Multi-Task Differential Privacy under Distribution Skew

For multi-task learning under user-level differential privacy, the researchers propose a general algorithm based on adaptive reweighting of empirical losses to address distribution skew across tasks. The work systematically analyzes the optimal allocation of user privacy budget across tasks, and can quantifiably improve excess empirical risk when task distributions are skewed and some tasks have scarce data. On recommendation problems with a long tail of small tasks, the method significantly improves utility, reaching SOTA on two standard benchmarks.