Logo image
Back
Camouflage Adversarial Attacks on Multi-agent Reinforcement Learning Systems
Journal article   Peer reviewed

Camouflage Adversarial Attacks on Multi-agent Reinforcement Learning Systems

Ziqing Lu, Guanlin Liu, Lifeng Lai and Weiyu Xu
IEEE transactions on signal processing, Vol.74, pp.589-604
2026
DOI: 10.1109/TSP.2025.3644869

View Online

Abstract

The multiple agent reinforcement learning systems (MARL) based on the Markov Game (MG) have emerged in many critical applications. To improve the robustness/defense of MARL systems against adversaries, studying various adversarial attacks on reinforcement learning systems is very important. Previous works on adversarial attacks considered some possible features to attack in the MG, such as action poisoning attacks, reward poisoning attacks, and state perception attacks. In this paper, we propose a new form of perception attack, called the camouflage attack in MARL systems. In the camouflage attack, the attackers change the appearances of some objects in the environment but without changing the actual objects; and the camouflaged appearances may look the same to all the targeted recipient (victim) agents. The camouflaged appearances can mislead the recipient agents to follow misguided policies. We evaluate the effect of camouflage attacks in two different scenarios: Camouflage attacks were performed during the learning (training-time attacks) and were performed during the test of agents' policies (test-time attacks). Our numerical and theoretical results show that camouflage attacks can rival the more conventional, but likely more difficult state perception attacks, by comparing their effect on reducing agents' global benefits. We also investigated cost-constrained camouflage attacks, compared them with cost-constrained state perception attacks, and showed how cost budgets affect attack performance numerically.
 Games Optimization Adversarial attacks Costs Indexes Markov decision process Markov decision processes Multi-agent reinforcement learning Perturbation methods Reinforcement learning Robots Signal processing algorithms Uncertainty

Details

Metrics

13 Record Views
Logo image