Evolutionary neural networks for anomaly detection based on the behavior of a program

Area of Science:

• Computer Science
• Cybersecurity
• Machine Learning

Background:

• Intrusion detection systems (IDS) commonly use machine learning on system-call audit data.
• Neural networks excel at learning system-call sequences but determining optimal structures and weights is time-consuming.
• Existing analytical solutions for neural network optimization in intrusion detection are limited.

Purpose of the Study:

• To propose a novel intrusion-detection technique using evolutionary neural networks (ENNs).
• To address the challenge of efficiently determining neural network structures and weights for intrusion detection.

Main Methods:

• Utilizing evolutionary algorithms to optimize neural network architectures and parameters simultaneously.
• Applying the proposed ENN technique to system-call audit data for behavior learning.
• Evaluating the method using the 1999 Defense Advanced Research Projects Agency (DARPA) Intrusion Detection Evaluation (IDEVAL) dataset.

Main Results:

• ENNs significantly reduce the time required to develop effective neural networks for intrusion detection.
• The proposed ENN technique demonstrates promising performance on benchmark intrusion detection data.
• Simultaneous discovery of structures and weights leads to superior neural network models.

Conclusions:

• Evolutionary neural networks present a more efficient and effective approach to intrusion detection.
• ENNs offer a viable solution for real-world application of machine learning in cybersecurity.
• The method shows potential for improving the accuracy and speed of detecting network intrusions.