"Practical Federated Learning Systems: Privacy-Preserving Training Across Devices and Orgs"
Federated learning is no longer a research novelty—it's a production discipline where privacy, distributed systems reality, and machine learning optimization collide. This book is written for experienced ML engineers, platform engineers, and security-minded practitioners who need to ship federated training across millions of devices or multiple organizations, not just run a demo. It treats communication, client unreliability, and governance as first-class constraints, and it shows why "data stays local" is only the beginning of privacy.
You'll learn to formulate FL problems under non-IID data and partial participation, design end-to-end architectures (coordinator, client runtimes, transport), and make FedAvg work under real budgets and failure modes. The book then builds a rigorous threat model and translates it into deployable controls, diving deep into secure aggregation and differential privacy—including accounting over rounds and tuning workflows. It also covers robustness to malicious or faulty clients, communication efficiency techniques, and how to preserve observability without breaking privacy guarantees.
Throughout, the emphasis is on operational readiness: framework selection, decentralized data pipelines, privacy-preserving telemetry, staged rollouts and rollbacks, and cross-silo governance and incident response. Readers should be comfortable with modern ML training, distributed systems fundamentals, and basic applied cryptography and privacy terminology.
