A Framework for Moving Target Defense based on Federated Semi-Supervised Learning

Abstract

In recent years, with the rapid development of the Internet, it has penetrated into all areas of daily lives. However, due to the complexity of the Internet architecture, there are inevitably some inherent security threats, which could be exploited by adversaries to cause great damage. Moving Target Defense (MTD) has been proposed to solve this problem by building a dynamic, heterogeneous and redundant system architecture. Unfortunately, most of the existing data arbitration algorithms for MTD are based on the majority consensus voting algorithm, which cannot cope with common mode escape. Therefore, in this paper, we propose a framework for moving target defense based on Federated Semi-Supervised Learning (FSSL), called FedDA. In addition to the output data of heterogeneous executives, FedDA leverages their behavior data to assist in data arbitration. Meanwhile, we consider a more realistic assumption that the behavior data of heterogeneous executives is not annotated with ground-truth lables, and FSSL is used for model training. Finally, a data arbitration algorithm combined with historical confidence is proposed to identify malicious executives. Extensive experimental results show that our method can well resist common mode escape and outperform the state-of-the-art data arbitration algorithms.