Decentralized Federated Learning for Intrusion Detection in IoT-based Systems: A Review

Abstract

Internet of Things-based systems are typically distributed systems and thus inherit all issues related to the need to guarantee confidentiality, integrity, and availability. Moreover, IoT-based systems are vulnerable to several attacks, mainly due to the weakness of IoT devices, which have little computational and memory power, necessary for more sophisticated security features. To build robust infrastructures, one of the traditional strategies to deal with these problems involves intrusion detection and prevention techniques. Implementing them in a centralized way is usual, which leads to not being scalable for IoT systems with an increasing number of devices. Moreover, it implies an unacceptable single point of failure. Besides, sending all collected data to a centralized server in the cloud poses a significant risk to the privacy of information. Recently, machine learning techniques, as decentralized federated learning, combined with distributed ledger technologies, have been used to implement more robust and privacy-preserving intrusion detection systems for IoT-based systems. However most current decentralized federated learning approaches depend on a centralized server, and so have a single point of failure. The centralized server is used to aggregate the local trained models obtained by means of deep learning algorithms performed on edge and fog devices. This paper reviews state of the art on intrusion detection based on totally decentralized federated learning and distributed ledger techniques applied to minimize IoT security threats, identifies open problems, and recommends future research directions to cope with them.