Detection of distributed denial of service ((DDoS) attacks using artificial intelligence methods

The Internet penetrates more and more areas of our lives every day. However, security problems have started to be encountered more and more. According to the report published by Nexuesquard for the first quarter of 2023, Distributed Denial of Service (DDoS) attacks increased by 183% (Anonymous, 2023a). At the same time, as the number of devices joining the Internet increases with Internet of Things applications, the capacity of attackers' attack networks is expanding. In order to ensure security in the Internet environment, artificial intelligence-based intrusion detection systems are being built in accordance with the techniques of the period. The aim of this study is to contribute to the development of advanced intrusion detection systems by selecting the features used in intrusion detection systems with the mutual information method and analyzing the success rates of artificial intelligence-based algorithms in detecting DDoS attacks. CICDDOS2019 dataset was used in the study. DDoS attacks in the dataset were detected using machine learning algorithms such as K-Nearest Neighbor, Decision Tree, Naive Bayes and Random Forest algorithms. In all experiments conducted within the scope of the study, an acceptable decrease was observed in the classification success after feature selection, and a significant decrease was observed in detection times. In real-life network traffic management, it is expected that attacks are blocked regardless of their type and non-attack packets are included in the network traffic. In the binary classification process performed using all features, the accuracy rate in detecting non-attack packets in the test set was obtained in the K-Nearest Neighbor algorithm with 99.34%, and after the feature selection was made, the highest accuracy rate was obtained in the Naive Bayes algorithm with 96.29%. Reducing the number of features used in classification is expected to reduce memory usage and CPU processing time. In the experimental study, in parallel with this, a significant reduction in the classification processing time after feature selection was observed.