Cyber Attack Intensity Prediction Using Feature Selection and Machine Learning Models

Cybercrimes are becoming increasingly more sophisticated and dangerous as we rely more on technology in all aspects of our lives. Crimes, such as data breaches, cyber extortion, and identity theft are more common than ever. It is estimated to cost the world billions of dollars and no country is immune to it. This paper aims to investigate the possibility of using various machine learning techniques, such as stochastic gradient descent and random forest in order to forecast potential cyberattacks. This is done by training the chosen machine learning model using the UNSW-NB15 dataset. This dataset contains nine types of network-based cyberattacks along with normal network activities. Information Gain Attribute Evaluation (IGAE) is used for feature selection with a rank cutoff 0.15. For the cross-validation task, 10-fold cross-validation is used. Results show that applying feature selection marginally increased the accuracy of all models used. The accuracy of the models ranged between 92.4% and 99.9%. The highest accuracy is obtained when using the random forest algorithm and a combination of random forest and logistic regression.