Reward Poisoning Attack on RIS-Aided Wireless Control Systems
Researchers propose a Disagreement-Guided Reward Poisoning (DGRP) attack targeting Soft Actor-Critic (SAC) agents in Cognitive Radio Networks (CRN) with Reconfigurable Intelligent Surfaces (RIS). The attack corrupts rewards when SAC dual critics disagree, especially in high-leverage, high-uncertainty states, distorting value estimations and leading to suboptimal actions. DGRP significantly reduces RIS performance gains and degrades transmission quality. The study explores key attack parameters. This is a technical paper on arXiv (2605.20037) about adversarial machine learning in wireless communications.
Key facts
- DGRP attack targets SAC agents in RIS-assisted CRN.
- Attack exploits disagreement between SAC dual critics.
- Corrupts rewards in high-leverage, high-uncertainty states.
- Degrades transmission quality and RIS performance gains.
- Published on arXiv with ID 2605.20037.
- Focuses on reward-poisoning in learning-based wireless control systems.
Entities
Institutions
- arXiv