以分群客製及分群重新加權提升聯邦學習在域名轉移下的效能及公平性

Abstract

聯邦學習（FL）已經成為一種協作訓練框架，用於去中心化數據，但它面臨著客戶之間的數據異質性帶來的挑戰。現有方法主要集中於解決標籤偏斜或非獨立同分佈特徵變化的問題，卻忽視了客戶分群以及每個群集的數據不平衡。在這項工作中，我們提出了FedGCR（具有分群客製和分群重新加權的聯邦學習）的分層FL方案，以解決這些限制。在FedGCR中，我們使用「分群客製（GC）」來將具有相似特徵分佈的客戶進行分組，使它們可以相互學習領域特定知識，並從共享服務器模型中學習領域不變知識。此外，我們通過一種新穎的「分群重新加權（GR）」算法來解決群集大小不平衡的問題，該算法增強了不同群組之間的性能一致性。在多個數據集上的實驗評估表明，FedGCR在準確性和性能一致性方面優於現有方法。所提出的方法促進了聯邦學習的進步，使得在具有客戶相似性和群集不平衡的場景中能夠更有效地進行知識共享並提高性能。

Federated learning (FL) has emerged as a collaborative training framework for decentralized data, but it faces challenges due to data heterogeneity among clients. Existing approaches primarily focus on addressing label skew or non-IID feature shift, while neglecting client clustering and the data imbalance of each cluster. In this work, we propose FedGCR (Federated learning with Group Customization and Reweighting), a stratified FL scheme, to address these limitations. In FedGCR, we use Group Customization (GC) to group clients with similar feature distributions, allowing them to learn domain-specific knowledge from one another and domain-invariant knowledge from the shared server model. Additionally, we tackle the issue of imbalanced cluster sizes through a novel Group Reweighting (GR) algorithm, which enhances performance uniformity among different groups. Experimental evaluations on multiple datasets demonstrate that FedGCR outperforms state-of-the-art methods in terms of accuracy and performance uniformity. The proposed approach contributes to the advancement of federated learning by enabling more effective knowledge sharing and improved performance in scenarios with client similarity and imbalanced clusters.