IoT-FKGDL-SL: Anomaly Detection Framework Integrating Knowledge Distillation and a Swarm Learning for 5G IoT
Lun Tang, Enqiao Kou, Weijia Zhang, Qianlin Wu, Qianbin Chen · IEEE Internet of Things Journal · 2024
Anomaly detection using multivariate time series (MTS) is critical for detecting abnormal traffic and device failures in 5G Internet of Things (IoT) devices. The current anomaly detection framework lacks the ability to model multidimensional long time series and to address issues, such as resource overhead, privacy protection, and data security in distributed learning modes within the IoT. Therefore, this article proposes an anomaly detection framework integrating knowledge distillation and swarm learning for 5G IoT (IoT-FKGDL-SL). First, to model the correlations between different variables, a new method for capturing correlations between variables through clustering is proposed. Second, to perform long-term modeling of MTS, a long-time-series anomaly detection model called IoT-FKGD is proposed, based on multiscale dilated convolution and locality-sensitive hashing (LSH) attention. Finally, a framework based on IoT-FKGD is proposed to detect traffic anomalies of IoT devices under a swarm learning architecture that incorporates knowledge distillation. The effectiveness of the IoT-FKGDL-SL framework is demonstrated by comparing it with advanced anomaly detection methods on real data sets. Experimental results show that on a long time scale, the precision, recall, and F1-score of anomaly detection using this framework all surpass those of baseline methods.