Bias Mitigation in Federated Learning for Edge Computing

Federated learning (FL) is a distributed machine learning paradigm that enables data owners to collaborate on training models while preserving data privacy. As FL effectively leverages decentralized and sensitive data sources, it is increasingly used in ubiquitous computing including remote healthcare, activity recognition, and mobile applications. However, FL raises ethical and social concerns as it may introduce bias with regard to sensitive attributes such as race, gender, and location. Mitigating FL bias is thus a major research challenge. In this paper, we propose Astral, a novel bias mitigation system for FL. Astral provides a novel model aggregation approach to select the most effective aggregation weights to combine FL clients' models. It guarantees a predefined fairness objective by constraining bias below a given threshold while keeping model accuracy as high as possible. Astral handles the bias of single and multiple sensitive attributes and supports all bias metrics. Our comprehensive evaluation on seven real-world datasets with three popular bias metrics shows that Astral outperforms state-of-the-art FL bias mitigation techniques in terms of bias mitigation and model accuracy. Moreover, we show that Astral is robust against data heterogeneity and scalable in terms of data size and number of FL clients. Astral's code base is publicly available.