Security Vulnerabilities and AI‐Driven Intrusion Detection in 5G Network Slicing Architectures

Network slicing enables 5G networks to establish many virtualized, on-demand networks on one shared infrastructure, allowing for a wide range of use cases like the Internet of Things (IoT) and Ultra-Reliable Low-Latency Communications (URLLC). Although such flexibility in architectures is advantageous, it also vastly expands the attack surface, offering new forms of vulnerabilities like cross-slice attacks and shared resource exploitation. This paper examines these security vulnerabilities and suggests an AI-based intrusion detection system tailored to sliced 5G architectures. The scheme utilizes a Transformer-based model incorporating multi-head self-attention mechanisms to effectively recognize complex temporal relationships in network traffic. The model is trained and evaluated on typical 5G datasets, i.e., the 5G NIDD dataset, in varied realistic attack settings. The model performs multi-class classification - it both detects malicious traffic and classifies it into attack types (e.g., DDoS, port scan, protocol exploit). Comparative experiments on baseline models, i.e., Convolutional Neural Networks (CNN), Long Short-Term Memory networks (LSTM), ensemble Autoencoder-Support Vector Machines (AE/SVM), and Gradient Boosting, validate the enhanced performance of the Transformer-based intrusion detection system. Our Transformer model becomes approximately 99% accurate in detection, which is better than the CNN-based method (performed with ~92% accuracy), ensemble techniques (89.33% accuracy), and even traditional machine learning techniques such as Gradient Boosting (99.3% accuracy). These improvements are given in the tables provided, which illustrate the superiority of Transformer models for solving the new security issues of 5G network slicing. The results confirm that sophisticated AI methods-i.e., Transformer models-are a good solution to counter security threats in 5G networks. Future work will improve model interpretability and investigate integration into live operational network environments.