Adversarial Pilot Injection Against AI-Based CSI Feedback in FDD Massive MIMO Systems

Artificial intelligence (AI)-based channel state information (CSI) feedback reduces overhead in frequency division duplexing (FDD) massive multiple-input multiple-output (MIMO) systems, yet its reconstructed CSI depends on the integrity of pilot acquisition. This paper identifies and evaluates the attack path of adversarial pilot injection against this feedback. A transmitted disturbance can damage the user-side estimation before encoding; correctly delivered quantized bits can still represent a contaminated channel estimate. We model this path through attacker propagation, linear minimum mean square error estimation, compression, quantization, and decoding under a total transmit-energy constraint. An adversarial perturbation generator (APG) implements the attack using channel statistics without current legitimate CSI or instantaneous attacker-channel observations at deployment. At a Rician attacker-channel factor of 10~dB and a transmit perturbation-to-signal ratio of $-4$~dB, APG produces a normalized mean-square error (NMSE) of 0.84~dB in base-station CSI reconstruction, compared with $-10.31$~dB for Gaussian interference. Estimator-targeted and reconstruction-targeted attacks yield different error orderings, and reconstruction damage also degrades beamforming. As a countermeasure, alternating adversarial training (AAT) adapts the feedback encoder and decoder to contaminated pilots. With condition-matched training and APGs reoptimized for the final receiver, AAT reduces NMSE by 2.40--3.79~dB across six representative channel--power conditions. The results demonstrate the importance of protecting pilot acquisition alongside quantized feedback delivery.