[論文レビュー] Mitigating Bias in Federated Learning

Annie Abay, Yi Zhou+4 more|arXiv (Cornell University)|Dec 4, 2020

Privacy-Preserving Technologies in Data参考文献 27被引用数 41

ひとこと要約

本論文は連邦学習におけるバイアスの源を分析し、3つのバイアス緩和手法—local reweighing、global reweighing with differential privacy、そして federated prejudice removal—を提案し、IIDおよび非IIDデータ分布の下でプライバシーを損なうことなくバイアスを低減できることを示す。

ABSTRACT

As methods to create discrimination-aware models develop, they focus on centralized ML, leaving federated learning (FL) unexplored. FL is a rising approach for collaborative ML, in which an aggregator orchestrates multiple parties to train a global model without sharing their training data. In this paper, we discuss causes of bias in FL and propose three pre-processing and in-processing methods to mitigate bias, without compromising data privacy, a key FL requirement. As data heterogeneity among parties is one of the challenging characteristics of FL, we conduct experiments over several data distributions to analyze their effects on model performance, fairness metrics, and bias learning patterns. We conduct a comprehensive analysis of our proposed techniques, the results demonstrating that these methods are effective even when parties have skewed data distributions or as little as 20% of parties employ the methods.

研究の動機と目的

• Identify and analyze causes of bias in federated learning settings.
• Adapt centralized bias mitigation techniques to FL without compromising data privacy.
• Propose and evaluate three bias-mitigation methods under varying data heterogeneity.
• Assess fairness metrics and privacy trade-offs in FL scenarios.

提案手法

• Analyze causes of bias in FL including party selection, data heterogeneity, and fusion algorithms.
• Adapt Reweighing and Prejudice Remover from centralized ML to FL in two privacy-preserving forms: local reweighing (pre-processing) and global reweighing with differential privacy (pre-processing) plus federated prejudice removal (in-processing).
• Propose three mitigation approaches: Local Reweighing, Global Reweighing with DP, Federated Prejudice Removal.
• Evaluate methods using IID and non-IID partitions on UCI Adult and ProPublica Compas datasets, using logistic regression with fairness metrics.
• Compare against centralized baselines and analyze privacy-performance-fairness trade-offs.

実験結果

リサーチクエスチョン

• RQ1 What are the main sources of bias in federated learning?
• RQ2 Can centralized bias-mitigation techniques be adapted to FL without sacrificing privacy?
• RQ3 How effective are pre-processing and in-processing bias mitigation methods in FL under IID and non-IID data?
• RQ4 What is the impact of partial party participation and varying data distributions on bias mitigation effectiveness in FL?

主な発見

• Local reweighing reduces bias and maintains prediction accuracy across IID settings.
• Federated prejudice removal reduces bias in several metrics but may trade off with accuracy, and shows more stable behavior than local reweighing.
• Global reweighing with differential privacy can mitigate bias under DP budgets; effectiveness depends on privacy parameter epsilon.
• Partial participation (only a subset of parties) still yields bias reduction, showing method robustness to dropouts.
• In highly imbalanced data, local reweighing remains effective as long as parties have some representation from both groups; federated prejudice removal is less consistently effective.
• Some fairness metrics may be unreliable in FL settings, and the choice of metric matters for evaluating fairness.