Stealing Client Data in Federated Learning: A Novel Attack Evading Detection

A novel attack framework called SEER can effectively steal client data from federated learning models, even with large batch sizes and secure aggregation, while avoiding detection by clients.