FedHeur: Multi-Heuristic Client Selection for Task Offloading in Federated Learning

Federated Learning (FL) enables decentralized model training across multiple clients while preserving data privacy. However, traditional client selection strategies, such as random sampling in FedAvg, may be inefficient in heterogeneous environments where client reliability, load, and data freshness vary significantly. In this paper, we propose Multi-Heuristic Client Selection for Task Offloading in Federated Learning (FedHeur), a novel task offloading-aware federated averaging framework that incorporates a multi-heuristic client selection strategy. Clients are scored based on a weighted combination of model similarity (cosine similarity to the global model), system load, failure rate, and data freshness. To evaluate the effectiveness of our approach, we conducted preliminary experiments using LSTM and CNN models in simulated federated settings. Results show that FedHeur achieves stable convergence and effective model training across heterogeneous clients, demonstrating the potential benefits of integrating task offloading and client selection heuristics in federated learning.