BACSA: A Bias-Aware Client Selection Algorithm for Privacy-Preserving Federated Learning in Wireless Healthcare Networks

TL;DR

This work proposes the Bias-Aware Client Selection Algorithm (BACSA), which detects user bias and strategically selects clients based on their bias profiles, and demonstrates that BACSA improves convergence and accuracy, compared to existing benchmarks, through evaluations on various data distributions.

Abstract

Federated Learning (FL) has emerged as a transformative approach in healthcare, enabling collaborative model training across decentralized data sources while preserving user privacy. However, performance of FL rapidly degrades in practical scenarios due to the inherent bias in non Independent and Identically distributed (non-IID) data among participating clients, which poses significant challenges to model accuracy and generalization. Therefore, we propose the Bias-Aware Client Selection Algorithm (BACSA), which detects user bias and strategically selects clients based on their bias profiles. In addition, the proposed algorithm considers privacy preservation, fairness and constraints of wireless network environments, making it suitable for sensitive healthcare applications where Quality of Service (QoS), privacy and security are paramount. Our approach begins with a novel method for detecting user bias by analyzing model parameters and correlating them with the distribution of class-specific data samples. We then formulate a mixed-integer non-linear client selection problem leveraging the detected bias, alongside wireless network constraints, to optimize FL performance. We demonstrate that BACSA improves convergence and accuracy, compared to existing benchmarks, through evaluations on various data distributions, including Dirichlet and class-constrained scenarios. Additionally, we explore the trade-offs between accuracy, fairness, and network constraints, indicating the adaptability and robustness of BACSA to address diverse healthcare applications.

Figures (7)

• Figure 1: Illustration of FL in a wireless IoT for healthcare scenario: Clients have different health conditions causing severely non-IID data. A subset of clients need to be selected for each communication round due to system and device limitations: Clients 5 and 6 are omitted in this round.
• Figure 2: Data distributions based on Dirichlet and CCDD using CIFAR-10 dataset. (a) Dirichlet distribution with $\alpha = 0.1$. (b) Dirichlet distribution with $\alpha = 1$. (c) CCDD with $\Phi = 2$ and $\Gamma = 10$. (d) CCDD with $\Phi = 5$ and $\Gamma = 10$.
• Figure 3: Class estimation accuracy using proposed weight initialization (WI) and WI as mentioned in pmlr-v9-glorot10a (a) Dirichlet $\alpha = 0.5$. (b) CCDD $\Phi = 5$. (c) NIH-CXR Dirichlet $\alpha = 0.2$.
• Figure 4: NIH-CXR dataset wang2017hospital
• Figure 5: Comparing legacy and proposed weight initialization methods on class estimation accuracy using CIFAR-10 dataset