Adversarial Attacks and Defenses for Vision-based mmWave Beam Prediction in 6G UAV Networks

In the dynamic environment of 6G Unmanned Aerial Vehicle (UAV) networks, vision-based beam prediction is a key enabler for advanced AI-based beam management to realize highspeed millimeter-wave (mmWave) communications. However, the reliance on deep neural networks (DNNs) for visual inference introduces significant security vulnerabilities, thus posing a critical threat to the overall 6G network management and operational integrity. This paper provides, to our knowledge, the first systematic study of white-box adversarial attacks on vision-based UAV beam prediction models. We demonstrate this vulnerability by proposing two practical and label-free universal adversarial perturbations (LF-UAPs) for DNN-based beam prediction. The first is an untargeted attack that maximizes divergence from the clean prediction, and the second is a targeted attack that steers beams to the farthest angular extreme. Both attacks are shown to effectively misdirect beam selection and degrade link quality in real-time. To counter this threat, we propose CLARA, a supervised contrastive learning framework that improves adversarial resilience while preserving clean performance. Unlike traditional defenses relying solely on adversarial training or gradient masking, CLARA aligns semantic representations of clean and adversarial inputs from the same class. By combining class-oriented representation learning with adversarial training, it enforces robust intra-class feature consistency across clean and perturbed views, thereby reducing the decision boundary's sensitivity to imperceptible image-level changes. Extensive experiments on a real-world 6G dataset show that standard models are easily compromised and baseline defenses fail abruptly, whereas the proposed framework successfully preserves signal integrity and significantly reduces power loss under unseen and stronger perturbations.