Paper 2017/1110

The secrets of profiling for side-channel analysis: feature selection matters

Stjepan Picek and Annelie Heuser and Alan Jovic and Lejla Batina and Axel Legay

Abstract

Profiled side-channel attacks feature a number of steps one needs to take. One significant step, importance of which is sometimes ignored, is selection of the points of interest (features) within side-channel measurement traces. A large majority of the related works on profiling in side-channel analysis starts with an assumption that the features are somehow selected and distinct attack methods are compared in order to find the best approach for the key recovery. Contrary to this, in this work we concentrate on the feature selection step and show that if an optimal selection is done, most of the attack techniques perform well i.e., result in the key recovery. Consequently, in this paper, we investigate in details how more advanced feature selection techniques stemming from the machine learning domain can be used to improve the attack efficiency. To this end, we look into relevant aspects and we provide a systematic evaluation of machine learning methods of interest. Our results show that the so-called Hybrid feature selection methods perform with the best classification accuracy over a wide range of test scenar- ios and number of features selected. The experiments are performed on several real-world data sets containing software and hardware implemen- tations of AES, and even including the random delay countermeasure. We emphasize the L1 regularization technique, which consistently performed well and in many cases resulted in significantly higher accuracy than the second best technique. Further on, we consider even Principal Compo- nent Analysis (PCA) as a typical dimensionality reduction method and show that feature selection combined with the ML classification remains the method of choice (when confronted with PCA).