FL-finder: Detecting Unknown Network Anomaly in Federated Learning

Abstract

The emergence of federated learning has ensured data and privacy security in deep learning models while enabling models to train more efficiently. However, the transmission of network parameters in federated learning may be subject to attacks by unknown anomalies. In this paper, we attempted to detect unknown anomalies in transmitted parameters in federated learning. We designed and implemented F1-finder, an unknown network anomaly detection framework in federated learning, which detects anomalies based on incremental learning. It retains the unknown anomalies to its prior knowledge base using the network updater, and adopts an online mode that reports new anomalies in a real-time. Extensive experimental results show that our model increased the average accuracy of unknown anomaly detection by 10.4% and the average F1-Score improved to 19%.