Adversarial Robustness of Deep-Learning-Based Near-Field Beam Prediction In XL-MIMO Systems

Deep learning is increasingly used in sixth generation (6G) extremely large-scale MIMO (XL-MIMO) systems to predict beams directly from channel observations to alleviate the overhead of exhaustive nearfield beam search.However, these channel-based observations also provide an attack surface to learned beam-selection models.In this paper, we consider the adversarial robustness of a deep learning-based nearfield beam predictor for XL-MIMO.We train a fully connected neural network to classify 45 near-field candidate beams generated for a simulated 8x8 planar array and evaluate its vulnerability under the Fast Gradient Sign Method (FGSM) and Projected Gradient Descent (PGD).The beam classification accuracy drops from 74.07%under clean condition to 28.70% under FGSM and 26.11% under PGD at perturbation budget of ε=0.05.Adversarial training improves the respective accuracies to 36.11% and 33.70% while maintaining comparable accuracy on clean data.Specifically, this classification-level degradation does not yield a proportional calculated communication-level impact: the beamforming gain ratio stays close to unity and the derived effective SNR, theoretical BPSK BER and Shannon-based spectral-efficiency estimate only change slightly, with a maximum reported effective-SNR loss of 0.014 dB.Results show that the near-field XL-MIMO beam predictors are vulnerable at the classification level under the input-space threat model considered, while the impact on communication-level is highly dependent on the spatial structure of the beam codebook.