IoT Malware Detection based on Latent Representation

This paper proposes a new approach for IoT malware detection system based on the analysis of IoT network traffic features. First, we use an autoencoder network to gather latent presentation of the input data. This is followed by a classifier to identify whether an IoT network traffic is malware or benign. We carry out a comprehensive comparison of different input feature sets and figure out that using latent representation is more effective than the original features. This proves that autoencoder network can compress the IoT network traffic features and keep only the most meaningful features. The model latent representation and classifies IoT malware and benign with high performance. Another finding is that our trained model can detect new types of abnormal IoT network traffics which do not appear in the training process.