DPFedSAM-Meas: Comparison Of Differential Privacy Federated Learning In Medical Image Classification

Machine learning is widely used in medical image classification tasks. However, medical images often exhibit uneven distribution and high sensitivity to noise. A feasible solution involves using federated learning (FL) with differential privacy (DP), a distributed training method that protects patient privacy. In this work, many studies have proposed improved solutions for localized differential-private federated learning (DP-FL) frameworks. However, these studies are isolated, which would be detrimental to privacy practitioners in designing and using the algorithms. To provide a comprehensive analysis of these algorithms, we propose DPFedSAM-Meas, as a framework for comprehensive utility analysis of DP-FL. In this framework, we employed the state-of-the-art federated learning framework FedSAM. Moreover, we categorize DP algorithms into Laplace DP and Gaussian DP by the underlying DP mechanisms, and into Gradient DP and Parameter DP by the DP position in FL train. DPFedSAM-Meas allows a comparative analysis of these four DP techniques, measuring their model utility, privacy leakage, and overhead when FL uses different network structures. Finally, we evaluate DPFedSAM-Meas on datasets of Pneumonia, Blood, and Path, aiming to investigate the performance of different DP techniques on mainstream deep learning algorithms, including Convolutional Neural Networks (CNN) and Vision Transformers (ViT).