Biologically Inspired Defenses Against Computer Viruses

IJCAI 1995

TL;DR

This paper introduces biologically inspired techniques for computer virus defense, moving beyond traditional signature-based methods. It presents a neural network for detecting boot sector viruses and a prototype "computer immune system" that automatically analyzes and neutralizes new viruses, aiming to automate tasks traditionally done by human experts.

What Problem Does This Paper Address?

The paper addresses the escalating arms race between computer virus creators and anti-virus software developers. Current anti-virus technology relies heavily on human experts to analyze new viruses, create signatures, and develop removal tools. This process is slow and struggles to keep pace with the daily creation of new viruses. Furthermore, the increasing interconnectedness of computers and the rise of mobile intelligent agents are predicted to accelerate virus spread, making traditional methods obsolete. The authors argue that biological immune systems, with their adaptive and rapid response capabilities, offer a promising model for developing more robust and automated computer defenses.

Methods

The paper details two primary biologically inspired techniques:

  • Neural Network Virus Detector: This approach uses a neural network to learn to distinguish between infected and uninfected programs. Specifically, they focus on boot sector viruses, extracting "trigrams" (3-byte sequences) that are indicative of viral behavior. A single-layer neural network is trained on these trigrams to classify boot sectors. This method aims for a more generic detection than traditional signature matching.
  • Computer Immune System (CIS): This is a more comprehensive, multi-component system designed to automate virus analysis and response. Key components include:
  • Anomaly Detection: Similar to the innate immune system, this uses activity and integrity monitors to flag suspicious behavior.
  • Decoys: Programs designed to attract and infect unknown viruses, providing samples for analysis. This mimics how biological immune systems present antigens.
  • Algorithmic Virus Analysis: Automates the process of dissecting captured virus samples to understand their structure and how they infect hosts.
  • Automatic Signature Extraction: Generates signatures for new viruses based on the analysis, aiming to minimize false positives and negatives.
  • Virus Removal: Develops automated methods to remove identified viruses from infected systems.
  • Kill Signal: A mechanism for infected machines to broadcast information about new viruses to neighboring machines, akin to an adaptive immune response.

Datasets

  • Training/Testing Data: The authors used a collection of approximately 350 boot sector examples, comprising 250 known boot sector viruses and 100 legitimate boot sectors. For feature selection in the neural network, they also utilized a collection of over 10,000 DOS and OS/2 programs (half a gigabyte of code) to estimate "self" code statistics.
  • New Benchmarks: The paper implicitly introduces a benchmark for boot sector virus detection by evaluating their methods on this specific dataset.

Key Insights and Ideas

  • Biology as a Blueprint: The core insight is that biological immune systems provide a powerful and effective model for computer virus defense, offering mechanisms for rapid adaptation, broad recognition, and automated response that current methods lack.
  • Generic Detection is Key: Moving beyond specific signatures to detect patterns of behavior or code structures (like trigrams) is crucial for handling novel viruses.
  • Automation is Essential: The sheer volume and speed of virus creation necessitate automating the analysis and defense process, a task human experts are increasingly unable to manage.
  • Decoys as a "Trap": Using decoy programs to lure and capture unknown viruses is an ingenious way to obtain samples for analysis, mirroring biological antigen presentation.
  • "Kill Signal" for Network Immunity: Broadcasting threat information across a network can create a collective, adaptive defense, similar to how an immune system response can spread.

Key Results

  • Neural Network Classifier: Achieved a false-negative rate of 10-15% and a false-positive rate of 0.02% on artificial boot sectors. On genuine boot sectors, there were no false positives. The classifier, incorporated into IBM Antivirus, successfully identified several new viruses.
  • Hand-crafted Classifier: Achieved an 82% detection rate with a negligible false-positive rate on its test set.
  • Computer Immune System: Components are described as functioning prototypes, providing useful data for IBM Antivirus and demonstrating the feasibility of automated analysis and signature extraction.

Limitations and Future Work

  • Limitations:
    • The neural network classifier's 15% false-negative rate means some viruses will still escape detection.
  • Obscured code can hinder feature extraction for the neural network.
  • The "generic classifier" is viewed as analogous to the "innate" immune system, which is less sophisticated than the adaptive immune system.
  • The paper acknowledges that algorithmic analysis is less sophisticated than human expert analysis.
  • Future Work:
    • Incorporating the full computer immune system into IBM Antivirus.
  • Deploying these systems in networks with mobile software agents.
  • Further research into more sophisticated biological analogies for defense.
  • Developing more advanced methods for virus removal and handling obscured code.
  • Exploring the "kill signal" mechanism for broader network-wide defense.

Summary updated

BibTeX
@inproceedings{1995biologically,
  title = {{Biologically Inspired Defenses Against Computer Viruses}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1995},
}
Figure