Preserving Inference Privacy in Multi-Agent Systems

Abstract

A multi-agent system (MAS) involves autonomous agents observing and communicating with a fusion center, which aggregates the data to infer one or more system parameters with the highest possible accuracy. However, besides the system parameters, an honest-but-curious fusion center may infer secondary sensitive information or create a profile of an agent without its consent. These attacks constitute an inference privacy breach. This thesis aims to develop robust and efficient mechanisms to prevent such breaches in MASs. Since any entity other than the data owner can be a potential adversary, preserving inference privacy requires decentralized data sanitization. Moreover, the limited computational resources of autonomous agents necessitate lightweight sanitization mechanisms. As data sanitization inevitably modifies the raw observations and reduces utility, privacy mechanisms must preserve sufficient utility for accurate distributed inference at the fusion center.