An enhanced attention and dilated convolution-based ensemble model for network intrusion detection system against adversarial evasion attacks

Abstract Network Intrusion Detection System (NIDS) is a system for recognizing suspicious activities in the network traffic. Numerous machines learning and deep learning-aided IDSs have been implemented in the past, however, most of these techniques face challenges based on class imbalance issues and high false positive rates. Other primary problems of the conventional techniques are their vulnerability to adversarial attacks and also there is no analysis done on how NIDS sustain their performance over various attacks. Moreover, recent studies have demonstrated that while handling the attackers in real-time, the deep learning-based IDS shows slight variations in accuracy. To defend against adversarial evasion attacks, an enhanced deep learning-based NIDS model is designed in this work. For this purpose, at first, the required data is collected from available websites. From the collected data, effective features are extracted to improve the accuracy of the process. To select the optimal features, this work employed the Improved Cheetah Optimizer (ICO) that eliminates the unwanted features efficiently. Further, an Attention and Dilated Convolution based Ensemble Network (ADCEN) is implemented to detect the intrusions from the optimal features. The Deep Temporal Convolutional Neural Network (DTCN), Long Short Term Memory (LSTM), and Gated Recurrent Unit (GRU) models are integrated to develop the ADCEN. The outcomes from each technique are considered for the fuzzy ranking mechanism to generate the final detected outcome. Thus, recognized intrusion is attained as the outcome and to demonstrate how well the recommended deep learning-based NIDS defends against adversarial evasion assaults, experiments are conducted against conventional models. The accuracy and the FPR values of the recommended model are 95 and 4.9 when considering the first dataset which is superior to the conventional techniques. Thus, the findings indicated that the implemented NIDS against adversarial evasion attacks attained more effective solutions than the baseline approaches.