Machine Learning Vulnerabilities in 6G: Adversarial Attacks and Their Impact on Channel Gain Prediction and Resource Allocation in UC-CFmMIMO
Résumé
Machine learning (ML) models integrated into physical-layer functions in wireless systems are increasingly vulnerable to adversarial attacks. Although prior research has investigated such threats in conventional massive MIMO architectures, the security risks in future 6G topologies, particularly user-centric cell-free massive MIMO (UC-CFmMIMO) deployed in vehicular environments, remain largely unexplored. These architectures depend heavily on frequency-domain channel gain estimation, which opens new attack surfaces. In this work, we present a black-box adversarial framework tailored to UC-CFmMIMO networks operating in dynamic vehicular environments. The attacker passively collects RF data to train a surrogate model and crafts perturbations using the FGSM attack. A local anomaly detector is integrated to assess stealth prior to uplink injection via pilot contamination. Our method significantly disrupts channel gain estimation and subband allocation, while requiring no access to the target model's internals. These results underscore emerging vulnerabilities in ML-enabled wireless systems and highlight the need for robust, context-aware defenses.
| Origine | Fichiers produits par l'(les) auteur(s) |
|---|---|
| Licence |