Parameters security strategy formulated by hyperchaos in federal learning
摘要
Federated Learning (FL) is a machine learning framework that effectively provides multiple organizations with data usage and model training while meeting the requirements of privacy protection, data security and government regulations. However, the leakage of parameters frequently occurred in the process of their exchange. Existing parametric encryption algorithms, such as differential privacy and homomorphic encryption, can significantly elevate privacy protection levels but always bring negative effect to the convergence performance of the final model or huge time consuming for practical application. Therefore, this paper propose a new encryption and decryption algorithm based on the hyperchaotic system, called bit and parameter correlation permutation (BCP), which helps to protect the parameters through the upload and download process, and such algorithm is compatible with any hyperchaotic map. With this method, we could guarantee the security without any sacrifices of model accuracy in a shorter period of time. Finally, the proposed algorithm achieves a time complexity of O(