Unsupervised Anomaly Detection for IoT Data based on Robust Adversarial Learning

Yan Qiao, Benchu Zhang, Zeyu Zhang

Abstract: Anomaly detection based on Generative Adversarial Networks (GANs) has been demonstrated its superiority on in imaging area. However, most existing GAN-based methods require noise-free datasets to accurately train the network, whereas neither collecting an absolute clean dataset or labeling the anomalies in the dataset are quite difficult for Internet of Things (IoT) applications due to the harsh environment and the high dimensionality of the data. In this paper, we propose a robust GAN-based model for anomaly detection in IoT, which can accurately learn the data pattern unsupervisedly even with polluted training data. The new model adopts the framework of Bidirectional GAN (BiGAN) to enable an efficient representation mapping, and integrates the mechanism of Robust Principal Component Analysis (RPCA) to rule out the noises of data from the training set. To improve its performance, a new objective function and scoring function are elaborately designed, while a proximal method and Alternating Direction Method of Multiplier (ADMM) are incorporated into the training process. Comparing our method with four state-of-the-art anomaly detection methods, the experimental results strongly confirm the superiority of our method with polluted IoT datasets.