As computer networks expand, security and unauthorized access prevention have become increasingly critical. Traditional Intrusion Detection Systems (IDS) generally fall into two categories: misuse detection (searching for known "signatures" of attacks) and anomaly detection (flagging behavior that deviates from a user's established norm).
However, traditional anomaly detection faces significant hurdles. Hand-coding rules for "normal" behavior is labor-intensive and practically impossible to maintain. Furthermore, many systems attempt to predict the sequence of commands in real-time. Because the order in which a user opens a mail client, runs a compiler, or checks a directory is highly variable, sequence prediction is computationally expensive and prone to high error rates.
To address these challenges, researchers introduced NNID (Neural Network Intrusion Detector). By shifting the focus from real-time sequence prediction to offline statistical distributions, NNID successfully identifies computer users by their "command fingerprints," achieving high accuracy with minimal false alarms.
The NNID Approach: Distribution Over Sequence
The core insight of NNID is that what a user does is a much more stable identifier than the order in which they do it. For instance, a programmer will consistently use gcc and make, while an administrator might frequently use su and chmod.
To capture this, the authors proposed an offline approach using a standard three-layer backpropagation neural network. Instead of tracking when a command was typed, the system aggregates how often it was used over a 24-hour period to build a "User Vector."
Constructing the User Vector:
* Feature Selection: The system tracks the 100 most common commands on the system (e.g., ls, gcc, emacs, mail).
* Non-linear Scaling: To prevent high-frequency commands from overwhelming the model, command counts are mapped into 11 non-linear intervals. For example, the difference between using a command 0 times versus 2 times is more "telling" than the difference between 500 and 502 times. These intervals are represented as values from 0.0 to 1.0.
* Network Architecture: The NNID utilizes 100 input units (one for each tracked command), 30 hidden units, and 10 output units (one for each authorized user on the test system).
Because the model only looks at command names and frequencies—ignoring arguments like filenames or email recipients—it provides a robust layer of security without invasive monitoring of private data.
Experimental Setup and the "Digital Print"
The study utilized real-world data collected over 12 days from a NetBSD machine at the University of Texas at Austin. The dataset comprised 10 users, ranging from heavy daily users to intermittent ones, totaling 89 "user-days" of data. Because the system was "clean" (containing no actual intrusions), the authors generated 100 random user vectors to simulate anomalous behavior. These vectors maintained the same first-order statistics as the real data but lacked the specific correlations of a real user's fingerprint.
The authors liken user behavior to a thumbprint left at a crime scene. You don't need to watch a person leave the print to know they were there; you just need to recognize the pattern they left behind. At the end of each day, the system administrator runs the logs through NNID.
If the network cannot confidently match the day's activity to the authorized user's established output profile, or if it strongly suggests the wrong user, it flags a potential security breach.
Key Results
The NNID demonstrated that user profiles are highly distinct and learnable by a neural network:
- High Anomaly Detection: The system rejected 63% of random vectors outright. When factoring in the low probability (1/10) of an intruder accidentally mimicking a specific user perfectly, the overall anomaly detection rate reached 96%.
- Accurate Identification: The model correctly identified legitimate users 93% of the time.
- Low False Alarms: The system maintained a 7% false alarm rate. Notably, almost all false alarms were triggered by a single user who only had three days of training data, highlighting that the model's accuracy is directly tied to the consistency and depth of the training history.
Limitations and Future Work
While highly effective on a small scale, the authors note a few limitations:
- Scalability: The experiment was conducted with only 10 users. In a corporate environment with thousands of users, the network would need to make much finer distinctions, which could increase the false alarm rate or require a more complex architecture.
- Behavioral Drift: User behavior naturally changes over time. A student might spend a month coding heavily and the next month writing a thesis. To combat this, the model requires periodic retraining—a task that takes only about 90 seconds on their test system.
- Infrequent Users: Users with very little data are difficult to profile accurately, posing a challenge for "guest" accounts or occasional collaborators.
Ultimately, NNID proves that learning offline statistical user profiles is a highly effective, privacy-respecting method for detecting network intrusions.
Summary updated