A Modular Multiple Classifier System for the Detection of Intrusions in Computer Networks

Giorgio Giacinto, Fabio Roli, Luca DidaciMultiple Classifier Systems 2003 · DOI

TL;DR

This paper proposes a novel modular Multiple Classifier System (MCS) architecture for network intrusion detection. By breaking down the problem into service-specific modules and fusing diverse feature sets, their approach significantly improves upon existing methods, achieving better performance in terms of classification cost and false alarm rates, making it more practical for real-world deployment.

What Problem Does This Paper Address?

Computer networks are constantly under threat from evolving cyberattacks. Intrusion Detection Systems (IDS) are crucial for identifying these threats, but they face a fundamental challenge: balancing the ability to detect novel, unknown attacks with the need to minimize false alarms. Traditional signature-based IDS are good at detecting known attacks but struggle with new variants. Anomaly detection, which relies on statistical profiles of normal traffic, can detect novel attacks but often generates a high rate of false alarms, which is problematic given the sheer volume of normal network traffic. This paper aims to bridge this gap by developing an IDS that can effectively detect a wider range of attacks, including novel ones, while keeping false alarms at an acceptable level.

Methods

The core of this paper is the proposal of a Modular Multiple Classifier System (MCS). Instead of a single, monolithic IDS, they advocate for a system composed of specialized modules, each dedicated to a specific network service (e.g., HTTP, FTP, Mail). This modularity is based on the observation that different services exhibit distinct traffic patterns, and thus, attacks targeting them will also have unique characteristics.

Within each service-specific module, the system employs an MCS that fuses information from multiple classifiers. These classifiers are trained on distinct feature sets extracted from network connections. The paper categorizes these features into three types:

  • Content features: Information from packet payloads.
  • Intrinsic features: Characteristics of the current connection (e.g., duration, protocol).
  • Traffic features: Statistics derived from past connections (e.g., number of connections to the same destination).

The fusion of these diverse feature representations is key. The authors explore several decision fusion functions, including simple majority voting and average of classifier outputs, as well as more sophisticated trainable methods like Naive-Bayes and decision templates. This allows the system to leverage the complementary strengths of classifiers trained on different data perspectives.

Datasets

The experiments were conducted using a subset of the DARPA 1998 Intrusion Detection Evaluation Program database, specifically the version pre-processed by Columbia University and distributed as part of the UCI KDD Archive (kddcup99). This dataset contains network connections labeled as either normal traffic or belonging to one of five classes: Denial of Service (DoS), Remote to Local (R2L), User to Root (U2R), and Probing attacks. The dataset features a 41-dimensional vector for each connection. The authors mention using a "balanced" training set to address class imbalance issues, particularly with high-volume attack types like "Smurf" and "Neptune."

Key Insights and Ideas

  • Service-specific modules are crucial: Recognizing that different network services have unique traffic patterns and attack vectors is fundamental. A one-size-fits-all approach is less effective than specialized modules.
  • Feature diversity is power: Training classifiers on distinct feature sets (content, intrinsic, traffic) and then fusing their outputs allows the system to capture a broader range of attack characteristics that might be missed by any single feature set alone.
  • Fusion is more than just combining: The choice of fusion rule significantly impacts performance. While simple rules can work, trainable fusion methods can better adapt to the varying accuracies and correlations of individual classifiers.
  • MCS offers a better trade-off: The MCS approach, by combining diverse perspectives, provides a more effective balance between detecting novel attacks (generalization ability) and minimizing false alarms, a critical requirement for practical IDS.

Key Results

The paper demonstrates that their proposed modular MCS architecture outperforms previous state-of-the-art methods, particularly the winner of the KDD99 contest.

  • Improved Performance Metrics: The modular MCS achieved better results in terms of classification cost and the sum of false and missed alarm rates. For instance, one configuration of the modular MCS achieved a classification cost of 0.1688, significantly lower than the best KDD99 result of 0.2331.
  • Reduced False Alarms: The system shows a marked reduction in false and missed alarms, which is a critical factor for operational acceptance of IDS.
  • Modular Advantage: The modular approach, when optimized for each service, consistently outperformed a single classifier trained on the entire feature set.
  • Fusion Rule Effectiveness: The majority voting fusion rule proved effective for most services, suggesting that classifiers trained on independent feature sets can be treated as largely independent.

Limitations and Future Work

  • Dataset Specificity: The proposed division of services into groups and the feature selection process are tailored to the specific dataset used (KDD99) and the network environment it represents. The authors acknowledge that this division might need to be adapted for different networks.
  • Feature Selection Dependency: The effectiveness of the system relies on a robust feature selection process, which is described as using a forward selection algorithm. The performance could be sensitive to the quality of this selection.
  • Future Work: The paper suggests that the modular MCS architecture represents a significant step forward. Future work could involve further exploring different fusion techniques, optimizing the feature selection for various network environments, and potentially extending the modularity to finer-grained service classifications. The authors also highlight the potential for this approach to increase the acceptance of pattern recognition techniques in network security.

Summary updated

BibTeX
@inproceedings{giacinto2003modular,
  author = {Giorgio Giacinto and Fabio Roli and Luca Didaci},
  title = {{A Modular Multiple Classifier System for the Detection of Intrusions in Computer Networks}},
  booktitle = {Multiple Classifier Systems},
  year = {2003},
  doi = {10.1007/3-540-44938-8\_35},
}
Figure