An Immune Inspired Approach to Anomaly Detection

Jamie Twycross, Uwe Aickelin

Abstract

The immune system provides a rich metaphor for computer security: anomaly detection that works in nature should work for machines. However, early artificial immune system approaches for computer security had only limited success. Arguably, this was due to these artificial systems being based on too simplistic a view of the immune system. We present here a second generation artificial immune system for process anomaly detection. It improves on earlier systems by having different artificial cell types that process information. Following detailed information about how to build such second generation systems, we find that communication between cells types is key to performance. Through realistic testing and validation we show that second generation artificial immune systems are capable of anomaly detection beyond generic system policies. The paper concludes with a discussion and outline of the next steps in this exciting area of computer security.

Introduction: Beyond Simple Pattern Matching

Traditional Intrusion Detection Systems (IDS) are frequently plagued by high false-positive rates and a vulnerability to "zero-day" attacks. While early Artificial Immune Systems (AIS) attempted to address these issues, they often relied on a simplistic model of adaptive immunity (like T-cells) while neglecting the innate immune system's role in providing environmental context. This paper introduces a "second-generation" AIS that mimics the complex, multi-cell interactions of biological immunity to create a more robust and selective anomaly detection mechanism.

The libtissue Framework

A core contribution of this work is the libtissue framework, a C-based, client/server architecture designed to facilitate AIS research. This framework separates the data collection process from the algorithm itself, allowing for greater flexibility and modularity.

The libtissue framework architecture, separating data collection from the AIS algorithm.

As shown in the architecture, "clients" act as sensors on monitored hosts, collecting two types of data:

  1. Antigens: Representations of system behavior, specifically system calls.
  2. Signals: Environmental data such as CPU and memory usage.

These data streams are sent to the libtissue server, where the AIS algorithm operates within a "tissue compartment." This separation allows researchers to easily swap out different data sources or AIS algorithms without re-engineering the entire system.

The twocell Algorithm

Building on the libtissue framework, the authors developed the twocell algorithm. This approach utilizes two distinct artificial cell types to process information:

  • Type 1 Cells (Innate/APC): These emulate Antigen Presenting Cells. They ingest system calls and monitor environmental signals. The "context" provided by these signals (e.g., a sudden spike in CPU load) determines the duration for which they "present" system calls to the rest of the system.
  • Type 2 Cells (Adaptive/T-cell): These cells attempt to bind with Type 1 cells. If a Type 2 cell matches a system call currently being presented by a Type 1 cell, it triggers a response, such as logging an alert or initiating a defensive action.

By using Type 1 cells to "gate" the response based on environmental context, the system becomes sensitive to the effects of an attack rather than just a static signature.

Experimental Validation

The researchers validated their approach using a target system running the rpc.statd daemon, which contains a known "format string" vulnerability. The "antigens" monitored were the specific system calls made by the process.

Examples of system calls monitored as antigens within the framework.

The testing involved six distinct datasets covering successful attacks (resulting in a remote root shell), failed attacks (causing process crashes), and normal usage. This realistic testing environment allowed the authors to compare the twocell algorithm against standard whitelisting policies.

Key Results and Insights

The experimental results demonstrated the effectiveness of the second-generation AIS:

  • High Selectivity: The twocell algorithm was significantly more discerning than a naive whitelisting policy. In successful attack scenarios, the twocell policy denied 52% of system calls, compared to only 9% caught by a standard policy.
  • Low Computational Overhead: Despite the complexity of simulating cellular interactions, the system is lightweight, typically consuming only 1% of CPU resources and never exceeding 3%.
  • Operational Integrity: The system successfully permitted all system calls identified in the "normal" usage baseline, indicating a low risk of disrupting legitimate functionality.

Conclusion and Future Work

This research demonstrates that incorporating environmental context and multi-cellular interactions significantly improves the performance of Artificial Immune Systems in computer security. While the current prototype focuses on system call numbers, future iterations will incorporate system call arguments and instruction pointer addresses to detect more sophisticated evasion techniques. The researchers are also moving toward the "Dendritic Cell Algorithm" (DCA) and testing against larger, noisier datasets to further refine the system's capabilities in complex environments.

Summary updated

Abstract

The immune system provides a rich metaphor for computer security: anomaly detection that works in nature should work for machines. However, early artificial immune system approaches for computer security had only limited success. Arguably, this was due to these artificial systems being based on too simplistic a view of the immune system. We present here a second generation artificial immune system for process anomaly detection. It improves on earlier systems by having different artificial cell types that process information. Following detailed information about how to build such second generation systems, we find that communication between cells types is key to performance. Through realistic testing and validation we show that second generation artificial immune systems are capable of anomaly detection beyond generic system policies. The paper concludes with a discussion and outline of the next steps in this exciting area of computer security.
BibTeX
@misc{twycross2009immune,
  author = {Jamie Twycross and Uwe Aickelin},
  title = {{An Immune Inspired Approach to Anomaly Detection}},
  year = {2009},
  eprint = {0910.3117},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/0910.3117},
}
Figure