Hybrid Makes Better: Privacy-Preserving Medical Image Classification Based on Federated Learning

The power of machine learning makes it available for medical image classification. However, machine learning requires large medical datasets to improve accuracy, which will involve patients' private information and lead to their privacy leakage. Federated learning is a trending technique to both protect patients' privacy and improve the accuracy of medical image classification. Unfortunately, current research shows that federated learning faces the risk of privacy leakage. In this paper, we propose a privacy-preserving scheme FHDM. Specifically, we construct a local hybrid differential privacy algorithm (LHDP) against patients' privacy leakage. This hybrid algorithm utilizes both Gaussian and Laplace differential privacy without enlarging the privacy budget. We prove that the algorithm applies to the model parameter. Moreover, we design loss optimization and global optimization strategies on the algorithm to achieve higher accuracy in medical image classification. Finally, we validate FHDM in terms of privacy-preserving and model accuracy on real datasets. Experiments show that FHDM effectively protects privacy and improves the average accuracy by 10.0% compared to adopting the LDP-based scheme with the same medical image dataset and privacy budget.