Optimized Cyber-Attack Detection and Classification Techniques Using Deep Learning

Cybersecurity in the Internet of Things (IoT) has become the fastest developing technology, with a significant impact on social life and commercial environments. The number of threats is increasing every day, and attacks are becoming more numerous and complicated. However, existing techniques lead to information loss and slow learning during feature extraction, making the process difficult and increasing computational complexity. Hence, a novel Cascaded Denoising Autoencoder (DAE)-based optimized Back Propagation (BP) algorithm is introduced to enhance feature extraction efficiency, reduce noise during feature overlap, accelerate training speed, and improve accuracy, while minimizing computational complexity. The existing techniques suffer from the multiclass classification problem which leads to lower performance in classification and detection. Hence, a novel Softmax-based Multilayer Perceptron Neural Network with Support Vector Machine (MLPNN-SVM) model is introduced, which combines the neural network with a classifier to enhance the accuracy of classification and detection, reducing the false positive rate and computational cost. The proposed model uses the Network Security Laboratory-KDD (NSL-KDD) dataset for evaluating intrusion detection systems. This dataset contains various cyber-attacks and normal traffic data. As a result, the MLPNN-SVM model detects and classifies cyber-attacks, reducing the false positive rate and computational cost, leading to an achievement of an accuracy of 99.23%, precision of 99.24%, and recall of 99.22%, significantly improving overall system performance.