Private Multi-Task Learning - Formulation and Applications to Federated Learning

Learn about private multi-task learning and its applications in federated learning through this 18-minute research presentation from the Symposium on Foundations of Responsible Computing (FORC) 2022. Explore how Carnegie Mellon University researcher Shengyuan Hu addresses privacy challenges in machine learning applications across healthcare, finance, and IoT computing. Gain insights into task-level privacy formalization using joint differential privacy, a relaxation of differential privacy for mechanism design and distributed optimization. Discover a novel algorithm for mean-regularized multi-task learning that provides certifiable guarantees on both privacy and utility. Examine improved privacy/utility trade-offs compared to global baselines across federated learning benchmarks, with detailed coverage of introduction, federated learning setup, private multi-task learning, differential privacy concepts, joint differential privacy, and privacy guarantees.