Federated Machine Learning for Privacy-preserving Data Analytics in Distributed Systems

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).

By keeping the data separate, federated learning (FL) enables multiple data owners, such as hospitals, edge devices, or financial institutions, to jointly develop a shared machine-learning model, which is of particular relevance in applications of distributed systems that have to be operated under privacy regulation and underlying narrow bandwidth restrictions. We implement both two basic FL mechanisms (Federated Averaging, FedAvg, and a mechanism at the client-side that applies differential-privacy, DP) from the ground-up in NumPy, and test both in a uniform experimental protocol on two different distributed-systems scenarios: a network of 10 simulated healthcare data-silos with the real Wisconsin Diagnostic Breast Cancer benchmark (n = 569) and a network of 30 simulated edge/IoT devices with the real UCI handwritten-digits benchmark (n = 1,797). Client data was partitioned both in the near-IID and Dirichlet-skewed NON-IID (Hsu, Qi, & Brown, 2019) regime. In both communication rounds (40 in healthcare and 30 in edge/IoT), federated training achieved 97.7% and 96.7% test accuracy, respectively, (plus or minus 0.6 and 0.4 percentage points, respectively, compared to an equivalent centralized model trained on a “pooled” dataset), but the edge/IoT task suffered significant, monotonic loss of accuracy as the degree of client heterogeneity increased (96.7% for near-IID to 95.4% for the most skewed partition, whereas there was nearly no loss in the healthcare task for which the client models were simpler and near-linearly-separable). The systematic sweep over the differential-privacy noise multiplier confirmed that there is a significantly higher privacy-utility tradeoff for the harder high dimensional edge/IoT task than for the healthcare task (privacy-accuracy tradeoff varied from 96.3% accuracy to 51.1% accuracy as noise varied in the former and 97.7% accuracy to 78.4% accuracy as noise varied in the latter), hence the "cost of privacy" is not an intrinsic property of a differential privacy mechanism but is rather dependent on task difficulty and dimensionality. These results are placed in the context of the federated learning literature in a PRISMA-compliant review of 29 studies. Finally, we present practical suggestions for provisioning communication rounds, client participation percentages, and privacy budgets when rolling out FL across a set of distributed systems with different types and privacy restrictions

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

  Eduschool International Journal of Data Science and Machine learning (EIJDSML) 2026 · 0 citations

• Privacy-Preserving Federated Learning for Artificial Intelligence Agriculture Disease Detection Under Heterogeneous Farm Data Silos: A Computer Differential-Privacy Utility Evaluation

  Zenodo (CERN European Organization for Nuclear Research) 2026 · 0 citations