Multi-agent Perspective of Fake Feedback Attacks on Stochastic Multi-armed Bandits

The problem of false feedback attacks on stochastic Multi-Armed Bandits (MAB) algorithms is the focus of this article from the perspective of Multi-Agent Systems (MAS). We present the roles, beliefs, interactions, and goals of the agents involved in false feedback attacks according to performance metrics to evaluate the MAB algorithms’ performance. Three types of attacks illustrate the problem: constant, adaptive, and Jun’s adversarial attacks relaxed version, specifically built to exploit vulnerabilities in the e-Greedy and UCB1 algorithms. Experiments exploiting the vulnerabilities inherent in stochastic MAB algorithms present results revealing the useful characteristics for those who design MAS requiring defenses against false feedback attacks.