Abstract
Federated Learning (FL) in edge environments faces significant challenges due to resource volatility. While greedy client selection optimizes for speed, it often compromises fairness and long-term model robustness. This paper evaluates the trade-offs between accuracy, training time, and fairness, demonstrating that reputation-based fair selection provides a sustainable and effective alternative for real-world Edge AI.
Introduction
Federated Learning (FL) allows edge devices, such as smartphones and IoT sensors, to train a shared model without transmitting raw data to a central server. However, the edge is inherently volatile; devices possess heterogeneous processing powers and internet speeds that fluctuate based on battery levels, mobility, and varying workloads.
Standard FL protocols often employ "greedy" selection, prioritizing the fastest or most resource-rich devices to accelerate training. This approach introduces two critical issues: bias against slower devices and poor sustainability due to battery drain on high-performing nodes. This study investigates whether fairness-aware selection can mitigate these issues without significantly degrading model accuracy or training efficiency.
Methodology
The researchers conducted an empirical evaluation comparing traditional greedy algorithms with fairness-based selection methods:
- Random Selection: The baseline where every device has an equal chance.
- Greedy (Computational/Communication): Prioritizes devices based on current speed.
- RBFF (Reputation-Based Fair FL): A scoring system balancing current capabilities with past participation history.
- RBCSF (Reputation-Based Client Selection with Fairness): An extension of RBFF that penalizes over-selected devices.
Volatility was modeled by dynamically changing compute and communication capabilities in every training round. The study utilized three benchmarks: EMNIST (character recognition), FashionMNIST (clothing items), and CIFAR-10 (complex color images).
Key Results and Discussion
The experimental results highlight a fundamental shift in FL dynamics under volatile conditions. While greedy algorithms can be up to 25% faster in wall-clock time, they suffer a 2-3% accuracy penalty by ignoring diverse data from slower devices.
As shown in the figure above, fairness-based algorithms like RBFF and RBCSF maintain a high Jain’s Fairness Index (JFI), consistently exceeding 0.99. In contrast, greedy approaches exhibit significantly lower and more volatile fairness scores. Surprisingly, the fairness mechanisms provide a "robustness benefit" in Non-IID scenarios, where fair algorithms occasionally outperformed greedy ones by ensuring the model learned from the entire data distribution rather than a privileged subset.
Conclusion
Fairness in Edge AI is not merely an ethical consideration but a technical necessity for robustness and sustainability. Reputation-based selection (RBFF) offers a viable middle ground, providing competitive accuracy and superior fairness compared to greedy methods. Future work will focus on integrating these fairness mechanisms with privacy-preserving techniques like Differential Privacy and exploring long-term cumulative fairness metrics across the lifecycle of federated networks.
Summary updated