Abstract
While most security projects have focused on fending off attacks coming from outside the organizational boundaries, a real threat has arisen from the people who are inside those perimeter protections. Insider threats have shown their power by hugely affecting national security, financial stability, and the privacy of many thousands of people. What is in the news is the tip of the iceberg, with much more going on under the radar, and some threats never being detected. We propose a hybrid framework based on graphical analysis and anomaly detection approaches, to combat this severe cybersecurity threat. Our framework analyzes heterogeneous data in isolating possible malicious users hiding behind others. Empirical results reveal this framework to be effective in distinguishing the majority of users who demonstrate typical behavior from the minority of users who show suspicious behavior.
Introduction
Traditional cybersecurity efforts have long prioritized the "perimeter," focusing on fending off external hackers and unauthorized access. However, the "insider threat"—comprising malicious employees, contractors, or partners—often poses a far greater risk. These individuals already possess authorized access, allowing them to bypass many standard security measures. Detecting these threats requires moving beyond simple log file analysis to a deeper understanding of the complex interrelationships between users, the devices they access, and their underlying psychological profiles.
This paper introduces a hybrid framework designed to identify the "known unknown" of authorized users acting maliciously. By combining graph-based structural analysis with the Isolation Forest anomaly detection algorithm, the system successfully isolates suspicious outliers from the "normal" majority within an organization.
The Proposed Framework
The proposed framework consists of two primary logical components: a Graphical Processing Unit (GPU) and an Anomaly Detection Unit (ADU). It is important to note that in this context, GPU refers to a logical processing unit for graph data rather than hardware acceleration.
The process begins with Graph Construction, where the system builds a large-scale bipartite graph. Nodes represent users and devices, while edges represent "Logoff" events. Logoff events were specifically chosen for their lack of ambiguity compared to logon or unlock events, providing a cleaner signal of user-device interaction.
Feature Engineering and Parameterization
The core of the framework's analytical power lies in its ability to extract 50 distinct parameters across several heterogeneous data categories. These parameters capture both technical behavior and psychological traits.
The extracted features include:
- Graph/Subgraph Properties: The system calculates vertex count, edge count, density, diameter, and the number of peers for subgraphs up to the 5th order of neighborhood. This captures the complexity of a user's digital "neighborhood," revealing clusters of similar behavioral patterns.
- Time-Dependent Parameters: Statistical summaries (minimum, maximum, mean, and mode) for logon/logoff times and USB device usage.
- Web Access Patterns: Tracking the number of unique URLs visited to identify unusual browsing habits.
- Psychometric Data: Scores based on the "Big Five" personality traits (Openness, Conscientiousness, Extroversion, Agreeableness, and Neuroticism), attempting to capture the human element that often precedes a technical attack.
Anomaly Detection via Isolation Forest
These 50 parameters are fed into the Anomaly Detection Unit, which utilizes the Isolation Forest (iForest) algorithm. Unlike traditional methods that attempt to define "normal" behavior and flag deviations, iForest explicitly seeks to isolate anomalies. It operates on the principle that anomalies are "few and different," making it highly efficient for processing large, high-dimensional datasets without the need for a perfect baseline of normal activity.
Experimental Results and Discussion
The study utilized the CERT Insider Threat Dataset (R4.2), a comprehensive "dense needle" dataset containing 17 months of data for 1,000 employees. The analysis focused on three specific job roles: Engineering, Software Management, and Research.
Key findings from the empirical evaluation include:
- Effective Separation: The framework successfully distinguished over 79% of users as demonstrating typical behavior, effectively isolating the small minority of suspicious outliers.
- Anomaly Scoring: By applying a threshold based on maximum deviation, the system identified users whose behavior significantly diverged from their peers.
- Multi-Dimensional Stealth: Interestingly, no users were flagged as suspicious across more than three main parameter categories simultaneously. This suggests that malicious insiders often attempt to blend in across most dimensions, deviating only in specific areas to avoid detection.
Conclusion
This graph-based approach demonstrates that insider threat detection must be behavior-driven rather than purely data-driven. By incorporating higher-order subgraph analysis and psychometric data, the framework captures nuances that traditional tools miss. Future work will focus on incorporating content analysis (e.g., email and web page text), deeper temporal dynamics, and organizational social graphs to further refine detection capabilities.
Summary updated