Privacy-Preserving Data Sharing Scheme with FL via MPC in Financial Permissioned Blockchain
Jingwei Liu, Xinyu He, Rong Sun, Xiaojiang Du, Mohsen Guizani
Abstract
Each bank has different clients and each client may have transactions with multiple banks. Hence, clients’ data in a single bank may be partial and incomplete. If the data can be combined, each bank obtains comprehensive information, so as to better carry out business and enhance the quality of service, such as recommending financial products and inquiring about personal credit records. However, after the promulgation of GDPR by European Union in 2018, it is illegal to directly consolidate data crossing enterprises due to privacy and security concerns, especially for privacy-sensitive industries. Emerging federated learning(FL) is very suitable for secure data sharing for distributed banks in privacy. To prevent from connection of clients’ data and the certain bank, we adopt anonymity mechanism to hide the real identity of banks. In this paper, we first propose blockchain-empowered secure federated learning for distributed banks based on multi-party computation(MPC) with multi-key fully-homomorphic encryption(FHE) scheme. Then, we give detailed description of multi-key FHE based MPC protocol, anonymity mechanism and permissioned blockchain consensus protocol. Finally, we analyze the security and compare our scheme with several existed schemes. Numerical results show that the proposed data sharing scheme has good performance in terms of computational overhead and model accuracy.