Interactive Learning for Network Anomaly Monitoring and Detection with Human Guidance in the Loop

With the advancement in big data and cloud computing technology, we have witnessed tremendous developments in applying intelligent techniques in network operation and management. However, learning- and data-based solutions for network operation and maintenance cannot effectively adapt to the dynamic security situation or satisfy administrators' expectations alone. Anomaly detection of time-series monitoring indicators has been a major challenge for network administrative personnel. Monitored indicators in network operations are characterized by multiple instances with high dimensions and fluctuating time-series features and rely on system resource deployment and business environment variations. Hence, there is a growing consensus that conducting anomaly detection with machine intelligence under the operation and maintenance personnel's guidance is more effective than solely using learning and modeling. This paper intends to model the anomaly detection task as a Markov Decision Process and adopts the Double Deep Q-Network algorithm to train an anomaly detection agent, in which the multidimensional temporal convolution network is applied as the principal structure of the Q network and the interactive guidance information from the operation and maintenance personnel is introduced into the procedure to facilitate model convergence. Experimental results on the SMD dataset indicate that the proposed modeling and detection method achieves higher precision and recall rates compared to other learning-based methods. Our method achieves model optimization by using human-computer interactions continuously, which guarantees a faster and more consistent model training procedure and convergence.