An Unsupervised Convolutional Adversarial Anomaly Detection Model for IoT Data Infrastructure

Peian Wen, Zhenyu Yang, Lei Wu, Sibo Qi, Juan Chen

Abstract: Anomaly detection is critical to ensure the IoT (Internet of Things) data infrastructures' Quality of Service. However, due to the complexity of incon-spicuous(indistinct) anomalies, high dynamicity, and lack of anomaly labels in the operational IoT systems and cloud infrastructures, multivariate time series anomaly detection becomes more difficult. Existing approaches are rarely effective in meeting these challenges. In this paper, We propose a novel convolutional adversarial model based on CNN adversarial training with POT (Peak Over Threshold) dynamic threshold selection (CAT-POT) for multivariate time series anomaly detection in the IoT data. We build on the original autoencoder, replacing the fundamental linear changes with a convolutional neural network and adding a convolutional decoder to make an adversarial training architecture. The ability of the model to identify “slight” anomalies is enhanced through adversarial training, combined with POT dynamically selected thresholds to detect anomalies. We conduct ex-periments on five subsets of various open datasets, and the overall performance of CAT-POT is better than that of the baseline method. The average F1 value of CAT-POT on all datasets is 0.889, which is 5.1% higher than the suboptimal model, with the best improvement reaching 21.5%.