Using a Structured Composition of Correlation Coefficient Functions to Optimize Machine Learning-Driven Cyber Attack Detection Models

The growing number of cybersecurity attacks impacting millions of people highlight the strong need to enhance the use of artificial intelligence techniques in the detection of cybersecurity attacks in real-time. However, designing security attack detection systems that maintains a fine balance between increasing the accuracy of attack detection and minimize false positives is difficult. Feature selection algorithms play a crucial role in designing such effective machine learning systems. An area that has not been explored yet is understanding the effectiveness of using a structured composition of correlation coefficient functions to create a hybrid feature selection algorithm that selects only the most useful features from a dataset. In this paper, we have theoretically proved that it is possible to have such a hybrid feature selection algorithm. On top of that, we showed experimentally that the hybrid feature selection algorithm selected the most useful features from a dataset such that the accuracy of various models trained on CICIDS2017, CICIoV2024, and CIC-UNSW-NB15 turned out to be $99.992 \%, 99.65 \%, 98.28 \%$ respectively with the size of each model being less than 0.27 MB. To ensure the models trained were explainable, we highlighted the top 3 features that impacted the decisions of the model the most. Overall, this research provides an effective way to optimize the performance of Machine Learning models for Security Attack Classification. The code for the paper is available at: https://github.com/space-math/Machine-Learning-Security-Attack-Classification.