Fair Federated Learning under Domain Skew with Local Consistency and Domain Diversity

TL;DR

This work tackles performance fairness in Federated Learning under domain skew by uncovering Parameter Update Consistency (PUC) during local updates. It introduces FedHEAL, a two-component framework consisting of Federated Parameter-Harmonized Learning (FPHL), which discards unimportant parameter updates to prevent conflicts, and Federated Aggregation-Equalized Learning (FAEL), which reduces global-model bias across diverse domains by minimizing the variance of distances to local models. The method is model-agnostic and compatible with existing FL approaches, and demonstrates improved cross-domain fairness and accuracy on multi-domain datasets (Digits and Office-Caltech) with ablation and compatibility studies. By focusing on update significance and domain diversity, FedHEAL provides a practical, privacy-preserving path to公平 performance across heterogeneous clients.

Abstract

Federated learning (FL) has emerged as a new paradigm for privacy-preserving collaborative training. Under domain skew, the current FL approaches are biased and face two fairness problems. 1) Parameter Update Conflict: data disparity among clients leads to varying parameter importance and inconsistent update directions. These two disparities cause important parameters to potentially be overwhelmed by unimportant ones of dominant updates. It consequently results in significant performance decreases for lower-performing clients. 2) Model Aggregation Bias: existing FL approaches introduce unfair weight allocation and neglect domain diversity. It leads to biased model convergence objective and distinct performance among domains. We discover a pronounced directional update consistency in Federated Learning and propose a novel framework to tackle above issues. First, leveraging the discovered characteristic, we selectively discard unimportant parameter updates to prevent updates from clients with lower performance overwhelmed by unimportant parameters, resulting in fairer generalization performance. Second, we propose a fair aggregation objective to prevent global model bias towards some domains, ensuring that the global model continuously aligns with an unbiased model. The proposed method is generic and can be combined with other existing FL methods to enhance fairness. Comprehensive experiments on Digits and Office-Caltech demonstrate the high fairness and performance of our method.

Figures (11)

• Figure 1: Problem illustration of Federated Learning under domain skew. Conventional FL methods () exhibit potential performance disparities due to Parameter Update Conflicts and Model Aggregation Bias. The former indicates that varying parameter importance and inconsistent update directions lead to an unfair decline in aggregated performance. The latter suggests biased convergence objective, resulting in performance disparities. Our method () achieves more equitable performance across different domains while enhancing overall performance.
• Figure 2: Illustration of Parameter Update Consistency. The consistency of parameter updates is displayed over 10 and 100 rounds for a randomly selected layer of the client model update. A significant proportion of parameters maintain a consistent update direction, i.e., almost half of the parameters show the same direction for over 90 of the 100 rounds, indicating a persistent tendency to steer the global model in a fixed direction.
• Figure 3: Architecture illustration of FedHEAL. Clients send local model updates to the server. In FPHL (\ref{['sec:fphl']}), the server maintains a consistency table, computes the consistency of current updates with past directions and discards updates with low consistency. Then in FAEL (\ref{['sec:fael']}), server minimized the variance of distance between global model and client model to mitigate model aggregation bias.
• Figure 5: Comparison of convergence of average accuracy with counterparts on Digits. Please see details in \ref{['sec:sota']}.
• Figure 7: Comparison of convergence of average accuracy with and without the integration of FedHEAL, across selected FL methods. Please see details in \ref{['sec:hyperparam']}.