Hunter in the Dark: Discover Anomalous Network Activities Using Deep Ensemble Networks

Machine learning (ML)-based network intrusion detection system (NIDS) plays a critical role in discovering unknown and novel threats in a large-scale cyberspace. It has been widely adopted as a mainstream hunting method in many organizations, such as financial institutes, manufacturing companies and government agencies. However, there are two challenging issues in the existing designs: 1) achieving excellent performance of threat detection is often at the cost of a large number of false positives, leading to the problem of alert fatigue and 2) the interpretability of detection results is low, making it difficult for the security analyst to obtain the insight of threats and take promptly actions against the attacks. To tackle the above issues, in this paper we propose a defense mechanism, DarkHunter, that includes three parts: stream processor, detection engine and incident analyzer. The stream processor converts raw network packet streams into data records of a set of statistical features that can be effectively used for learning; The detection engine leverages an efficient ensemble neural network (EnsembleNet) to identify anomalous network traffic; The incident analyzer applies a correlation analysis to filter out the mis-predictions from EnsembleNet, traces each detected threat from its statistical representation back to its source traffic flow to enhance its intelligibility and prioritizes the threats to be responded to minimize security risks. Our evaluations, based on the UNSW-NB15 testbed, show that DarkHunter significantly outperforms state-of-the-art ML-based NIDS designs by achieving higher accuracy, higher detection rate, higher precision, higher F1 score while keeping lower false alarm rate.