Federated Learning-Based Intelligent Resource Management for Privacy-Preserving 6G Networks

Abstract

The evolution toward sixth-generation (6G) wireless networks demands ultra-low latency, massive device connectivity, and native artificial-intelligence (AI) support, yet centralized machine-learning pipelines conflict with the stringent data-privacy and bandwidth constraints of dense edge environments. This paper proposes FL-IRM, a Federated Learning-based Intelligent Resource Management framework that jointly optimizes client-side model training and network-side resource allocation for privacy-preserving 6G networks. The framework introduces a three-tier architecture spanning edge devices, an edge/fog aggregation layer, and a 6G core secure-aggregation server, in which a lightweight learning-based resource manager adaptively allocates bandwidth and transmit power to reduce straggler-induced client dropout, while local differential privacy (LDP) and secure aggregation protect client updates from inference attacks. A deadline-aware scheduling policy, formulated as a sequential decision process, equalizes per-device completion latency and thereby minimizes the round-level makespan that governs synchronous federated learning (FL). Simulation studies on a non-IID 30-client testbed show that FL-IRM improves global classification accuracy from 56.74% to 60.52% and macro-F1 from 0.550 to 0.604 relative to a static-allocation FedAvg baseline, while a companion 25-device resource-allocation study shows a 46.9% reduction in round makespan versus equal allocation and an 81.8% reduction versus greedy channel-based allocation, with inter device latency variance reduced by more than three orders of magnitude. A differential-privacy sweep further shows that the proposed framework retains within 1–3 percentage points of its non-private accuracy for privacy budgets ε ≥ 1.0. These results indicate that jointly reasoning about learning dynamics and radio-resource scheduling offers a practical path toward trustworthy, resource-efficient intelligence at the 6G edge.The evolution toward sixth-generation (6G) wireless networks demands ultra-low latency, massive device connectivity, and native artificial-intelligence (AI) support, yet centralized machine-learning pipelines conflict with the stringent data-privacy and bandwidth constraints of dense edge environments. This paper proposes FL-IRM, a Federated Learning-based Intelligent Resource Management framework that jointly optimizes client-side model training and network-side resource allocation for privacy-preserving 6G networks. The framework introduces a three-tier architecture spanning edge devices, an edge/fog aggregation layer, and a 6G core secure-aggregation server, in which a lightweight learning-based resource manager adaptively allocates bandwidth and transmit power to reduce straggler-induced client dropout, while local differential privacy (LDP) and secure aggregation protect client updates from inference attacks. A deadline-aware scheduling policy, formulated as a sequential decision process, equalizes per-device completion latency and thereby minimizes the round-level makespan that governs synchronous federated learning (FL). Simulation studies on a non-IID 30-client testbed show that FL-IRM improves global classification accuracy from 56.74% to 60.52% and macro-F1 from 0.550 to 0.604 relative to a static-allocation FedAvg baseline, while a companion 25-device resource-allocation study shows a 46.9% reduction in round makespan versus equal allocation and an 81.8% reduction versus greedy channel-based allocation, with inter device latency variance reduced by more than three orders of magnitude. A differential-privacy sweep further shows that the proposed framework retains within 1–3 percentage points of its non-private accuracy for privacy budgets ε ≥ 1.0. These results indicate that jointly reasoning about learning dynamics and radio-resource scheduling offers a practical path toward trustworthy, resource-efficient intelligence at the 6G edge.