Simplified Single-Trace Side-Channel Attacks on Elliptic Curve Scalar Multiplication using Fully Convolutional Networks

We aim to simplify the worst-case horizontal attack on scalar multiplication published at CHES 2017 [1] by making use of deep learning techniques, and to automate the critical steps of this previous work, namely the information identification, information extraction and information combination steps. For this purpose, we gradually increase the number of automated steps, targeting a very challenging assembly-level regular Montgomery ladder scalar multiplication implementation on a BeagleBone Black (BBB) Board running at 1 GHz. Our results demonstrate that the latter two steps can be simplified using deep learning techniques and lead to similar results as previous work. By contrast, the first step still requires some additional engineering. By showing that points of interest (POIs) selection can play an important (sometimes necessary) role for attacking asymmetric cryptography algorithms using deep learning techniques, we bring a more contrasted view on the advantage and limitations of such techniques. To the best of our knowledge, this is the first public report of deep learning based attack on ECC implementations. Besides, we propose the use of Fully Convolutional Networks as an alternative (deep) learning tool for side-channel analysis.