Private Empirical Risk Minimization

Private empirical risk minimization (ERM) focuses on training machine learning models while preserving the privacy of the training data, typically using differential privacy mechanisms. Current research emphasizes improving the accuracy of privately trained models, particularly for high-dimensional data and non-convex loss functions, employing techniques like greedy coordinate descent, variance-reduced gradient descent, and output perturbation. These advancements aim to reduce the trade-off between privacy guarantees and model utility, impacting fields like healthcare and finance where data privacy is paramount.