Attention-Based LSTM for Insider Threat Detection

Abstract

Insider threat is an important cyber security issue for businesses and organizations. Existing insider threat detection methods can be roughly divided into two categories, statistical features based detection methods and action sequence based detection methods. The first kind of method aggregates all actions that a user has performed over one day and uses these aggregated features to find insider threat. This kind of coarse-grained analytics of user behavior may miss anomalous behavior happening within that day. The second kind of method overcomes the coarser-grained problem and uses fine-grained detection to identify insider threat through user actions. However, the second kind of method considers all user operations to be equally important, without highlighting malicious user actions. To solve this problem, we present an attention-based Long Short-Term Memory (LSTM) model to detect insider threat. In our model, we apply the LSTM to capture the sequential information of user action sequence and employ an attention layer that can learn which user actions contribute more to insider threat detection. Extensive studies are conducted on the public dataset of insider threat. Our results demonstrate that the proposed model outperforms other deep learning models and can successfully identify insider threat.