Trustworthy AI-Driven Dynamic Hybrid RIS: Joint Optimization and Reward Poisoning-Resilient Control in Cognitive MISO Networks

AI systems controlling wireless networks can be vulnerable to "reward poisoning" attacks, where malicious actors manipulate the feedback signals that train machine learning models, potentially degrading network performance. This paper presents the first systematic study of such attacks in reconfigurable intelligent surface (RIS) networks and proposes a defense mechanism using reward clipping and anomaly detection.

Dynamic switching between operational modes improves energy efficiency: Rather than operating wireless infrastructure components in a fixed way, the researchers' "hybrid RIS" system intelligently alternates between passive and active modes based on available energy, achieving better trade-offs between data throughput and power consumption compared to static approaches.

Real-world hardware limitations matter for AI deployment: The study accounts for practical impairments like hardware constraints and signal degradation that often get overlooked in theoretical models, demonstrating that trustworthy AI systems must be tested under realistic conditions rather than idealized scenarios.

Soft actor-critic (SAC) reinforcement learning outperformed other AI methods for this wireless optimization task, suggesting that algorithm selection significantly impacts performance in dynamic, continuous environments—an important consideration when deploying AI to critical infrastructure.

Security and robustness must be built into AI systems from the start: The paper's focus on defending against reward poisoning attacks highlights that responsible AI deployment requires proactive threat modeling and lightweight defense mechanisms, not just optimizing for accuracy.