Abstract
Federated Learning (FL) typically relies on a central server to coordinate model updates from edge devices, often leading to significant communication overhead and inefficiencies in non-IID data environments. This paper presents a decentralized approach to user selection that leverages standard wireless random access protocols. By mapping model importance to transmission probability via contention window manipulation, the proposed system prioritizes informative updates while maintaining fairness through a local counter mechanism. Experimental results on Fashion-MNIST and CIFAR-10 demonstrate that this decentralized method achieves convergence speeds and accuracy comparable to centralized priority-based selection, effectively mitigating the negative impacts of data correlation.
Introduction
In Federated Learning, a central server coordinates training across numerous edge devices, such as smartphones or autonomous vehicles. However, bandwidth limitations present a major bottleneck. Traditional user selection is handled centrally, adding complexity and communication overhead. Furthermore, real-world training data is often non-IID (not independent and identically distributed), meaning users possess varying data biases. Random selection in such scenarios can lead to biased global models or missed critical updates. This research explores an "intelligent" user selection method that operates without a centralized controller by leveraging the intrinsic properties of wireless networks.
Distributed User Selection Mechanism
The authors propose a distributed mechanism operating at the Medium Access Control (MAC) layer to handle user selection efficiently.
The process involves several key components:
- Priority-Based Random Access: Instead of server-side polling, users compete for channel access using Carrier Sensing Multiple Access (CSMA). Winning the competition grants the right to upload a model update.
- Contention Window (CW) Manipulation: The backoff timer—the random delay before transmission—is modified based on a priority score. This score is calculated using the Euclidean distance between a user's local model and the previous round's global model. Users with "novel" information (larger distance) receive shorter CWs, increasing their statistical probability of selection.
- Fairness Counter: To prevent high-information users from dominating and biasing the model, a local counter tracks selection frequency. If a user exceeds a specific threshold (set at 16% in tests), they voluntarily abstain from the current round.
- FedAvg Integration: The server aggregates the first $K^t$ received models using the standard Federated Averaging (FedAvg) algorithm.
Experimental Evaluation
The framework was evaluated using two standard computer vision benchmarks: Fashion-MNIST (using MLP and CNN architectures) and CIFAR-10 (using CNN). These were tested under both IID and non-IID scenarios to simulate realistic data distributions.
Key Results and Insights
The core innovation lies in treating the wireless channel as a natural selection filter. By mapping model value to transmission probability, the need for the server to pre-screen user data is eliminated.
Key findings include:
- Performance Parity: The decentralized random access approach achieved convergence speeds and accuracy nearly identical to centralized priority-based selection.
- Superiority over Random Selection: In non-IID scenarios, the priority-based method significantly outperformed standard random selection, reaching higher accuracy more rapidly.
- Impact of Window Size: The base contention window size ($N$) is a critical hyperparameter. Larger windows (e.g., 2048) provide better resolution for prioritization, allowing for more effective differentiation between users.
- Effective Fairness: The local counter mechanism successfully balanced contribution frequency, ensuring model stability and preventing over-representation of specific data subsets.
Conclusion and Future Work
While the proposed method effectively decentralizes user selection, several areas remain for future exploration. The current study focused on 10-user scenarios; scaling to high-density environments where wireless collisions increase is a necessary next step. Additionally, the sensitivity of hyperparameters like the fairness threshold and window size suggests a need for dynamic adaptation mechanisms. Finally, the energy trade-off between active carrier sensing in decentralized competition and centralized polling warrants further investigation to optimize for battery-constrained edge devices.
Summary updated