Federated Unlearning with Clustered Asynchronous Aggregation and Ensemble Learning for Efficient Privacy-Preserving Document Analysis

Ahmad Sarmad Ali, Momina Moetesum, Faisal Shafait, Adnan Ul-Hasan

Abstract: Federated Learning (FL) enables decentralized model training while preserving privacy, but compliance with data erasure regulations poses significant challenges. Existing federated unlearning methods suffer from high computational costs or incomplete data removal. This paper introduces a novel federated unlearning framework that integrates Asynchronous Clustered Aggregation and Federated Ensemble Learning, ensuring efficient unlearning while maintaining model accuracy. By retraining only affected clusters instead of the entire model, our approach reduces retraining overhead by 2\(\times \) while preserving 96–99% accuracy. Additionally, targeted noise injection prevents residual data influence, enhancing privacy protection. Evaluations on ICDAR-POD2017 and Invoices datasets demonstrate superior performance, achieving a 4–5% accuracy improvement over standard clustered aggregation. Our findings bridge the gap between privacy compliance and high-performance document analysis, offering a scalable and effective solution for federated unlearning.

External IDs:doi:10.1007/978-3-032-04617-8_6