Ensemble Strategy for Insider Threat Detection from User Activity Logs
Shihong Zou, Huizhong Sun, Guosheng Xu, Ruijie Quan
Abstract
In the information era, the core business and confidential information of enterprises/organizations is stored in information systems. However, certain malicious inside network users exist hidden inside the organization; these users intentionally or unintentionally misuse the privileges of the organization to obtain sensitive information from the company. The existing approaches on insider threat detection mostly focus on monitoring, detecting, and preventing any malicious behavior generated by users within an organization’s system while ignoring the imbalanced ground-truth insider threat data impact on security. To this end, to be able to detect insider threats more effectively, a data processing tool was developed to process the detected user activity to generate informationuse events, and formulated a Data Adjustment (DA) strategy to adjust the weight of the minority and majority samples. Then, an efficient ensemble strategy was utilized, which applied the extreme gradient boosting (XGBoost) model combined with the DA strategy to detect anomalous behavior. The CERT dataset was used for an insider threat to evaluate our approach, which was a real-world dataset with artificially injected insider threat events. The results demonstrated that the proposed approach can effectively detect insider threats, with an accuracy rate of 99.51% and an average recall rate of 98.16%. Compared with other classifiers, the detection performance is improved by 8.76%.