Global and Local Differential Privacy for Collaborative Bandits

Collaborative bandit learning has become an emerging focus for personalized recommendation. It leverages user dependence for joint model estimation and recommendation. As such online learning solutions directly learn from users, e.g., result clicks, they bring in new challenges in privacy protection. Despite the existence of recent studies about privacy in contextual bandit algorithms, how to efficiently protect user privacy in a collaborative bandit learning environment remains unknown. In this paper, we develop a general solution framework to achieve differential privacy in collaborative bandit algorithms, under the notion of global differential privacy and local differential privacy. The key idea is to inject noise in a bandit model's sufficient statistics (either on server side to achieve global differential privacy or client side to achieve local differential privacy) and calibrate the noise scale with respect to the structure of collaboration among users. We study two popularly used collaborative bandit algorithms to illustrate the application of our solution framework. Theoretical analysis proves our derived private algorithms reduce the added regret caused by privacy-preserving mechanism compared to its linear bandits counterparts, i.e., collaboration actually helps to achieve stronger privacy with the same amount of injected noise. We also empirically evaluate the algorithms on both synthetic and real-world datasets to demonstrate the trade-off between privacy and utility.