Federated and Privacy-Preserving Machine Learning: Methods and Trade-offs

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Federated learning trains a shared model across many decentralized clients while keeping raw data on the device, which addresses data silos and tightening privacy regulation such as GDPR and HIPAA. This paper reviews the methods and the practical trade-offs of federated and privacy-preserving machine learning. The discussion begins with the FedAvg protocol and the distinction between cross-device settings, with millions of intermittent participants, and cross-silo settings, with a small set of reliable institutions. Statistical heterogeneity, where client data is non-IID, slows convergence and biases the global model; proximal correction in FedProx and control variates in SCAFFOLD reduce this client drift. Because the uplink is often the bottleneck, the survey covers gradient compression, quantization, and sparsification that cut communication cost by an order of magnitude. Sharing model updates is not by itself private: membership inference and gradient leakage attacks can recover sensitive information, so the paper examines differential privacy, secure aggregation, and homomorphic encryption as defenses, together with the accuracy cost each one carries. Personalization methods that adapt the global model to local distributions are summarized, followed by applications in healthcare, mobile keyboards, and finance. Illustrative experiments on a non-IID benchmark show SCAFFOLD reaching a target accuracy in fewer rounds than FedAvg, and a differential-privacy sweep quantifies the privacy-utility curve. The results are illustrative and do not describe a deployed production system.