Cryptology ePrint Archive: Report 2018/880

Spread: a new layer for profiled deep-learning side-channel attacks

Christophe Pfeifer and Patrick Haddad

Abstract: Recent publications, such as [10] and [13], exploit the advantages of deep-learning techniques in performing Side-Channel Attacks. One example of the Side-Channel community interest for such techniques is the release of the public ASCAD database, which provides power consumption traces of a masked 128-bit AES implementation, and is meant to be a common benchmark to compare deep-learning techniques performances. In this paper, we propose two ways of improving the effectiveness of such attacks. The first one is as new kind of layer for neural networks, called "Spread" layer, which is efficient at tackling side-channel attacks issues, since it reduces the number of layers required and speeds up the learning phase. Our second proposal is an efficient way to correct the neural network predictions, based on its confusion matrix. We have validated both methods on ASCAD database, and conclude that they reduce the number of traces required to succeed attacks. In this article, we show their effectiveness for first-order and second-order attacks.

Category / Keywords: implementation / Deep-learning,Side-channel attacks,Spread layer,ASCAD,Confusion matrix,Bayesian correction

Date: received 19 Sep 2018, last revised 12 Dec 2018

Contact author: patrick haddad at st com

Available format(s): PDF | BibTeX Citation

Note: Changes: • Formula 5 • Added information about how x is remapped to x' • minor grammar changes

Version: 20181212:095716 (All versions of this report)

Short URL: ia.cr/2018/880