Evaluating Robustness of Cooperative MARL: A Model-based Approach

Abstract

In recent years, a proliferation of methods were developed for cooperative multi-agent reinforcement learning (c-MARL). However, the robustness of c-MARL agents against adversarial attacks has been rarely explored. In this paper, we propose to evaluate the robustness of c-MARL agents via a model-based approach. Our proposed formulation can craft stronger adversarial state perturbations of c-MARL agents(s) to lower total team rewards more than existing model-free approaches. In addition, we propose the first victim-agent selection strategy which allows us to develop even stronger adversarial attack. Numerical experiments on multi-agent MuJoCo benchmarks illustrate the advantage of our approach over other baselines. The proposed model-based attack consistently outperforms other baselines in all tested environments.

Certified Policy Smoothing for Cooperative Multi-Agent Reinforcement Learning

“…Lin et al (2020) proposed to train a policy network to find a wrong action that the victim agent is expected to take and set it as the targeted adversarial example. Pham et al (2022) then proposed to craft a stronger adversary by using a model-based approach.…”