Transformer-Based Security Anomaly Detection for Wireless Sensor Networks

Abstract

Although Wireless Sensor Networks (WSNs) are already being used in the development of cyber-physical systems, they are susceptible to several security risks due to their open architecture and few resources. In order to handle high-dimensional time-series data and different kinds of assaults, this study introduces a security anomaly detection technique for WSNs based on Transformer architecture. Eleven different types of attacks and over 240,000 labeled time-series windows were created by merging real data from several sensors and simulated attack situations in multiple networks. Richer information can be extracted from sensor nodes and sequential data using a Multi-Head Self-Attention-based Model and Spatial-Temporal Feature Encoding. With an average area under the ROC curve (AUC) of 0.976, an F1-score of 0.942, and more than 93% recall on an unbalanced test set, the Transformer-based framework outperforms the basic model, Long Short-Term Memory network, and Isolation Forest, according to the experiment results. Additionally, the detection rate decreases by less than 7% under adversarial attacks or increased signal noise, and the system operates steadily in the presence of noise and variations in the environment or parameters. As a result, it is evident that the Transformer approach has improved detection stability and accuracy and is still viable for high-speed use in WSN defense. In order to enhance the security of wireless sensor networks for monitoring vital infrastructure and intelligent environments, this study suggests an effective and comprehensive solution.