User Level Differential Privacy

User-level differential privacy (DP) aims to protect the privacy of all data contributed by an individual user in machine learning, a stronger guarantee than protecting individual data points. Current research focuses on developing and analyzing algorithms for achieving user-level DP in various settings, including federated learning and large language model fine-tuning, often employing techniques like per-user gradient clipping and adaptive noise injection. This area is crucial for ensuring robust privacy in collaborative machine learning and data analysis, particularly when dealing with sensitive personal information, and is driving the development of new privacy-preserving algorithms and theoretical frameworks.