Comparing Energy Efficiency and Robustness in Federated Learning

In this article, we present a novel approach to incorporate energy efficiency considerations into the clients’ selection process in federated learning. Our proposed algorithm, CA-AFL, leverages a tuning parameter to transition between two established algorithms: AFL and top-K greedy selection. We demonstrate through extensive simulations that CA-AFL outperforms existing baselines in terms of both energy efficiency and distributional robustness.
At the core of our algorithm is the concept of a "gating mechanism," which serves as a switch to control the selection process based on the value of the tuning parameter C. When C approaches infinity, the energy-efficient expert takes precedence, while when C approaches zero, the robustness expert becomes the default choice. For values between these extremes, the combined PMF represents a blend of the two insights.
We show that as C increases, our algorithm tends to select the client requiring the lowest energy for model upload. This dimensionality reduction within the simplex space is articulated through relationships between the tuning parameter and the resulting selection probabilities. Specifically, we derive expressions for lim C → ∞ rown[2] i and lim C → ∞ m i / rown[1].
These insights demonstrate that CA-AFL achieves a balance between energy efficiency and distributional robustness by selecting clients with the lowest energy requirements while ensuring that the resulting model is robust across different distributions. Remarkably, at C = 8, CA-AFL matches the energy efficiency of GCA (a robustness benchmark) without compromising performance, while surpassing GCA in terms of standard deviation (STD).
By leveraging a tuning parameter to control the selection process, CA-AFL offers a flexible and adaptive approach to incorporating energy efficiency considerations into federated learning. Our proposed algorithm not only achieves better performance metrics than existing baselines but also provides a promising direction for future research in this area.