FedReview: Reviewing and Removing Poisoned Updates in Federated Learning Without a Validation Dataset or Historical Knowledge

Researchers propose a federated learning review mechanism, FedReview, which can identify and remove poisoned updates without requiring a server-side validation dataset or prior client knowledge. Each round, the server randomly assigns some clients as reviewers, who evaluate and rank updates on their own training data, remove suspicious updates by majority vote, and tolerate incorrect rankings by malicious reviewers. Compared with FLTrust, it allows clients to participate flexibly; experiments show that it can train a well-performing global model in adversarial environments.