2026-08-28T03:18:41.246Z http://ezaposleni.singidunum.ac.rs/rest/sciNaucniRezultati/oai

ezaposleni.singidunum.ac.rs/rest/sciNaucniRezultati/oai:2:12160 2026-08-22T23:03:16Z 2

Tackling edge devices security: Integrated approach using two-level machine learning framework tuned by metaheuristics and generative adversarial networks 2026 http://ezaposleni.singidunum.ac.rs/rest/sciNaucniRezultati/oai/record/2/12160 https://www.sciencedirect.com/science/article/pii/S2666827026001477 N. Bacanin M. Zivkovic M. Djuric Jovicic M. Antonijevic M. Abdel-Salam V. Simic M. Cajic B. Nikolic J. Perisic Intrusion detection is essential in Internet of Things (IoT) and edge-device networks for protecting connected devices and data from unauthorized access and malicious activities. As IoT technologies continue to expand into domains such as the Metaverse, autonomous vehicles, and industrial automation, ensuring the security of these devices has become increasingly important due to their resource-constrained nature and heterogeneous architectures. This study addresses the challenge of intrusion detection in edge networks through a two-level framework comprising convolutional neural networks (CNNs) and a light gradient boosting machine (LightGBM) classifier. A generative adversarial network (GAN) was employed to generate synthetic data and investigate whether such samples can support classifier training when real attack data are limited. The experiments assess the task-oriented utility of the generated samples and do not constitute an independent validation of their intrinsic quality, distributional fidelity, or privacy-preserving properties. To improve the performance of both levels of the framework, a novel adaptation of the particle swarm optimization (PSO) algorithm was proposed for hyperparameter optimization. Comprehensive comparative experiments were conducted, benchmarking the proposed method against several established and recently introduced metaheuristic optimizers under identical experimental conditions. The results demonstrate the effectiveness of the proposed approach and its potential for addressing security challenges in edge-device environments. The best-performing model achieved an accuracy of 99.4278% in a multi-class classification setting. The best-performing models were further analyzed through statistical validation and feature importance assessment using state-of-the-art explainable artificial intelligence techniques. article 10.1016/j.mlwa.2026.100982 100982 2666-8270 Machine Learning with Applications