Privacy Leaks In Multi-Agent Mediators

Mechanism design typically assumes the presence of a principal who wields significant influence over the regulations governing interactions among agents to attain a favorable result. Nevertheless, in various practical scenarios, such a level of control is often unattainable. For instance, a market designer might be constrained by the pre-existing conditions of a market, making it impossible for her to ensure that players engage with her designed mechanism or adhere to the proposed outcomes, even if they do participate. Similarly, in many situations, the option of making transfers may be off the table. To distinguish these principals with limited authority from the conventional principals explored in academic literature, we refer to them as "mediators." Recent years have seen important advances in the theory of mediators for multi-agent systems, as well as the learning of mediator policies using deep reinforcement learning. However, by proposing actions for agents to take, mediators may give away private information about the agents' private preferences, thus allowing for opportunities for other agents to exploit them. In this project, we will be conducting a detailed theoretical and empirical assessment of differential privacy violations in multi-agent mediation settings, from idealised repeated general-sum games, to larger-scale market clearing settings.