When Privacy Meets Partial Information - Privacy-Utility Trade-offs in Bandits

Attend a research seminar exploring the intersection of differential privacy and bandit algorithms, where Prof. Debabrata Basu from INRIA France examines how privacy constraints affect sequential learning systems. Discover how bandits serve as fundamental models for decision-making under uncertainty, where algorithms must balance exploration and exploitation while learning about decision utilities through limited interactions. Learn about the critical challenges of implementing differential privacy in data-sensitive applications like adaptive clinical trials, hyperparameter tuning, and recommender systems. Explore three key research questions: how to properly define differential privacy in progressive bandit settings, how privacy requirements change the fundamental difficulty of bandit problems, and how to modify existing algorithms to achieve both privacy guarantees and optimal performance. Understand new information-theoretic quantities and generic algorithms that demonstrate how ε-differential privacy can often be achieved with minimal performance cost in bandit scenarios. Gain insights into the speaker's broader research on developing theoretically-grounded responsible AI systems, including work on robust, private, fair, and explainable algorithms for online learning and reinforcement learning problems.