Algorithmic Bias Detection in Federated Learning

Abstract

Federated learning (FL) offers a promising approach to training machine learning models on decentralized data sources, preserving data privacy. However, the inherent data heterogeneity across clients introduces a significant challenge: algorithmic bias. This paper presents a novel method for detecting and quantifying bias within the federated learning process. Our approach leverages statistical measures and differential privacy techniques to monitor model updates and identify bias amplification. We define key metrics such as variance of model updates, divergence between client models, and the sensitivity of model predictions to sensitive attributes. The core of our method is the adaptive adjustment of learning rates based on these metrics, aiming to mitigate bias while maintaining model convergence. We demonstrate the effectiveness of our approach through a theoretical analysis and a conceptual framework, highlighting its potential for robust and fair FL systems. The proposed framework provides a quantifiable assessment of bias risk and suggests strategies for bias mitigation during the training process.