New adaptive attack targets wireless control systems using critic disagreement

Researchers have developed a new adaptive reward poisoning attack called Disagreement-Guided Reward Poisoning (DGRP) that targets learning-based wireless control systems. This attack specifically exploits disagreements between dual critics in Soft Actor-Critic agents, particularly in high-uncertainty states. By corrupting reward signals, DGRP distorts value estimations and leads the agent's policy toward suboptimal actions, significantly degrading performance in RIS-assisted networks. AI

IMPACT Highlights a new vulnerability in reinforcement learning agents, necessitating more robust security measures for AI-controlled systems.

RANK_REASON Academic paper detailing a novel attack methodology on a specific AI application. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• Deep Reinforcement Learning (DRL)
• Disagreement-Guided Reward Poisoning (DGRP)
• Reconfigurable Intelligent Surfaces (RIS)
• Soft Actor-Critic (SAC)

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →