Talk · IEEE International Conference on Big Data Analytics · April 1, 2026
An Interpretable Deep Ensemble Learning Framework for Vision-Based Threat Classification to Enhance IoT Security in Smart Cities
Tamara Kokenovna Zhukabayeva, Zulfiqar Ahmad, Дина Сатыбалдина, Yerik Mardenov, Awad M. Awadelkarim, Nurdaulet Karabayev
Abstract
The rapid proliferation of heterogeneous IoT devices in smart cities has heightened the risk of sophisticated cyber threats targeting critical urban infrastructures. This paper suggests an explainable deep ensemble learning framework for image-based threat classification to protect IoT networks in smart cities. The framework transforms the binary contents of IoT firmware into grayscale images, enabling it to combine four state-of-the-art convolutional neural networks, ResNet50, ResNet101, DenseNet121, and DenseNet169, into an ensemble model that leverages their respective strengths to improve detection performance and robustness. The ensemble model has a high classification accuracy of 95 percent, which is better than the individual base models of 84 to 90 percent. Moreover, it achieves a macro F1-score of 95 percent, indicating a balanced performance in detecting the classes of benign, grayware, hackware, and malware with low false positive and false negative rates. The framework uses Explainable Artificial Intelligence methods, namely Grad-CAM, to explain the decision-making process visually by highlighting important image parts that influence the predictions. This interpretability plays a vital role in assisting cybersecurity analysts in validating alerts and investigating incidents in complex smart city IoT ecosystems. Experimental tests demonstrate that the suggested framework provides scalable, accurate, and interpretable results for reducing attacks at the firmware level, which enhances the resiliency and safety of smart city IoT networks.
More by Yerik Mardenov
- Propagation-Aware Agentic Defense Framework for Resilient Connected Autonomous Vehicle Cyber-Physical SystemsIEEE Vehicular Technology Conference · 2026
- Analysis of vulnerability databases for wireless networks in the Internet of Things for threat modellingDMPedia Lecture Notes in Computer Science & Engineering · 2026
- Trustworthy Anomaly Detection in Industrial IoT-Enabled Cyber-Physical Systems Using Explainable Ensemble LearningProceedings of the AAAI Symposium Series · 2026
- AI-AUGMENTED PENTESTING: SECURING IIOT AND SMART CITIESBulletin of the National Engineering Academy of the Republic of Kazakhstan · 2026
