Mutual information neural estimation: algorithm optimization and application to wireless security

Details

Supervisors

Médard, Muriel ; Oestges, Claude ; Vandendorpe, Luc ; Yeh, Chia-Yi

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] : ingénieur civil électricien, à finalité spécialisée

Abstract

Mutual information is a fundamental concept in information theory that measures the dependency between two random variables. It has found application in many fields such as machine learning, signal processing, neuroscience, etc. The estimation of mutual information is however a challenging problem, especially when the probability distributions of the random variables are unknown. In this work, we present a new mutual information estimator called Mutual Information Neural Estimation (MINE) that was introduced in a paper published in 2018 by Belghazi et al. The latter is based on the use of neural networks and has been shown to outperform other mutual information estimators in many cases. We also present some optimizations we added to an already existing implementation of MINE and how they improve its performances and user-friendliness. Finally, we present an application of MINE to evaluate the security of a new wireless communication scheme as presented in a 2022 paper of Cohen et al. We also develop a statistical test in order to determine whether the system is secure or not.