2026-05-04T19:52:45.165Z http://ezaposleni.singidunum.ac.rs/rest/sciNaucniRezultati/oai

ezaposleni.singidunum.ac.rs/rest/sciNaucniRezultati/oai:2:11621 2025-10-11T22:15:55Z 2

Tackling smart city security: deep learning approach utilizing feature selection and two-level cooperative framework optimized by adapted metaheuristics algorithm 2025 http://ezaposleni.singidunum.ac.rs/rest/sciNaucniRezultati/oai/record/2/11621 https://link.springer.com/article/10.1007/s10207-025-01137-6 K. Kumpf M. Cajic V. Zeljkovic M. Mravik M. Zivkovic J. Mani V. Simic N. Bacanin The development of smart cities is heavily dependent on the Internet of Things (IoT), which facilitates uninterrupted data transmission between different systems and services. IoT networks enable real-time data collection and communication between interconnected devices, optimizing urban infrastructure and services like traffic management, energy usage, and public safety. Nevertheless, due to their lightweight hardware configurations and exposure to online networks, IoT devices are particularly susceptible to cybersecurity breaches, making protection a paramount concern for maintaining a secure operational framework. Conventional defense mechanisms are becoming increasingly inappropriate against the rapidly evolving digital threats, necessitating the adoption of dynamic, AI-driven countermeasures. This study introduces a composite architecture that merges convolutional neural networks (CNNs) with advanced machine learning (ML) classifiers, namely XGBoost and AdaBoost, augmented through metaheuristic optimization techniques to elevate performance outcomes. A layered, two-stage system is introduced to accurately detect and classify intrusions within IoT ecosystems in smart cities. The proposed system incorporates optimization of CNN in the first tier, and a simultaneous optimization of boosting classifiers and feature selection in the second tier of the framework. A real-world IoT dataset serves as the foundation for a comparative evaluation, illustrating the framework’s efficacy in multi-class attack categorization, with the top-performing configurations achieving an accuracy of 97.96% in multi-class attack identification. Furthermore, explainable artificial intelligence (XAI) methods were employed to interpret the reasoning behind model outputs, providing valuable insights for refining data acquisition strategies and enhancing cybersecurity protocols in future deployments. article 10.1007/s10207-025-01137-6 24 221 1615-5270 International Journal of Information Security