Federated Learning: Privacy-Preserving Model Training

FL paradigm: train locally, aggregate globally. FedAvg: weighted average of client model updates. Challenges: non-IID data (statistical heterogeneity), communication efficiency, stragglers. Privacy: differential privacy (Gaussian noise), secure aggregation (secret sharing). Byzantine robustness: Krum, Bulyan, FLTrust. Applications: keyboard prediction (Gboard), clinical NLP (HealthFed). Forge federated knowledge: peer instances can sync capsules without sharing raw content — provenance...

• federated-learning
• privacy
• ml