Intelligent Clustering as a Means to Improve K-means Based Horizontal Attacks

Abstract

Machine learning approaches have a high potential for improving the success rate of side channel analysis attacks. In this paper we present horizontal side channel analysis attacks against three crypto-implementations suffering from different levels of leakage using a single power and a single electromagnetic trace. We show the effectivity of attacks using $k-means$ as analysis tool. In addition we introduce a new approach that we call intelligent clustering that enables attackers to select the start centroids in such a way that the ability of $k-means$ to extract the key bits is increased up to 38.56 % compared to $k-means$ starting the farthest neighbors centroids and up to 66.66 % compared to the mean correctness for $k-means$ starting with random centroids.