Randomized Least Squares Value Iteration itself is Joint Differentially Private
Researchers have developed a new privacy analysis for reinforcement learning algorithms, specifically focusing on Randomized Least Squares Value Iteration (RLSVI). Their work demonstrates how the inherent noise used for exploration in RLSVI can simultaneously offer differential privacy protection. The study provides a mathematical characterization of this privacy guarantee, showing that RLSVI is $(\varepsilon(\delta),\delta)$-joint differentially private in tabular Markov Decision Processes. AI
IMPACT This research could enable the use of reinforcement learning in sensitive domains by providing a formal privacy guarantee.