Abstract: As a distributed optimization paradigm, federated learning enables a large number of resource-constrained client nodes to collaboratively train models without sharing their data. However, traditional federated learning methods, such as fedAvg, often fail to adequately address fairness issues. In practical scenarios, data distributions are typically highly heterogeneous, and conventional aggregation operations may introduce biases towards certain clients, resulting in significant performance disparities across clients for the global model. To tackle this challenge, a novel algorithm named Federated Learning for Personalization and Fairness (fedPF) is proposed. fedPF aims to effectively mitigate inefficient aggregation behaviors in federated learning and, by exploring the correlations between the global model and local models, distributes personalized models among clients, ensuring a balanced performance distribution among clients while maintaining the performance of the global model. fedPF was evaluated and analyzed on Synthetic, MNIST, and CIFAR10 datasets, and its performance was compared with three other federated learning algorithms: fedprox, qfedavg, and fedavg. The experimental results demonstrate that fedPF achieves notable improvements in both effectiveness and fairness.

Key words: federated learning, fairness, personalization, heterogeneous data, client selection

摘要： 作为一种分布式优化范式，联邦学习允许大量资源有限的客户端节点能够在不共享数据的情况下协同训练模型。然而，传统联邦学习(fedavg)通常未充分考虑公平性的问题。在实际场景中，数据分布通常具备高度异构性，常规的聚合操作可能会使模型对某些客户端产生偏见，导致全局模型在客户端本地的性能分布出现巨大差异。针对这一问题，提出了一种面向个性化与公平性的算法fedPF。fedPF旨在有效减少联邦学习中低效的聚合行为，并通过寻找全局模型与本地模型的相关性，在客户端之间分配个性化模型，在保证全局模型性能的同时，使客户端本地性能分布更均衡。fedPF在Synthetic、MNIST以及CIFAR10数据集上进行实验和性能分析，并与fedprox、qfedavg和fedavg三种联邦学习算法进行对比。实验结果表明，fedPF在有效性和公平性上均得到显著提升。

关键词: 联邦学习, 公平, 个性化, 异构数据, 客户端选择