Image-based and Graph-based Adversarial IoT Malware Detection and Classification



Esmaeili, Bardia

Journal Title

Journal ISSN

Volume Title


University of Guelph


With the emergence of the Internet of Things (IoT) various tasks can now be automated to reduce human involvement. The increasing demand for IoT networks has been followed by a rise in malicious programs attacking IoT devices. Hence, researchers have investigated software characteristics and algorithms that may assist in recognizing these harmful entities. Deep Learning (DL) and Machine Learning (ML) models have proven effective in classifying malware. However, these models are susceptible to adversarial attacks, where the attacker can manipulate the input and deceive the model. An adversarial malware bypassing a model could have catastrophic consequences. Therefore, in this study, we attempt to identify adversarial samples against DL-based IoT malware classifiers by utilizing adversarial detection models with image-based and graph-based input representations. We considerably improve the detection performance against adversarial examples on both mediums with 93.1% for the image-based adversarial detector and 98.96% for the graph-based adversarial detector.



Adversarial Detection, IoT Malware Classification, Adversarial Malware


B. Esmaeili, A. Azmoodeh, A. Dehghantanha, H. Karimipour, B. Zolfaghari and M. Hammoudeh, "IIoT Deep Malware Threat Hunting: From Adversarial Example Detection to Adversarial Scenario Detection," in IEEE Transactions on Industrial Informatics, vol. 18, no. 12, pp. 8477-8486, Dec. 2022, doi: 10.1109/TII.2022.3167672.