Federated Learning (FL) inverts the traditional ML pipeline by sending the model to the data rather than centralizing data for training. Each participating client trains a local model copy and returns only weight updates or gradients — never raw data — to a central coordinator that aggregates them via algorithms like FedAvg. The architecture is already in production at Google, Apple, healthcare institutions, and financial systems. Key frameworks include TensorFlow Federated, Flower, PySyft, and NVIDIA FLARE. FL does not automatically guarantee privacy; known attack vectors like model inversion, membership inference, and gradient leakage require additional defenses such as differential privacy, secure aggregation, and homomorphic encryption. Real trade-offs include coordination overhead, non-IID data distribution across clients, and slower convergence, but FL offers a meaningfully smaller compliance surface and attack footprint for systems handling sensitive distributed data.
