On the Feasibility of Real-Time Adversarial Attacks in 6G V2X: Pre-Computed Threats Against UC-CFmMIMO Resource Allocation

Machine learning is increasingly integrated into sixth-generation vehicle-to-everything systems for tasks requiring timely and accurate channel state information. Our previous work showed that such ML-based predictors are vulnerable to gradient-driven adversarial attacks. However, the high mobility of vehicle-to-everything scenarios yields channel coherence times (Tc) often below one millisecond, raising fundamental questions about the practical feasibility of these attacks. Widely used iterative methods, such as the Fast Gradient Sign Method and Projected Gradient Descent, may fail to meet these stringent real-time constraints, resulting in perturbations that target outdated channel states. This limitation motivates the investigation of more realistic adversarial threats that do not rely on per-sample gradient computation. Accordingly, we study in this paper pre-computed adversarial mechanisms, namely Universal Adversarial Perturbations and localized radio-frequency perturbations, which are generated offline and deployed instantaneously. Results show that these attacks remain stealthy and highly disruptive while satisfying submillisecond latency requirements. Despite being pre-computed, they consistently degrade channel gain prediction by exploiting distribution-level vulnerabilities and generalize effectively across channel realizations in high-mobility vehicle-to-everything environments, posing a credible threat to ML-driven physical-layer processing in future 6G systems.