Utility Aware Adaptive Privacy Budget Allocation for Streaming Multi-Agent Systems

Streaming Multi-Agent Systems (MAS) involve autonomous agents continuously sharing observations with a fusion center for coordination. Insuchscenarios, thedataflowismostlyconsistentallowing stronger privacy preservation at the cost of reduced utility. However, anomalies like accidents or emergencies require timely and accurate information sharing, necessitating lower privacy. Applying conventional differential privacy with a fixed and low 𝜖 value (high noise), risks obscuring the critical information. As each report consumes a part of the agents’ differential privacy budget, by sequential composition, the overall privacy reduces over time. Premature depletion of the privacy budget will lead to lower available budget for periods of high variability, reducing the accuracy and effectiveness of long term analytics. In this paper, we propose Adaptive Privacy Budget Allocation (APBA), a dynamic privacy budget allocation mechanism, leading to a tradeoff between utility and privacy. The allocation depends on (i) local signal uncertainty and (ii) the agent’s influence on the global estimate, concentrating privacyresourcesonthemostinformativetime-steps. Wetheoretically prove that APBA satisfies each agent’s global privacy budget under sequential composition while bounding estimation error. Experiments on real world sensor stream data demonstrate the efficiency and adaptability of APBA.

Authors

• Puspanjali Ghoshal Indian Institute of Technology Guwahati
• Ashok Singh Sairam Indian Institute of Technology Guwahati

Keywords

• Multi-Agent Systems
• Streaming Data
• Differential Privacy
• Adaptive Privacy Budget Allocation