Deep Autoencoder-Based Anomaly Detection for Intelligent Network Slice Monitoring in B5G Networks

Anomaly detection in 5G network slicing is critical for ensuring the security and reliability of next-generation telecommunications infrastructure. This study presents a deep autoencoder-based framework for unsupervised anomaly detection in network traffic, leveraging a comprehensive dataset of 210,786 samples with a realistic anomaly rate of 4.5%. The proposed approach incorporates advanced preprocessing techniques, including normalization, interpolation, and oversampling, and prioritizes key network features identified through correlation analysis. The autoencoder model is trained exclusively on normal traffic to learn baseline behaviour, with anomalies detected via reconstruction error analysis. Experimental results demonstrate that the model achieves robust separation between normal and anomalous samples, identifying 1,904 anomalies with a clear margin in reconstruction error statistics. Comparative evaluation with traditional unsupervised methods, such as Isolation Forest, PCA, and One-Class SVM, highlights the superior sensitivity and adaptability of the deep learning approach. The findings underscore the potential of autoencoder-based models for real-time, interpretable anomaly monitoring in highly dynamic and imbalanced 5G network environments, paving the way for more resilient and intelligent network management solutions.