Advanced Detection of IoT Malware Using Deep Learning Techniques on TON_IoT Network Traffic

Abstract

IOT malware detection can be challenging especially when trying to develop a solution that can also perform well in the low computing power of these devices. Deep learning is powerful in features extraction but unfortunately, it also requires high computing power. Gradient boosting and other traditional machine learning algorithms are fast at training and inference but can lack the precision of deep learning. Findings from the literature have shown that complex neural networks integrated with attention mechanisms and residual connections can reach the accuracy of ensemble models. This paper evaluates the accuracy of several deep neural network architectures on the TON_IoT dataset, 211,043 network traffic samples with 10 different attack labels. The performance gained by the neural networks was significant and the best accuracy was 94.816%. It was achieved by the MLP-4Layers+Attention with an F1 score of 0.9479 and AUC of 0.9972. This will potentially contribute in the design of neural networks that can attain ensemble level accuracy in low computation IoT environment