Enhanced Anomaly Detection in IoT Through Transformer-Based Adversarial Perturbations Model
Sahar Zia, Nargis Bibi, Samah Alhazmi, Nazeer Muhammad, Afnan Alhazmi · Electronics · 2025
Ensuring data security in IoT systems requires effective anomaly detection, particularly in multivariate time series data generated by sensor networks. This study introduces a transformer-based method to detect anomalies by capturing complex temporal patterns and long-range dependencies. The model adapts to diverse anomaly types across datasets, leveraging adversarial perturbations to enhance robustness and accuracy. Integration of the Streaming Peaks Over Threshold (SPOT) mechanism further improves thresholding. Experiments on MSL, SMD, NAB, and SWaT datasets validate the model’s effectiveness, demonstrating its competitive performance in strengthening IoT systems and ensuring data security in dynamic environments.