Effective Anomaly Detection in 5G Networks via Transformer-Based Models and Contrastive Learning

Abstract

In this paper, we present a novel approach to anomaly detection in 5G networks using contrastive learning and transformer-based models. Leveraging the power of self-supervision mechanisms, we aim to enhance the detection of anomalous network activities that can compromise the secu-rity and reliability of 5G networks. The methodology involves the use of pretrained transformer models (DistilBERT, BERT, RoBERTa, and ALBERT) as encoders, followed by a projection layer to reduce the dimensionality of the embeddings. We employ a contrastive learning objective to train the models, encouraging the separation of normal and anomalous data points. The trained models are evaluated on a custom 5G testbed dataset, which simulates various normal operations and attack scenarios. Our experimental results demonstrate that DistilBERT, RoBERTa, and ALBERT achieve high accuracy, precision, recall, and Fl-score, significantly outperforming BERT. We provide a comprehensive visualization analysis of each model's performance to illustrate their effectiveness. The findings underscore the effectiveness of contrastive learning combined with transformer-based architectures in achieving robust anomaly detection in 5G networks, offering valuable insights for future research and practical implementations in enhancing network security.