Evaluation of Diagonal Confidence-Weighted Learning on the KDD Cup 1999 Dataset for Network Intrusion Detection Systems

In this study, I evaluate the performance of diagonal Confidence-Weighted (CW) online linear classification on the KDD Cup 1999 dataset for network intrusion detection systems (NIDS). This is a compatible relationship due to the large number of instances in NIDS datasets, as well as the constantly changing feature distributions. CW learning achieves approximately 92% accuracy on the KDD dataset when optimized, which is higher than both Perceptron and the PassiveAggressive algorithm. CW learning also achieves faster convergence rates than both of these algorithms. Moreover, the accuracy of CW learning on the KDD dataset is comparable to several batch-learning algorithms. This challenges the assumption that batch learning should always be used when feasible. Due to shortcomings of the KDD dataset, a full generalization of CW learning to additional NIDS environments cannot yet be made. Nonetheless, this study shows that there is great promise to applying CW learning to future NIDS research.