MLP and CNN-based Classification of Points of Interest in Side-channel Attacks

Abstract

A trace contains sample points, some of which contain useful information that can be used to obtain a key in side-channel attacks. We used public datasets from the ANSSI SCA Database (ASCAD) as well as SM4 traces to learn whether a trace consisting of Points of Interest (POIs) have a positive effect using neural networks. Different methods were used on these datasets to choose POI or transform the traces into Principal Component Analysis (PCA) traces and forward-difference traces. The results show that two datasets are combined in different ways that improve the classification using neural networks. For example, for the ANSSI SCA database, PCA is a better approach to compress a 700-dimensional trace into a 100-dimensional trace. For SM4 traces, the amount of traces required can be reduced in side-channel attacks subsequent to forward-difference transformation.

Keywords

• Point of interest
• forward difference
• trace
• multi-layer perceptron
• convolutional neural network