A Feature Selection Approach for Network Intrusion Detections

Abstract

With the hype that can be seen regarding the application of Artificial intelligence not to places where needed, but rather in places where it is just possible, it is no wonder right now that the approaches in cyber security will also incorporate the utility of Artificial Intelligence. That is already the case for many network intrusion detection system which use many machine learning and deep learning classifiers to sniff out anomaly from a stream of traffics by analyzing different aspects of each network packets. But as the world becomes more digital, the classifiers will have to toil more and more consuming vast resources as their operating necessities a significant amount of computing power to run the complex classifier models. With the help of feature selection methods, the burden on these classifiers can be lessened to the extent of multiple times. Many other endeavors have been undertaken for this purpose. But very few work by keeping the features intact. In doing otherwise, they add another level of obscurity and complexity to the systems. The method proposed in this work is based on the mutual information which provides a minimum subset of them to choose for training the classifier without changing any feature. It is simpler than any dimension reduction technique, yet provides accuracy significantly closer to the far more complex mechanisms. The proposed method is independent of any classifier models and datasets as well, therefore, it can be extended to be used in any supervised learning process lessening the burden of computation without causing much accuracy loss.