Introduction
Network administrators often manage vast quantities of log data—including logins, file accesses, web traffic, and emails—without having access to the underlying "ground truth" metadata about the users themselves. Identifying a user’s organizational role (e.g., IT admin, HR, Engineer) is critical for several operational reasons. First, it enhances security by allowing administrators to define "normal" behavior per role, making the identification of anomalies—such as a salesperson accessing sensitive database schemas—more straightforward. Second, it aids in resource management by helping anticipate demand for specific tools. Finally, it allows for better environment customization based on actual user needs.
While user profiling has been explored in social media contexts using linguistic data, corporate environments present a unique challenge because the data is transactional and less structured. This research bridges that gap by treating a user's daily activity as a temporal sequence, utilizing Recurrent Neural Networks (RNNs) to predict roles from log patterns.
Methodology
The core of the approach is a stacked Long Short-Term Memory (LSTM) network. Unlike standard neural networks that process data points in isolation, LSTMs feature memory cells that capture temporal dependencies over extended periods.
- Input Representation: Each user is represented by a sequence of daily feature vectors ($x_1, x_2, \dots, x_T$). Each vector is 408-dimensional, summarizing activities across five categories: logons/logoffs, web traffic, file interactions, emails, and external device usage.
- Temporal Accumulation: The recurrent nature of the model allows the prediction for any given day $t$ to be informed by the user's entire history from day 1. This enables the model to "build confidence" as it accumulates evidence over time.
- Optimization: The researchers employed a stacked architecture and optimized hyperparameters—including layer count, hidden state dimensionality, and learning rate—using a random search strategy.
Dataset
The study utilized the CERT Insider Threat v6.2 dataset, a sophisticated simulation containing over 135 million log lines from 4,000 users over 516 days. For this analysis, the authors focused on the first 90 days of data. The users were categorized into 33 distinct roles (after merging rare roles into a catch-all category). The data was split into training (80%), development (10%), and test (10%) sets.
Results and Discussion
The model achieved a 38% classification accuracy on the 33-way task, a significant improvement over the 11% majority-class baseline. This represents a 30% relative reduction in role classification error compared to initial baseline models.
A key finding is the relationship between time and model confidence. As shown in Figure 1, the model's error (measured by cross-entropy) drops sharply during the initial weeks and continues to improve until approximately day 40, after which performance plateaus.
This "40-day horizon" suggests that roughly six weeks of log data is the optimal window required to distinguish professional roles in this environment. Initially, the model has high cross-entropy due to a lack of context; as more daily feature vectors are processed, the probability distribution over roles becomes increasingly peaked and accurate.
Conclusion
This research demonstrates that a user's professional role is defined not by isolated actions, but by the rhythmic patterns of their behavior over time. While the current model uses 408 hand-crafted features, future work will explore automated feature extraction and techniques to address class imbalance, such as resampling or replay learning. Additionally, the researchers plan to apply regularization methods like dropout to improve generalization and test the framework on non-simulated, real-world datasets. The methodology is designed to be extensible, potentially generalizing to other user attributes such as age or seniority levels.
Summary updated