This paper presents a novel approach to achieving differential privacy (DP). While conventional DP is implemented through noise injection, this study explores a method to achieve DP "for free" by leveraging the inherent randomness of existing algorithms. Specifically, by adding KL-regularization to the learning objective, we demonstrate that differential privacy can be achieved in three decision-making problems: multi-armed bandit, linear context bandit, and Reinforcement Learning from Human Feedback (RLHF). This novel approach guarantees privacy without noise injection while preserving the benefits of regularization, which contributes to improved performance.
