WHEN CLIENTS DRIFT: FEDERATED SLA-RISK FORECASTING ACROSS UNSEEN 6G RAN REGIMES

Abstract

In 6G radio access networks, the service patterns and operating environments of different clients vary significantly, posing a major challenge to the generalization of federated SLA risk prediction. This paper investigates federated next-window SLA risk prediction under unseen regimes using a leave-one-regime-out evaluation setting. We propose a federated method that combines client profile information and regime-aware aggregation to model cross-layer 6G RAN telemetry data. Experimental results show that centralized models remain stronger in AUROC, with Centralized-GRU reaching 0.882 and HistGB reaching 0.872; however, the proposed method performs best on metrics that more closely reflect practical decision quality, achieving the highest average balanced accuracy of 0.676, macro balanced accuracy of 0.676, and the best worst-regime F1 score of 0.485. Compared with vanilla FedProx-GRU, the method improves AUROC, F1, and balanced accuracy. Ablation experiments further show that client profiling and regime-aware weighting both contribute to robust prediction under unseen regimes. These results suggest that although the method is not the strongest overall ranker, it is more practically valuable for robust federated decision-making in unseen 6G RAN environments.