Malicious user classification in cognitive 5G networks using novel improved bidirectional encoder representations from transformers model

In cognitive 5G networks, identifying malicious users is essential for protecting dynamic spectrum access against attacks like jamming as well as spectrum sensing fraud. However, the complexity associated with many 5G settings, limited labelled information, as well as evolving attack methods make it extremely challenging to detect these individuals. In order to provide dependable effectiveness as well as confidence in cognitive radio-enabled 5G communication frameworks, these networks need real-time, efficient, and adaptable classification approaches that can reduce false alarms while generalizing successfully. Therefore, this paper performs the Malicious User Classification in Cognitive 5G Networks (MUC-C5GN) using novel intelligent machine learning-oriented optimization methodology. The data is first collected from the standard benchmark sources called 5G Network Intrusion Detection Dataset (5G‑NIDD). The pre-processing of this collected data is accomplished by the normalization and scaling methods. Next, the feature extraction of this pre-processed data takes place by the Self-Attention RNN-AE (Recurrent Neural Network-Autoencoder) approach. Finally, the classification of the malicious users in cognitive 5G networks is performed by the novel Improved Bidirectional Encoder Representations from Transformers (IBERT) model. The parameter tweaking in BERT is done by the nature inspired optimization algorithm called Revolution Optimization Algorithm (ROA). Accuracy maximization is considered as the fitness function for the overall MUC-C5GN model. Over seven types of attack as well as benign traffic, the proposed IBERT-ROA method is evaluated against LSTM-GRU, MLP, Chaotic DBN, and Detectron2 + YOLOv7. According to simulation results, IBERT-ROA achieves the best results with 99.74% accuracy, 98.48% sensitivity, 98.91% precision, 97.82% MCC, as well as 98.91% specificity-demonstrating improvements of up to 5.99% in sensitivity and 2.74% in accuracy over the state-of-the-art technique. These results demonstrate the effectiveness, scalability, as well as suitability of IBERT-ROA for real-time malicious user detection in dynamic cognitive 5G environments.