Continuous Trust Assessment of Power Terminals Based on Conditional Variational Autoencoders

The increasing scale of power IoT terminal access, in order to ensure that Web-based services are not damaged, automatic anomaly detection methods are needed to monitor hundreds of thousands or even millions of KPIs, detect anomalies and trigger troubleshooting in a timely manner, and at the same time, the zero-trust architecture concept is applied to the power IoT terminal security protection deployment on the basis of inheriting the existing security protection measures. However, when facing a large number of KPIs in practice, traditional statistical methods and supervised integration methods cannot meet this requirement. Donut, a leading supervised method, yields positive results but lacks sequential modeling capabilities and cannot effectively handle time-related anomalies. This paper introduces the Bagel algorithm, an unsupervised approach designed to address temporal information-related anomalies. Bagel utilizes the CVAE to incorporate temporal information and employs dropout layers to prevent overfitting. Experiments conducted with real data from internet companies demonstrate that Bagel enhances the top F1-score for anomaly detection by 0.08 to 0.43 when compared to Donut, highlighting its practical utility in anomaly detection.