Multivariate-Non-profiled-Leakage-Detection

This repository contains the practical implementation of several multivariate leakage detection tests considered in the IACR-Tches 2026, Volume 2 paper titled Multivariate Leakage Detection. The research was conducted by the Cybersecurity research group at the University of Klagenfurt, Austria. In this repository, we primarily focused on a comparative study of different non-profiled multivariate detection methods, along with multiplicity corrections for existing univariate detection methods.

General Introduction

• The project contains the Python implementation of three multivariate leakage detection tests, namely, the distance covariance estimator (aka MV-dcov), the Diagonal $T$-test, and Hotelling's $T^2$. We have also considered Bonferroni's multiplicity correction techniques for four univariate leakage detection tests: Welch's $t$-test, the $\chi^2$-test, the mutual information-based $G$-test, and the distance correlation-based test of independence (aka dcor).
• We have analysed both the False positive rate and the True positive rate of the aforementioned tests via p-value computation and then computing the statistical power of the tests for a certain number of iterations.
  User guide and detailed instructions of our Python implementations are provided in Code folder.

Datasets

We have considered both simulated and practical case studies for our implementation.

• In simulation experiments, we have considered different linear leakage models, like hamming weight, hamming distance, weighted hamming weight, and one non-linear model (by considering the double permutation). Along with leakage models, we also consider the Gaussian and non-Gaussian additive noises. The multivariate leakage simulation is provided in testnbr_dist_1.py Python script.
• We have considered a practical case study for the side-channel traces from an unprotected implementation of PRESENT block cypher as provided by DL-LA. The download instructions for this dataset are available in PRESENT-RC

Acknowledgement

This project is supported in part by the Austrian Science Fund (FWF) 10.55776/F85 (SFB SpyCode) and by the EU Horizon project (enCrypton, grant agreement number 101079319).