Protecting Trained Models in Privacy-Preserving Federated Learning

This NIST/RTA blog post describes techniques for providing output privacy in privacy-preserving federated learning, emphasizing differential privacy (adding noise to model updates), adaptations of FedAvg for horizontally partitioned data, complications for vertically partitioned data (entity alignment and homomorphic/MPC solutions), and the accuracy–privacy tradeoff including pretraining and fine-tuning strategies; it is an explanatory piece and does not report a security incident.