Cyberattack Detection: Performance Evaluation of Long Short-Term Memory Network and Convolutional Neural Network

The rapid growth, complexity, and widespread application of the Internet of Things has undoubtedly created room for the introduction of new efficient Intrusion Detection Systems (IDS) that the traditional methods of machine-learning algorithms might not be sufficient to cater to. Cyberattacks have become a serious online threat challenging the world which needs to be addressed to avoid internet extortion and blackmail and using some traditional methods of machine learning might not be sufficient to detect intrusion in a system. The research aims to evaluate the performance of Long Short-Term Memory Network (LSTM) and Convolutional Neural Network (CNN) CNN and LSTM Deep Learning Algorithms were used, and the models were trained and evaluated using the NSL-KDD dataset and the Malware dataset (N Saravana) dataset. The Convolutional Neural Network using the nsl-kdd dataset had a classification accuracy of 99.27%, a false-positive rate of 0.0085, and a true positive rate (Recall) of 99.14%, and using the Malware dataset (N Saravana) had an accuracy of 97.15%, false-positive rate of 0.011, and a true positive rate (Recall) of 95.47%. Long Short-Term Memory Network (LSTM) using the nsl-kdd dataset had a classification accuracy of 99.30%, false-positive rate of 0.0075, and a true positive rate (Recall) of 99.34%, and using the Malware dataset (N Saravana) had an accuracy of 98.37%, false positive rate of 0.012, and a true positive rate (Recall) of 97.89%. Comparing the evaluated models, the LSTM model generally outperformed the CNN model in most metrics, including accuracy, precision, recall, F1-score, error rate, false positive rate, false negative rate, ROC-AUC, PR-AUC, Kappa, and MCC. However, both models demonstrated high performance in classifying cyberattacks on their respective datasets.