Secure Federated Learning for Parkinson's Disease: Non-IID Data Partitioning and Homomorphic Encryption Strategies

• Published in: IEEE access Vol. 12; pp. 127309 - 127327
• Main Authors: Tanim, Sharia Arfin, Aurnob, Al Rafi, Mridha, M. F., Safran, Mejdl, Alfarhood, Sultan, Che, Dunren
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 2024; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Accuracy; Algorithms; Attention based fusion; Computational modeling; Data models; Data privacy; Datasets; Distributed databases; Encryption; Federated learning; Heterogeneity; Homomorphic encryption; Machine learning; Parkinson's disease; Partitioning; partitioning method; Privacy; Robustness; secure federated learning; Training
• ISSN: 2169-3536; 2169-3536
• DOI: 10.1109/ACCESS.2024.3454690

In this paper, we explore methods to enhance both the performance and privacy of federated learning models by implementing two key techniques: homomorphic encryption and attention-based fusion. Federated learning which involves client-side training of models using multiple users data suffers from its drawbacks such as privacy issues and the variability of data across different users, known as data heterogeneity. To address these challenges, we introduce a Secure Federated Learning (SFL) algorithm. This approach avoids problems associated with Non-Independent and Identically Distributed (Non-IID) data by the introduced prosed partitioning method of datasets and then using homomorphic encryption for securely summing encrypted model parameters. Additionally, we use attention-based fusion techniques to boost model accuracy and robustness. The experiments performed on the PD-BioStampRC21 dataset with a focus on the disease's progression demonstrated significant improvements in accuracy and robustness compared to traditional methods. We evaluated our approach using metrics such as accuracy and assessed the model's consistency before and after encryption. The results validate the resemblance and efficacy of the encryption process while indicating the possibility of reliable and fast implementation of federated learning in clinical scenarios. Overall, we present an impactful and comprehensive framework focusing on privacy-preserving model averaging in practice with case studies in healthcare and other sensitive contexts.