Anti-collusion Byzantine robust privacy protection federated learning optimization method

The invention relates to the technical field of federated learning and privacy protection, in particular to an anti-collusion Byzantine robust privacy protection federated learning optimization method. The method is composed of a malicious gradient screening mechanism and an anti-collusion Byzantine robust aggregation strategy. Firstly, clients participating in federated learning training select different random numbers to blind local gradients by using a malicious gradient screening mechanism. And then the central server can calculate cosine similarity under the condition of not knowing the local gradient plaintext to screen malicious gradients, and data privacy and model robustness are both considered. Secondly, designing an anti-collusion Byzantine robust aggregation strategy, and weighting local gradients according to cosine similarity calculated by a malicious gradient screening mechanism; and finally, performing weighted aggregation on different local gradients by using weights, and ensuring data privac