Published Paper Details:

For the easy and flexible management of large scale networks, Software-Defined Networking (SDN) is a strong candidate technology that offers centralisation and programmable interfaces for making complex decisions in a dynamic and seamless manner. On the one hand, there are opportunities for individuals and businesses to build and improve services and applications based on their requirements in the SDN. On the other hand, SDN poses a new array of privacy and security threats, such as Distributed Denial of Service (DDoS) attacks. For detecting and mitigating potential threats, Machine Learning (ML) is an effective approach that has a quick response to anomalies. In this article, we analyse and compare the performance, using different ML techniques, to detect DDoS attacks in SDN, where both experimental datasets and self-generated traffic data are evaluated. Moreover, we propose a simple supervised learning (SL) model to detect flooding DDoS attacks against the SDN controller via the fluctuation of flows. We verify the outcome through simulations and measurements over a real testbed. Our main goal is to find a lightweight SL model to detect DDoS attacks with data and features that can be easily obtained. In this study, DDoS attacks in SDN were detected using machine learning-based models. First, specific features were obtained from SDN for the dataset in normal conditions and under DDoS attack traffic. Then, a new dataset was created using feature selection methods on the existing dataset. Feature selection methods were preferred to simplify the models, facilitate their interpretation, and provide a shorter training time. Both datasets, created with and without feature selection methods, were trained and tested with K-Nearest Neighbors (KNN) classification models. The test results showed that the use of the wrapper feature selection with a KNN classifier achieved the highest accuracy rate Our results show that SL is able to detect DDoS attacks with a single feature. The performance of the analysed SL algorithms is influenced by the size of training set and parameters used. The accuracy of prediction using the same SL model could be entirely different depending on the training set. This project is developed using PYTHON, PyCharm as IDE, Kaggle dataset.