Hidden in Rounds: Predicting the Time Cost of 802.11 Contention in Federated Learning

The paper "Hidden in Rounds: Predicting the Time Cost of 802.11 Contention in Federated Learning" (arXiv:2609.12903, submitted September 11, 2026) by Satwat Bashir and Tasos Dagiuklas analyzes how 802.11 contention impacts communication time in Federated Learning (FL).

Key Findings: Rounds vs. Time: While the number of rounds to reach target accuracy remains relatively stable across different client densities and offered loads, the communication time-to-target increases by approximately two orders of magnitude as client density rises. Predictive Estimator: The authors propose a Bianchi-anchored estimator that uses frame-delivery ratio and saturation throughput to predict communication time, achieving a mean absolute percentage error (MAPE) of 2.3% to 10.2% on held-out configurations. Methodology Limitation: The estimator’s accuracy is validated against a constructed communication-time metric derived from the same round-duration equation, rather than against independently measured end-to-end FL completion time. Participation Bias: The study found no statistically distinguishable accuracy gap between uniform and persistent heterogeneous participation, though confidence intervals were wide.

The research suggests that reporting only rounds-to-target obscures the true cost of contention, and that communication time is a more critical metric for capacity planning in dense Wi-Fi FL deployments.

This material examines how IEEE 802.11 contention affects the communication time cost of federated learning (FL) rounds. Rather than treating wireless transmission as a simple function of model-update size, it focuses on the shared-channel behavior of Wi-Fi when many FL clients attempt to send updates simultaneously. The authors use ns-3 to measure frame-delivery ratio and saturation throughput under varying client densities and offered loads, capturing how MAC-layer backoff, collisions, and contention reduce the effective reliability and throughput of update transmission.

A key contribution is the coupling of these network measurements with a FedAvg training process. The measured frame-delivery ratio is used as a first-order proxy for the probability that a client’s update is admitted or successfully delivered, and an equation is then used to estimate the communication time for FL rounds. This makes the “hidden” cost of contention explicit: two FL deployments with identical model sizes can experience very different round latencies depending on how many clients contend for the channel and how heavily the wireless medium is loaded.

The work matters because federated learning over Wi-Fi is a common but latency-sensitive setting, especially in consumer, edge, or access-network environments where many devices share a single radio. By providing a measurement-based view of how 802.11 contention translates into FL communication overhead, the material offers a useful basis for predicting round duration, managing client participation, tuning offered load, or designing more efficient FL scheduling and update-admission policies under realistic wireless conditions.