Mask-Based Personalised Federated Learning and Its Application to Healthcare Privacy

With the increasing demand for personalisation and privacy protection in medical image analysis, federated learning (FL) techniques have become an important research area. PPPML-HMI has been proposed as a federated learning framework for personalisation and privacy protection designed for heterogeneous medical image data. Although existing PPPML-HMI frameworks have progressed in protecting privacy and handling heterogeneous data, their aggregation process still suffers from the security risk of client gradient leakage. In this paper, we propose an improved aggregation strategy that aims to enhance data protection for cyclic polymerisation processes in PPPML-HMI. Firstly, this paper proposes an attack against PPPML-HMI where a random initiator can compute the gradient of the next client after eavesdropping on the message sent by the next client. Secondly, we optimise the cyclic aggregation process of PPPML-HMI using AES based masking technique. Finally, the scheme's security analysis is performed, showing that the improved scheme can effectively resist malicious eavesdropping attacks and protect the client's gradient security. The theoretical analyses show that our improved method has stronger privacy protection capability under simulated attacks compared to the original PPPML-HMI framework. These theoretical analyses demonstrate the effectiveness and practicality of our approach. Our research provides new perspectives on the challenges of federated learning in preserving privacy during cyclic aggregation and offers an effective solution to achieve a more secure cyclic aggregated learning framework.