OracleFed: Latency-Aware Federated Learning via Dynamic Recovery and Causal Aggregation
Stefanos Tziampazis, Baran Can Gül, Nasser Jazdi, Michael Weyrich · IEEE Access · 2025
As federated learning (FL) continues to expand across wide-area, heterogeneous networks, preserving true training chronology grows complex, particularly in time-sensitive applications where network delays can obscure causal ordering and degrade global-model fidelity. Current aggregation schemes seldom account for temporal dynamics: they either stall on stragglers, discard late updates, or rely on arrival-based heuristics that conflate communication latency with training staleness. In light of this limitation, we introduceOracleFed—Order-Respecting Aggregation with Causality and Latency Equalization, a framework in which training-time lag is decoupled from network-transport delay and elevated as the primary driver of aggregation. Leveraging a hybrid design that integrates dual timestamping with system-wide synchronization, the framework continuously recalibrates the server’s waiting window and re-indexes arrivals by their generation timestamps, thereby preserving causal order. Rather than treating late arrivals as inherently stale,OracleFeddiscounts only genuine training staleness, allowing high-latency clients to contribute proportionally to their computational freshness. We benchmark the proposed framework against existing synchronous and asynchronous schemes across eight evaluation metrics in a distributed, in-vehicle emotion-recognition setting. Pairedt- andWilcoxon signed-rank analyses of twenty Monte-Carlo replicates verify thatOracleFedcombines the coordination advantages of both paradigms and achieves higher predictive quality, lower staleness, and fairer client participation than the baselines.