[Paper Review] Fair Federated Medical Image Segmentation via Client Contribution Estimation

Meirui Jiang, Holger R. Roth+7 more|arXiv (Cornell University)|Mar 29, 2023

Privacy-Preserving Technologies in Data4 citations

TL;DR

This paper proposes FedCE, a novel federated learning method for medical image segmentation that jointly optimizes collaboration fairness and performance fairness by estimating client contributions in both gradient and data space. By combining gradient direction differences and prediction error via an auxiliary model, FedCE uses these estimates as adaptive aggregation weights, achieving superior performance and robustness on real-world medical datasets with improved fairness and free-rider detection.

ABSTRACT

How to ensure fairness is an important topic in federated learning (FL). Recent studies have investigated how to reward clients based on their contribution (collaboration fairness), and how to achieve uniformity of performance across clients (performance fairness). Despite achieving progress on either one, we argue that it is critical to consider them together, in order to engage and motivate more diverse clients joining FL to derive a high-quality global model. In this work, we propose a novel method to optimize both types of fairness simultaneously. Specifically, we propose to estimate client contribution in gradient and data space. In gradient space, we monitor the gradient direction differences of each client with respect to others. And in data space, we measure the prediction error on client data using an auxiliary model. Based on this contribution estimation, we propose a FL method, federated training via contribution estimation (FedCE), i.e., using estimation as global model aggregation weights. We have theoretically analyzed our method and empirically evaluated it on two real-world medical datasets. The effectiveness of our approach has been validated with significant performance improvements, better collaboration fairness, better performance fairness, and comprehensive analytical studies.

Motivation & Objective

• To address the dual challenge of collaboration fairness and performance fairness in federated medical image segmentation.
• To develop a robust client contribution estimation method that is less biased by data quantity and distribution shifts.
• To design a federated learning algorithm (FedCE) that uses contribution estimates for adaptive global model aggregation.
• To empirically validate the method on real-world medical imaging datasets with comprehensive fairness and robustness analysis.
• To detect free-riding clients early in training using the contribution estimation metric.

Proposed method

• Estimate client contribution in gradient space by measuring directional differences between local and global model updates.
• Quantify contribution in data space using prediction error on client data via a client-excluded auxiliary model.
• Combine gradient and data space metrics using a multiplicative or additive strategy to form a comprehensive contribution score.
• Use the estimated client contributions as adaptive weights in global model aggregation to improve convergence and fairness.
• Theoretical analysis establishes convergence guarantees and robustness bounds under distribution shifts.
• Leverage naturally generated gradients and model outputs during training to avoid additional computational overhead.

Experimental results

Research questions

• RQ1 Can client contribution estimation that combines gradient and data space improve both collaboration and performance fairness in federated medical image segmentation?
• RQ2 How robust is the proposed contribution estimation metric to data distribution shifts and client heterogeneity?
• RQ3 Can the method detect free-riding clients early in the training process?
• RQ4 Does using contribution-based aggregation weights lead to better segmentation performance than standard FL methods?
• RQ5 How do individual components (gradient and data space metrics) contribute to the overall performance?

Key findings

• FedCE significantly outperforms state-of-the-art FL methods on two real-world medical image segmentation datasets in terms of segmentation accuracy and fairness metrics.
• The contribution estimation metric shows high robustness to data distribution shifts, with value changes below 1% when excluding one client from the dataset.
• Free-riding clients are detected with high confidence within 10 training rounds, demonstrating early detection capability.
• Abllation studies confirm that both gradient and data space metrics are essential, as using either alone leads to performance degradation.
• The combination of gradient and data space contributions yields superior performance, with marginal but consistent gains in low-heterogeneity settings like prostate segmentation.
• Theoretical analysis confirms convergence of FedCE and provides an upper bound on value changes under distribution shifts, validating robustness.