TL;DR
This paper introduces a fully unsupervised system for detecting insider threats by combining Word2Vec for automatic feature extraction and Bayesian Gaussian Mixture Models (BGMM) for behavior modeling. By training individual models for every user, the system achieves an 88% recall rate and a low 6.9% false positive rate without requiring labeled data or "clean" training sets.
What Problem Does This Paper Address?
Insider threats—malicious or negligent actions by authorized employees—are a nightmare for cybersecurity. Unlike external hackers, insiders already have the keys to the kingdom and the domain knowledge to hide their tracks.
Detecting them is a "needle in a haystack" problem: malicious actions often look like normal work, and they represent a tiny fraction (often <0.1%) of total activity. Most existing solutions rely on supervised learning, which requires historically labeled data that most companies simply don't have. Others require "One-Class" training, assuming the initial training data is 100% threat-free—a risky assumption in a real-world environment. This paper seeks a flexible, unsupervised solution that requires minimal domain knowledge to implement.
Methods
The authors propose a multi-stage pipeline that treats user activity like a language:
- User-Based Data Extraction: Activity logs (logons, emails, file access, etc.) are aggregated by user and grouped into daily "sentences" of events.
- Automatic Feature Extraction (Word2Vec): Instead of manually defining features, the authors use a Word2Vec Skip-gram model. This treats activity types as "words" and learns their context. For example, it captures the fact that "reading sensitive files" followed by "sending an external email" is a specific behavioral pattern. These are summed into a daily summary vector.
- Behavior Modeling (BGMM): The core innovation is using Bayesian Gaussian Mixture Models. While standard GMMs require the developer to pre-set the number of clusters ($K$), BGMM uses a Dirichlet process (often explained via a "stick-breaking" analogy) to automatically infer the right number of behavioral clusters for each specific user.
- Anomaly Scoring: The system calculates a log-likelihood score for each day. If a day’s activity is a statistical outlier compared to that specific user’s learned "normal" clusters, it is flagged as a potential threat.
Datasets
The study uses the CERT Insider Threat Dataset version 4.2, a standard benchmark in the field.
- Scope: 1,000 simulated employees over 502 days.
- Scale: Over 32 million events across five domains (Logon, Email, File, HTTP, and Device).
- Imbalance: Only 70 users are malicious, and malicious instances account for a mere 0.03% of the total data.
Key Insights and Ideas
- One Size Does Not Fit All: The "user-based" framework is critical. A system administrator’s normal day looks like a security breach for an accountant. By modeling users individually, the system reduces noise.
- The Power of Non-Parametric Clustering: By using BGMM, the model adapts to the complexity of a user's job. A multi-role manager might have five distinct "normal" modes of behavior, while a data entry clerk might only have one. BGMM identifies this automatically, preventing the underfitting or overfitting common in static models.
- Sequence Matters: Using Word2Vec allows the model to capture the temporality of events without the massive computational overhead of Recurrent Neural Networks (RNNs). It learns the "grammar" of a workday.
- Real-World Viability: The model does not require "data balancing" (oversampling the rare malicious events), making it much easier to deploy in environments where the ground truth is unknown.
Key Results
The proposed method was compared against several state-of-the-art techniques (like LSTM-Autoencoders and Isolation Forests) using the same CERT dataset:
- Recall: 88.38% (successfully identifying the vast majority of malicious days).
- False Positive Rate (FPR): 6.9%. This is a significant improvement over many deep learning approaches (e.g., the LSTM-Autoencoder in the study had a 9.84% FPR).
- Accuracy: 93.08%.
- AUC (Area Under Curve): 0.958, outperforming the 0.949 achieved by recent LSTM-based models.
Notably, these results were achieved using only the type and order of activities, ignoring more complex data like file content or specific timestamps.
Limitations and Future Work
- Feature Depth: The current model only looks at the type of activity. The authors acknowledge that incorporating the time of day, day of the week, or the actual content of emails/web searches could further sharpen detection.
- Real-Time Processing: The current embedding generation (Word2Vec) is performed on historical data. The authors aim to adapt the daily embedding process for real-time, streaming execution.
- Evolving Environments: While BGMM is flexible, future work is needed to study how the model handles "concept drift"—when a user’s job description or habits change legitimately over time.
Summary updated