PPsFLP: End-to-End Privacy Preserved and Secure Federated Learning Paradigm of Medical Images

PPsFLP: End-to-End Privacy Preserved and Secure Federated Learning Paradigm of Medical Images

연구 분야: Analysis

학회: SN Computer Science

Federated learning (FL) is a collaborative process that involves sharing information in a distributed control. Certain healthcare applications meant for prognosis, diagnosis, etc. in a federated setting require the sharing of sensitive information, especially medical images. The hospital data warehouse (data centers) or patients themselves can be the participants in such a federated framework. However, privacy preservation is crucial in such distributed control. This paper aims to provide a solution to the privacy preservation of medical images in FL. Firstly, the medical image obfuscation is guarded with a hashing key distribution that serves as seed values for the conservative S-modal robust chaos. Further, for authentication of the valid participants of learning, the authentication data is appended in the owner-specific encrypted images as secret shares using Shamir’s secret sharing strategy with a threshold (n, k). The secret shares are further extended to the progressive shares threshold (k1, k2). Most of the progressive shares are appended in the owner’s encrypted images with one essential share with the owner that is shared as consent for its participation in learning. Authentication of participants in the learning process is based on the successful recovery of progressive shares and then the shadow shares of the secret information/image. The decryption of images for learning is possible only after the authentication and consent of the owner. This solves the curious but honest server such that even if the data center is aware of most of the keys, still the information cannot be retrieved on its own. Results and analysis of the encryption stage substantiate the robustness of the proposed algorithm for encryption. The proposed access structure for authentication is discussed for the possibility of any malicious participant in the learning and is proved to be resistant to any such intruder as well as the honest but curious server.