Parameter optimization for detecting abnormal data traffic in computer networks

As a result of the widespread use of communication technologies and especially the internet on human life, ensuring the security of information systems and data has become a very important cyber security problem. In order to ensure that systems and users are affected as little as possible in the event of a possible attack, abnormal data traffic detection in computer networks should be reported as quickly and accurately as possible. In this direction, many different approaches and methods have been put forward by researchers. However, most of these studies are not sufficient when considering perception time or classification performance. In this study, the performance evaluation of a selected group of classification algorithms has been implemented in order to detect and prevent actions showing abnormal behavior on network traffic and harmful network traffic that may occur as a result of cyber-attacks. In addition, using data preprocessing steps such as ReliefF Feature Selection, Mahalanobis Distance and Chi-Square Test on high-successful classification algorithms, the effect of using less data features on the performance of classification algorithms is evaluated. As a result of the evaluations, a high classification success was achieved with a small number of features in less than half the time compared to the classification process without feature selection. While the highest classification success of a group of machine learning selected as a result of the feature selection processes belongs to the Random Forest Algorithm with %99.2187, the method with the lowest classification success is the Bayes Network Classifier Algorithm with a rate of %90.5778. NSL-KDD dataset was used for performance tests. It has been tried to achieve maximum performance by using less feature data in high performance and minimum time interval. Keywords: Abnormal Data Detection, Feature Selection, Machine Learning, Data Security, Intrusion Detection Systems