Private Prediction via PAC Privacy
Researchers have introduced a new framework called PAC privacy for privatizing machine learning model outputs, which is particularly suited for models served via APIs. This approach contrasts with differential privacy by focusing on instance-based privacy and calibrating noise to empirical stability to control mutual information leakage. The new method includes an efficient, adaptive composition technique that allows for linear accumulation of mutual information even under adversarial querying, enabling high utility with minimal per-query budgets. AI
IMPACT Introduces a novel privacy framework for ML models, potentially enabling more secure API-based predictions with high utility.