Algorithmic Bias Detection and Mitigation in Federated Learning

Abstract

Federated learning (FL) presents a promising paradigm for training machine learning models on decentralized data, preserving data privacy. However, this approach is increasingly recognized to be vulnerable to algorithmic bias. The inherent diversity and potential imbalances within the data contributed by numerous clients can significantly impact the fairness and equity of the resulting global model. This paper investigates the propagation and amplification of biases in federated learning systems. We propose a framework for detecting and mitigating these biases by integrating differential privacy and fairness constraints into the training process. The core claim is that FL systems are inherently susceptible to bias propagation. The proposed mechanism involves monitoring model performance across client subgroups, quantifying bias metrics, and adjusting the training process to enforce fairness. This work contributes to the development of more robust and trustworthy FL systems, paving the way for responsible and equitable AI applications.