A dynamic feature selection technique to detect DDoS attack

Abstract

Distributed Denial of Service (DDoS) attacks are a weapon of choice for hackers that are accountable for degradation in network performance, website downtime, and sometimes crashing of servers, etc. Because the feature selection mechanism is still naive, detecting DDoS attacks is difficult. There is a need to efficiently select features from the attack dataset, which also needs to be further analyzed. Simply placing the features as input to the classification model will result in an irrelevant class with poor performance. In the paper, in order to select the effective features, we proposed a voting-based hybrid feature selection technique that was cross-analyzed by three correlation methods. The hybrid method not only reduces the dimensions of the features and removes redundancy among them, but also provides the best relevant features for classification. The deployment of a multilayer perceptron with genetic algorithm (MLP-GA) as a classifier outperforms the conventional classifiers by furnishing an accuracy of 98.8%, a false positive rate of 0.6%, and the capability of early detection. The proposed method has been comprehensively validated using various publicly available benchmark datasets.