Analysis of Nonparametric Estimation Methods for Mutual Information Analysis

Abstract

Mutual Information Analysis (MIA) is a side-channel attack introduced recently. It uses mutual information, a known information theory notion, as a side-channel distinguisher. Most previous attacks use parametric statistical tests and the attacker assumes that the distribution family of the targeted side-channel leakage information is known. On the contrary, MIA is a generic attack that assumes the least possible about the underlying hardware specifications. For example, an attacker should not have to guess a linear power model and combine it with a parametric test, like the Pearson correlation factor. Mutual information is considered to be very powerful however it is difficult to estimate. Results of MIA can therefore be unreliable and even bias. Several efficient parametric estimators of mutual information are proposed in the literature. They are obviously very efficient when the distribution is correctly guessed. However, we loose the original goal of MIA which is to assume the least possible about the attacked devices. Hence, nonparametric estimators of mutual information should be considered in more details and, in particular, their efficiency in the side-channel context. We review some of the most powerful nonparametric methods and compare their performance with state-of-the-art side-channel distinguishers.