FedQoS, a federated system that predicts future quality-of-service (QoS) failure risk for candidate wireless links, stayed close to a centralized reference in a simulated campus network, according to an arXiv preprint. Among the federated methods tested, it had the highest displayed balanced accuracy and recall in the mild non-independent and identically distributed (non-IID) condition, and the highest displayed balanced accuracy, recall and F1 score in severe non-IID data. The framework supports access-node selection without centralizing user-level network data. But the result comes from physics-based synthetic network logs, not a real deployment.
FedQoS couples stabilized local training with QoS-aware aggregation that reflects both the amount of data at an access node and its exposure to QoS failures. That setup is built for heterogeneous, non-IID client logs. The model treats signal-to-interference-plus-noise ratio (SINR) as an explanatory feature rather than the sole target. Its QoS-failure label captures radio quality, resource contention, service demand and mobility over a short future horizon.
A system built around local network logs
The controller applied a hysteresis margin and a minimum dwell-time condition before triggering a handover, a design intended to avoid ping-pong handovers. In practical terms, the rules discourage switching access nodes after every small change in the estimate. Samples remained at the access nodes that observed them and were split locally into 70% training, 15% validation and 15% test sets. The evaluation used two naturally non-IID conditions.
The logs were generated with Sionna RT and Sionna SYS. The simulated campus measured 200 metres by 200 metres and contained three multi-floor buildings, one outdoor base station (BS) and two quasi-static unmanned aerial vehicle (UAV)-mounted access nodes. The comparison included centralized learning, Local-only training, FedAvg and FedProx, alongside conventional non-learning access-selection policies.
Prediction remained close to the centralized reference
In the mild condition, FedQoS recorded a balanced accuracy of 0.9440, with a standard deviation of 0.0049; recall of 0.9137, with a standard deviation of 0.0125; and an F1 score of 0.8672, with a standard deviation of 0.0025. It led the federated methods on displayed balanced accuracy and recall and remained close to the centralized reference.
In severe non-IID data, the corresponding figures were 0.9679 with a standard deviation of 0.0029 for balanced accuracy, 0.9617 with a standard deviation of 0.0085 for recall, and 0.9451 with a standard deviation of 0.0014 for F1 score. FedQoS had the highest displayed score among federated methods on all three measures and again stayed close to centralized performance.
Learning-based results were averaged over 10 random seeds and reported as a mean plus or minus standard deviation.
Selection gains were clearer in mild conditions
In mild non-IID data, FedQoS had the lowest displayed federated QoS-failure rate and block-error rate (BLER). Its failure rate was 0.0031, with a standard deviation of 0.0002; mean throughput was 17.6586 megabits per second (Mbps), with a standard deviation of 0.0476; and BLER was 0.0009, with a standard deviation of 0.0001. The paper describes these lower failure and BLER figures as coming with a small throughput tradeoff.
For the mild-condition comparisons, the paper reports approximate QoS-failure reductions of 13.9% versus FedAvg, 6.1% versus FedProx, 43.6% versus Local-only training and 96.6% versus Historical QoS. These are relative figures from the tested simulation.
In severe non-IID data, FedQoS tied FedAvg at the displayed precision for the lowest federated QoS-failure rate. Its failure rate was 0.1356, with a standard deviation of 0.0003; throughput was 19.1817 Mbps, with a standard deviation of 0.0346; and BLER was 0.0096, with a standard deviation of 0.0002. The paper describes the differences among federated methods as modest.
Across both settings, learning-based policies had lower QoS-failure rates and BLER than conventional heuristic policies. An Oracle reference had the lowest failure rate and BLER and the highest throughput, but it used future QoS outcomes and was included only as an unattainable reference.
The evidence stops at the simulation
A separate grid check varied q and lambda. Policy QoS-failure rates ranged from approximately 0.00302 to 0.00334 in mild data and from 0.13553 to 0.13594 in severe data. The main setting, q = 0.75 and lambda = 0.5, lay in the low-failure region, indicating a broad performance plateau across the tested grid.
Those results belong to a defined simulation. The evidence is limited to the Sionna-based synthetic campus, the tested topology, the listed baselines, the evaluated parameters, the two non-IID conditions and the random seeds. It does not establish how the framework would perform on real-world network logs or outside those tested settings.
The document is an arXiv preprint, version 1, dated 26 August 2026. The work was supported in part by the Luxembourg National Research Fund through CHIST-ERA SHIELD, Grant INTER/CHIST24/19023763/SHIELD, and in part by the FNR AFR program through FULFIL, Grant 2000141.
Paper data and sources
Original title: FedQoS: Federated QoS-Risk Learning for Heterogeneous Indoor-Outdoor Access Selection
Authors: Nguyen Van Thieu, Ti Ti Nguyen, Ons Aouedi et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text