Automated Differential Privacy Parameter Tuning via Reinforcement Learning with Dynamic Hyperparameter Adaptation

**Abstract:** This paper presents a novel framework for automating the tuning of differential privacy (DP) parameters in cryptographic algorithms, specifically focusing on enhancing computational efficiency. Current manual parameter selection methods are time-consuming and often lead to suboptimal privacy-utility trade-offs. Our method, Reinforcement Learning with Dynamic Hyperparameter Adaptation (RL-DHA), leverages reinforcement learning (RL) to learn optimal DP parameter settings, dynamically adapting the hyperparameter space during training. Leveraging recent advances in secure multi-party computation (SMPC), our system achieves a 3-5x performance increase while maintaining comparable privacy levels, poised for immediate commercial application in federated learning and secure data analytics.