Paper 2019/860

Machine learning and side channel analysis in a CTF competition

Yongbo Hu, Yeyang Zheng, Pengwei Feng, Lirui Liu, Chen Zhang, Aron Gohr, Sven Jacob, Werner Schindler, Ileana Buhan, and Karim Tobich

Abstract

Machine learning is nowadays supplanting or extending human expertise in many domains ranging from board games to text translation. Correspondingly, the use of such tools is also on the rise in computer security. Alongside CHES 2018, a side channel challenge was organised under the theme of ’Deep Learning vs Classical SCA’ to test whether Deep Learning is presently widely used in the SCA community and whether it yields competitive results. The competition had 58 participants, it ran for three months, and a quantity of 35GB of data was used as a test sample. This paper presents the solutions of the teams that captured a flag and then discusses the results. While deep learning was used by neither team, other machine learning methods turned out to be very useful. The first contribution is a snapshot in time of the expertise in the community and shows a clear bias towards classic SCA. The second contribution is the presentation of novel techniques for key extraction for the challenges proposed and a reference for a black-box evaluation of crypto primitives by experts in the field. The third contribution is a baseline which can be used to further improve upon. Based on the results of this competition, we conclude that human expertise remains very important in the design of successful SCA attacks and machine learning can be a useful tool. Section 2,3 and 4 of this report have been directly contributed by the winning teams; as a consequence, section 3 is essentially identical to the previous eprint https://eprint.iacr.org/2019/094/20190131:230649 [8] authored by A. Gohr, S. Jacob and W. Schindler.