A VAE Approach to Low-Power, Multi-Adversary Pilot Contamination Attacks in CF-mMIMO

This work investigates the vulnerability of user-centric cell-free massive MIMO (CF-mMIMO) systems to multi-adversary pilot contamination attacks (PCAs). We propose a lightweight Variational Autoencoder (VAE)-based detection framework that effectively identifies corrupted pilots using only the covariance matrices of the received pilot signals after correlation. Although distributed attacks can replicate the received power profile of legitimate user equipments (UEs), they produce a "whitening" effect on the spatial correlation structure. This feature manifests as an anomaly in the Kullback-Leibler (KL) divergence of the approximate posterior distribution, providing a reliable indicator of corrupted pilot transmissions. Simulation results confirm that the proposed approach significantly outperforms conventional norm-based detectors, achieving robust performance even in challenging low-power attack scenarios. [Code can be accessed in: https://github.com/DarielPereira/VAE-PilotContaminationAttack-CF_mMIMO ]