Personalized Federated Learning with Genetic Algorithm
Yu Huang, Zhijie Tang, Ang Li · 2023
Privacy concerns have heightened due to the risks associated with data breaches and surveillance, emphasizing the need for personal autonomy and trust in digital interactions. Federated Learning (FL) is a promising solution to this problem as it enables multiple parties to collaborate without disclosing their data. Nevertheless, the presence of non-identically and independently distributed (non-IID) user data can exert adverse effects on the convergence speed and efficacy of training. To address this predicament, this article proposes an evaluation mechanism known as "Genetic evaluation" which endeavors to strike a delicate balance between personalism and globalism while concurrently enhancing client personalism. The " Genetic evaluation " mechanism entails a preliminary genetic search conducted on the parameters derived from stochastic gradient descent (SGD). Subsequently, the deviation between the offspring and the global parameters is computed to quantify their disparities. By judiciously weighing the performance of each offspring and its respective deviation, a specific number of exceptional offspring are identified and selectively preserved in the repository. Finally, this article introduces a novel personalized FL algorithm called FedGA. Empirical evaluations convincingly demonstrate that the proposed algorithm attains superior convergence rates, as evidenced by a remarkable tenfold performance enhancement over FedAvg, a threefold improvement over Per-FedAvg, and a twofold improvement over PFedMe on the widely used CIFAR-10 dataset. Furthermore, experimental results show that FedGA outperforms preceding FedAvg and Per-FedAvg in terms of accuracy.