SAB-Select Staleness-Aware Burst-Adaptive Client Selection for Federated Learning under Bursty Connectivity

Federated learning enables distributed devices to train a shared model without transmitting raw data to a central server. In real-world networks, devices intermittently disconnect and reconnect in bursts. When a device returns after a prolonged offline period, its update is computed from an outdated global model; such stale updates introduce noise, increase gradient variance, and slow convergence. Existing client selection methods either ignore staleness or address it post hoc through aggregation, and therefore fail to jointly optimize staleness and bursty connectivity. This paper proposes SAB-Select, a client selection method that scores each client using three signals: staleness, burst availability, and gradient diversity. Theoretical convergence analysis demonstrates that SAB-Select reduces the number of rounds required to reach a target accuracy. This paper also introduces an optional audit log for accountability, which can be instantiated as a signed append-only log or as a permissioned blockchain ledger. Experiments on MNIST and Fashion-MNIST demonstrate that SAB-Select reaches 85% accuracy faster (7 vs. 12 rounds), reduces average staleness, and maintains fairness comparable to the baselines. A cost-based analysis (communication bytes and time proxy) further demonstrates that faster convergence translates into reduced bandwidth and latency requirements for reaching the target accuracy. The results demonstrate that staleness-aware client selection provides a practical, theoretically grounded solution for federated learning on realistic edge networks.