Adaptive Differential Privacy via Dynamic Thresholds

Abstract

This paper introduces a novel adaptive differential privacy (DP) mechanism designed to mitigate the limitations of traditional DP approaches, particularly concerning excessive noise addition in scenarios with small datasets. The core idea is to dynamically adjust the noise scale based on the sensitivity of the query and the current state of the data. This is achieved through the integration of a reinforcement learning (RL) agent that learns optimal thresholds for noise addition. The RL agent's objective is to minimize privacy loss while simultaneously preserving data utility. We demonstrate that this dynamic thresholding approach offers a significantly improved trade-off between privacy and utility compared to fixed noise scales commonly employed in DP. The theoretical analysis provides insights into the convergence properties of the RL agent and the overall privacy guarantees offered by the mechanism. The proposed method represents a crucial step towards more efficient and effective DP implementations, particularly in resource-constrained settings where minimizing noise is paramount.