Federated Learning (FL) has increasingly been promoted as a practical way to train models over distributed edge data while avoiding raw-data centralization, yet most published evaluations have remained confined to a single task and therefore have provided limited evidence about how FL behaves across substantially different workloads. In this paper, a multi-domain empirical study of FL is presented in two representative edge settings: vision-based object detection for connected vehicles and time-series energy forecasting for smart-grid assets. A Flower-based system with explicit client, server, and coordination layers is implemented, and multiple data-partitioning strategies ranging from artificial to deployment-realistic heterogeneity are evaluated. Centralized baselines, isolated local training, federated training, and privacy-enhanced federated training are combined in the study. Across the evaluated domains, three recurring patterns are identified, especially under deployment-aligned client partitions. First, robustness to unseen data is improved by FL in the evaluated settings: in object detection, a 2.7% higher mAP50 and a 3.4% higher mAP50–95 are achieved by the federated global model compared with the average independently trained client on hidden validation data. Second, the value of FL is shown to depend strongly on how clients are formed: the benefit of collaboration can be obscured by random partitions, whereas substantial forecasting gains are produced by realistic seasonal and per-turbine partitions, including 16.6%–52.9% MSE reductions for the PV task. Third, much of the collaborative advantage is preserved by adaptive server-side clipping and noise addition in the evaluated PV forecasting scenario, with PV seasonalpartition MSE increased by 21.7% relative to non-private FL while still outperforming isolated training. A more precise picture is therefore provided of when FL is beneficial in heterogeneous edge environments and which evaluation choices most strongly affect the observed outcome.
Troulitaki et al. (Mon,) studied this question.