FGLIoT: IoT device identification via federated graph learning and spatio-temporal feature fusion

Abstract

The device silo problem poses a significant challenge to the management and security of the Internet of Things (IoT). The key to solving this issue is to accurately identify IoT devices connected to the network while protecting data privacy. However, existing solutions overlook inter-packet semantic correlations, a fact which renders them unable to fully explore the potential behavior patterns in device communication traffic. Therefore, we propose FGLIoT, a federated graph learning-based method for IoT device identification. FGLIoT first represents the communication traffic data generated by IoT devices as packet sequence graphs, preserving the semantic information of packets. It then employs a graph learning module to capture inter-packet semantic correlations and learn representations of device communication behaviors. Subsequently, the representations are processed by spatial and temporal feature extractors to capture their spatial correlations and temporal dependences, respectively. Finally, residual connections are used to fuse the behavior representations with their spatial and temporal features, generating behavioral fingerprints for IoT device identification. Experimental results on three public IoT device datasets demonstrate the effectiveness of FGLIoT in solving the device silo problem.