Cryptology ePrint Archive: Report 2019/578

Deep Learning based Side Channel Attacks in Practice

Houssem Maghrebi

Abstract: A recent line of research has investigated a new profiling technique based on deep learning as an alternative to the well-known template attack. The advantage of this new profiling approach is twofold: $(1)$ the approximation of the information leakage by a multivariate Gaussian distribution is relaxed (leading to a more generic approach) and $(2)$ the pre-processing phases such as the traces realignment or the selection of the Points of Interest (PoI) are no longer mandatory, in some cases, to succeed the key recovery (leading to a less complex security evaluation roadmap). The related published works have demonstrated that Deep Learning based Side-Channel Attacks (DL-SCA) are very efficient when targeting cryptographic implementations protected with the common side-channel countermeasures such as masking, jitter and random delays insertion. In this paper, we assess the efficiency of this new profiling attack under different realistic and practical scenarios. First, we study the impact of the intrinsic characteristics of the manipulated data-set (\emph{i.e.} distance in time samples between the PoI, the dimensionality of the area of interest and the pre-processing of the data) on the robustness of the attack. We demonstrate that the deep learning techniques are sensitive to these parameters and we suggest some practical recommendations that can be followed to enhance the profiling and the key recovery phases. Second, we discuss the tolerance of DL-SCA with respect to a deviation from the idealized leakage models and provide a comparison with the well-known stochastic attack. Our results show that DL-SCA are still efficient in such a context. Then, we target a more complex masking scheme based on Shamir's secret sharing and prove that this new profiling approach is still performing well. Finally, we conduct a security evaluation of a batch of several combinations of side-channel protections using simulations and real traces captured on the ChipWhisperer board. The experimental results obtained confirm that DL-SCA are very efficient even when a cryptographic implementation combines several side-channel countermeasures.

Category / Keywords: implementation / Deep Learning based Side-Channel Attacks, Data Dimensionality, Data Scaling, Artificial Noise, Side-Channel Countermeasures, Shamir's Secret Sharing, Combination of Countermeasures

Date: received 27 May 2019

Contact author: houssem mag at gmail com

Available format(s): PDF | BibTeX Citation

Version: 20190528:064443 (All versions of this report)

Short URL: ia.cr/2019/578


[ Cryptology ePrint archive ]