PrivacyMAS: A Privacy-Preserving Multi-Agent System Framework

The proliferation of multi-agent systems in sensitive domains such as healthcare and finance necessitates robust privacy-preserving mechanisms that do not compromise utility or coordination efficiency. We present \textbf{PrivacyMAS}, a novel framework that addresses the fundamental trilemma between privacy preservation, utility maximization, and coordination scalability in multi-agent systems. Our key innovation is the ADAPT (Adaptive Differential privacy with Agent-based Privacy budgeT) algorithm, which dynamically adjusts privacy budgets based on environmental feedback, attack detection, and coordination quality metrics. Unlike existing static approaches that have been shown to be suboptimal in dynamic environments, ADAPT learns from the coordination environment to optimize the privacy-utility tradeoff while maintaining $O(\log n)$ communication complexity. We evaluate PrivacyMAS on two real-world datasets: medical diagnosis coordination using the DrBenjamin AI-Medical-Chatbot dataset comprising 10,000 clinical dialogues, and financial trading using the Sujet-Finance-Instruct-177k dataset containing 177,597 financial instructions. Our experiments demonstrate that ADAPT achieves up to a 19.6\% improvement in utility compared to static differential privacy baselines while maintaining equivalent privacy guarantees with $\epsilon \in [0.1, 2.0]$. Furthermore, our framework exhibits superior resistance to membership inference and attribute inference attacks, reducing attack success rates by up to 52.9\% in medical domains and 38.0\% in financial domains. These results establish PrivacyMAS as a practical solution for deploying privacy-preserving multi-agent systems at scale, addressing critical challenges identified in recent surveys of the field. Our full implementation including training pipelines, and analysis tools are available at https://github.com/anonymous-gihub99/Trilemma