Research on Gradient Selection Grid Data Federation Learning System Based on Top-K Multi-Secret Sharing

The Top-K asymptotic combined with secret sharing to design a federated learning protocol, which can protect users' privacy and reduce communication costs. First, the algorithm uses the Top-K gradient selection algorithm to filter the required input gradients to reduce the number of required input gradients. A candidate quantization scheme and index fusion algorithm suitable for multiple boundaries are designed. The noisy data is distributed to multiple servers using a key-sharing mechanism. While ensuring the security of information transmission, balancing multiple servers' configurations maximizes the entire network's operational efficiency. The configuration of multiple servers maximizes the operational efficiency of the entire network. The experimental results show that the method studied in this project ensures the confidentiality of user information and improves its accuracy and execution speed.