a CNN model with respect to an input trace in order to accurately identify temporal moments where sensitive information leaks. In this paper, we theoretically show that this method may be used to efficiently localize Points of Interest in the SCA context. The efficiency of the proposed method does not depend on the particular countermeasure that may be applied to the measured traces as long as the profiled CNN can still learn in presence of such difficulties. In addition, the characterization can be made for each trace individually. We verified the soundness of our proposed method on simulated data and on experimental traces from a public Side Channel database. Eventually we empirically show that Sensitivity Analysis is at least as well as state-of-the-art characterization methods, in presence (or not) of countermeasures.

Category / Keywords: implementation / Side Channel Analysis, Profiling Attacks, Deep Learning, Points of Interest, Characterization

Date: received 11 Dec 2018, last revised 8 Jan 2019

Contact author: loic masure at cea fr

Available format(s): PDF | BibTeX Citation

Note: Fig.4 (right) page 12 modified (from pdf to png) to avoid printing issues

Version: 20190108:082936 (All versions of this report)

Short URL: ia.cr/2018/1196


[ Cryptology ePrint archive ]