Abstract
This paper introduces a novel framework for Clustered Federated Learning (CFL) designed to optimize the balance between model accuracy and energy efficiency in edge computing environments. By integrating Double Dueling Deep Q-Networks (D3QN) for edge association with hierarchical transfer learning for personalization, the proposed system addresses data heterogeneity and the energy-accuracy trade-off. Experimental results on the MNIST dataset demonstrate that the framework achieves high accuracy comparable to accuracy-optimal benchmarks while significantly reducing energy consumption and maintaining privacy.
Introduction
Federated Learning (FL) enables decentralized model training on user devices, preserving data privacy. However, two significant challenges persist: data heterogeneity (non-i.i.d. data) and the inherent trade-off between energy consumption and model accuracy. Users with diverse data distributions often require personalized models to achieve high performance. Yet, connecting to the most relevant base station for a specific data cluster may incur high energy costs if that station is geographically distant. Conversely, connecting to the nearest station might save power but result in a model that poorly fits the user's data. This work proposes a joint optimization of edge association and resource allocation to navigate these competing objectives.
Methodology
The authors formulate the problem as a Mixed-Integer Non-Linear Program (MINLP) aimed at maximizing a "System Profit" metric. This metric is defined as the weighted difference between the accuracy gains from federated training and the energy costs incurred during data transmission.
Edge Association via D3QN
To solve the complex association problem, a Double Dueling Deep Q-Network (D3QN) is deployed at the cloud level. The D3QN architecture separates the estimation of state values and action advantages, while the "double" mechanism mitigates the overestimation of Q-values. This reinforcement learning agent utilizes performance feedback—specifically, changes in model accuracy and energy consumption—to associate users with appropriate base station clusters. This approach preserves privacy by avoiding the direct exchange of raw model parameters for clustering purposes.
Optimal Bandwidth Allocation
Once edge association is determined, base stations must allocate limited bandwidth among their associated users. Using Karush-Kuhn-Tucker (KKT) conditions, the authors derive an optimal allocation strategy. A counter-intuitive but mathematically sound finding is that to minimize total system energy, more bandwidth should be allocated to users with poorer channel conditions. This "Worse is More" rule balances transmission times across the network, preventing devices with weak signals from draining their batteries during prolonged active radio states.
Hierarchical Transfer Learning
To mitigate performance drops for users who must connect to sub-optimal clusters due to extreme energy constraints, a hierarchical transfer learning approach is employed. Models are split into "base layers" that capture general features (shared and averaged within the cluster) and "personalization layers" that capture user-specific nuances (kept strictly local). This ensures that even "misplaced" users can maintain high local accuracy.
Experimental Setup
The framework was evaluated using the MNIST dataset under a label-swapped non-i.i.d. configuration. To simulate real-world diversity, users were assigned 500 samples, with specific labels swapped for different user groups (e.g., one group sees "3" as "8"). This creates distinct "personal preferences" that a single global model cannot satisfy, testing the system's ability to provide effective personalization.
Results and Discussion
The proposed solution was compared against several benchmarks, including "Greedy" association (signal-strength based), "Accuracy-optimal" association (ignoring energy), and standard DQN.
Performance Metrics
As illustrated in the experimental results, the D3QN approach demonstrated superior stability and faster convergence compared to standard DQN. The system profit consistently outperformed all benchmarks across varying user counts (from 3 to 15 users).
Key Insights
- Privacy-Preserving Clustering: Using accuracy feedback as a proxy for similarity effectively clusters users without exposing raw model weights, maintaining a high privacy standard.
- Energy Efficiency: The derived bandwidth allocation rule proved optimal for system-wide energy conservation, significantly reducing the energy overhead of federated training.
- Robustness: Hierarchical transfer learning acts as a safety net, ensuring that even when energy constraints force sub-optimal clustering, local model performance remains high.
Conclusion
The proposed energy-aware edge association framework effectively balances the needs of personalized federated learning with the practical constraints of mobile edge computing. By jointly optimizing association and resource allocation, the system achieves near-optimal accuracy with significantly improved energy efficiency. Future work will explore scaling the system to larger networks with thousands of devices and incorporating dynamic power control.
Summary updated