Sensitive behavior detection for intelligent measurement terminal applications based on domain adversarial neural networks

Abstract

Intelligent measurement terminals (IMTs) are key components in the power grid metering system. Detecting sensitive behaviors of the applications (Apps) installed on the IMTs is an effective means to prevent malicious behaviors and ensure the safety of the power grid. However, the traditional deep learning-based methods for sensitive behavior detection often suffer from performance degradation when the Apps change their behavior pattern after porting, migration, or upgrading. Retraining the network requires large quantities of labeled data, which is usually time-consuming and costly in practice. This paper proposes a new sensitive behavior detection method for IMT Apps based on the domain adversarial neural network (DANN). First, the DANN is utilized to extract the shared features of App behaviors before and after porting, migration, or upgrading, enabling the original data to be repurposed for training the new model, thereby alleviating the insufficient training data problem. Then, the distinctive features of the sensitive behavior in the new environment are extracted and fused with the shared features in the detection process to further improve the detection accuracy. Experimental results show that the proposed method attains 91.83% in detection accuracy, 91.21% in precision, 92.61% in recall, and 91.91% in F1-score, which are better than several traditional deep learning-based methods.