| Abstract | Federated Learning (FL) has enabled Machine Learning (ML) applications to capture a greater spectrum of training data by allowing all data to remain on-device, thereby providing privacy guarantees desirable in many industries. However, current FL approaches impose strict eligibility requirements on potential clients and suffer from the straggler problem, where client resource heterogeneity, unstable connectivity, and varying task-dedication result in client slowdown or dropout, thus introducing bias and potentially completely failing the training process. In this paper, we challenge the binary assumption that all training must take place either on-device (FL) or on-server (Centralized ML), instead observing the common occurrence of trust between clients and using this trust to allow offloading of data between trusted clients. We present the FedTeams framework that leverages trusted data offloading by (1) optimizing selection of training nodes that minimize training latency and maximize data captured, and (2) mitigating the straggler problem through round-based monitoring of client state. We develop and evaluate the FedTeams framework in a simulated environment, demonstrating an 81% decrease in training latency, 79 % decrease in energy consumption, and 4.5-11.5% increase in global model accuracy when compared to Cloud-based Standard FL and Edge/Fog-based Hierarchical FL.
|
|---|
