Security-aware IoT botnet attack detection framework using dilated and cascaded deep learning mechanism with conditional adversarial autoencoder-based features

The “Internet of Things (IoT)” technology has been utilized in various industries in the past few years. As the IoT technology has diverse and small devices connected to it, IoT gadgets are vulnerable to several attacks. These networks are heterogeneous and big, making it difficult to handle the security of the overall network. The methods of upgrading these network security methods to combat different security assaults such as keylogging, denial of service, and man-in-the-middle are also difficult. Due to the network’s heterogeneous nature and resource limitations, conventional high-end safety technologies are challenging to implement in IoT networks. Deep learning is an effective technique for identifying botnet attacks. However, the amount of internet activity required for this operation is typically substantial. Thus, implementing a deep learning technique in IoT devices with limited memory is practically difficult. Hence, this paper aims to decrease the dimensionality in the IoT network for efficient attack classification performance in memory constraint IoT devices using Conditional Adversarial Auto Encoder (CAAE) and generate realistic botnet traffic using CAAE to train the model. An efficient N-BaIoT Dataset are initially collected from the online datasets. The collected datasets are then given to the data cleaning process, which helps to avoid unnecessary information by refining essential data for attack detection. Further, the cleaned data is given to the CAAE, which is then incorporated for extracting efficient deep features in order to enhance the attack detection rate. Finally, the attack detection takes place with the extracted deep features, which are performed using the developed Dilated and Cascaded Recurrent Neural Network (DC-RNN) approach for accurately classifying the attacks in IoT devices. Throughout the analysis, the developed model shows 96%, 98%, 97%, and 96% in terms of accuracy, precision, F1-score, and recall. The research assessment is considered to analyze the suggested methods effectiveness by comparing it with conventional techniques.