Deep neural network for text anomaly detection in SIoT

Abstract

Unsupervised textual anomaly detection, which discovers anomalies from unlabeled texts, is critical to improve the cybersecurity and interaction ability among the objects in the Social Internet of Things (SIoT). Recently, detecting anomalies by deep neural networks has become a popular trend. Specially, context vector data description (CVDD) method shows the promising performance. However, CVDD has two limitations: (1) it uses an one-class classification objective to constrain the sentence embeddings, which leads the learned embeddings to lose content information of text. (2) Scalar-based attention weights, which are used to extract sentence features, fail to focus on dimensional properties in a word. Learning the text contents and the dimensional properties is important for detection task in SIoT, which can help detector capture the difference between normal and anomaly texts. To overcome these limits, this paper proposes a textual anomaly detection network. First, an adversarial training strategy is designed to fight against the problem of missing content information. Second, a textual anomaly detection module with multiple dimensional transformation matrices is constructed to learn dimensional properties of words in diverse semantic subspaces. Experimental results on several textual datasets show that our proposed method outperforms CVDD and other strong baselines.