A Novel Approach for Network Intrusion Detection using Probability Parameter to Ensemble Machine Learning Models

Existing algorithms that have been implemented by researchers to detect attacks, have used bagging, boosting and voting classifier as a method to ensemble multiple models. But the proposed approach uses the probability parameter as a key factor to ensemble two machine learning models and makes an efficient new model to detect attacks. This approach states that, if a particular model is not sure about its decision on a sample, then that decision should not be considered as final decision. The surety of a decision being correct can be calculated using probability parameter. Probability parameter tells with how much probability the decision will be correct. So, the ensemble of two models is done in such a way that, if Model 1 is not sure about the decision then Model 1’s decision will not be considered and then Model 2’s surety will be checked and if found surety is more than Model 1’s then the decision given by Model 2 will be the final decision and vice-versa. Eventually, the performance of final ensemble model will be boosted. After comparing with the existing models, the proposed model gave better performance. In this paper, two final models have been created which are trained on two different datasets, one using NSL-KDD dataset and other using Canadian Institute of Cyber Security’s dataset. The accuracy of the final model for NSL-KDD dataset is 99.32% and that for Canadian Institute of Cyber Security’s dataset is 99.68%. In short, this paper proposes a new hyperparameter related to probability parameter to ensemble two existing Machine Learning models. Future work is to work on zero-day attacks and also detect the type of attack more accurately.