Scholarlink

<p>Federated learning has emerged as a prominent system to safeguard users’ privacy by processing data locally. Despite its wide acceptance, it experiences inaccurate predictions due to inherent bias in non-independently and identically distributed (non-IID), raising fairness concerns particularly toward minority demographic groups. Moreover, recent federated learning literature does not adequately address privacy and fairness simultaneously, despite their critical importance. In this paper, we propose to incorporate federated learning with fair decision-making system that are ethical, trustworthy, and free from unintentional bias. Proposed FedEQ relies on client-side data equilibration to address imbalance based on sensitive attributes within a federated system. To calibrate system bias, disparate impact, statistical parity difference, and equalized odds difference are used as fairness metrics. For user’s data privacy, a bias-based metric masking (BMM) technique is proposed that allows weighted updates containing metric scores to be shared with server to enhance fairness within system. The proposed data-driven approach has demonstrated to be more effective in terms of accuracy and fairness trade-off (approx. 2%), and ensures equal treatment across demographic groups (approx. 12.7%) based on models prediction compared to existing model-based fairness schemes in federated system.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Toward a socially fair data equilibration approach for responsible federated learning

  • 期刊论文

  • 发表日期：

    2025年9月20日

查看全文
  • Jaya Pathak,
  • Jagat Sesh Challa,
  • Amitesh Singh Rajput

摘要

Federated learning has emerged as a prominent system to safeguard users’ privacy by processing data locally. Despite its wide acceptance, it experiences inaccurate predictions due to inherent bias in non-independently and identically distributed (non-IID), raising fairness concerns particularly toward minority demographic groups. Moreover, recent federated learning literature does not adequately address privacy and fairness simultaneously, despite their critical importance. In this paper, we propose to incorporate federated learning with fair decision-making system that are ethical, trustworthy, and free from unintentional bias. Proposed FedEQ relies on client-side data equilibration to address imbalance based on sensitive attributes within a federated system. To calibrate system bias, disparate impact, statistical parity difference, and equalized odds difference are used as fairness metrics. For user’s data privacy, a bias-based metric masking (BMM) technique is proposed that allows weighted updates containing metric scores to be shared with server to enhance fairness within system. The proposed data-driven approach has demonstrated to be more effective in terms of accuracy and fairness trade-off (approx. 2%), and ensures equal treatment across demographic groups (approx. 12.7%) based on models prediction compared to existing model-based fairness schemes in federated system.