Meta-heuristic optimization and lightweight residual strategy assisted progressive convolutional deep absolute transformer for cyber-attack detection

Internet of Things (IoT) has become more standard across various intelligent applications, and its services has been helps various organizations, high-tech institutions, and possess better reliability. At the same time, the security of this network is a major concern due to increased utilization of these applications. An efficient cyber-attack detection framework is necessary to address the safety concerns. Numerous research studies have been introduced by researchers to ensure the security of IoT environments. However, they offered some limitations, including poor detection accuracy, high computational complexity, information loss and so on. Hence, it is necessary to propose a novel cyber-attack detection framework, so this research implemented an advanced deep learning (DL) technique. Initially, NSL-KDD, UNSW-NB15, CIC IoT 2023 and Ton-IoT datasets are collected to implement the proposed cyber-attack detection. Further, to balance the data samples from publicly available datasets, an Adaptive Synthetic Minority Oversampling Technique (Adp-SMOTE) is developed that increases the number of samples. Further, a hybrid guided mutation strategy based on the rabbit optimization algorithm (Hyb-GuMRO) is proposed to select relevant attack features. After feature selection, a novel lightweight Residual strategy assisted progressive convolutional deep Absolute transformer (ResP-ConDATr) is proposed. This model significantly combines the strengths of lightweight convolution, residual function, progressive attention and transformer model that facilitate effective attack detection by analyzing deep and long-range dependencies. Moreover, chaos-based artificial protozoa optimization (Cha-APO) is proposed to deliberately tune the parameters of the attack detection model and enhance the classification accuracy. Experimental results demonstrated that the proposed cyber-attack detection framework attained better classification accuracies of 99.7%, 99.69%, 99.47% and 99.81% for all four datasets.