Abstract

The problems with the current researches on intrusion detection using data mining approach are that they try to minimize the error rate (make the classification decision to minimize the probability of error) by totally ignoring the cost that could be incurred. However, for many problem domains, the requirement is not merely to predict the most probable class label, since different types of errors carry different costs. Instances of such problems include authentication, where the cost of allowing unauthorized access can be much greater than that of wrongly denying access to authorized individuals, and intrusion detection, where raising false alarms has a substantially lower cost than allowing an undetected intrusion. In such cases, it is preferable to make the classification decision that has minimum cost, rather than that with the lowest error rate.For this reason, we examine how cost-sensitive machine learning methods can be used in Intrusion Detection systems. The performance of the approach is evaluated under different experimental conditions and different models in comparison with the KDD Cup 99 winner resultsin terms of average misclassification cost, as well as detection accuracy and false positive ratesthough the winner used original KDD dataset whereas for this research NSL-KDD dataset which is new version of the original KDD cup data and it is better than the original dataset in that it has no redundant data is used. For comparison of results of CS-MC4, CS-CRT and KDD winner result, it was found that CS-MC4 is superior to CS-CRT in terms of accuracy, false positives rate and average misclassification costs. CS-CRT is superior to KDD winner result in accuracy and average misclassification costs but in false positives rate KDD winner result is better than both CS-MC4 and CS-CRT classifiers.