Recoverable Autoencoder-Based Intrusion Detection with Trust Level Driven Slice Access Control in 5G/B5G networks

Abstract

Network slicing improves service customization but expands the attack surface. This paper proposes Recoverable AutoEncoder-Based Intrusion Detection System with Trust Level Driven Slice Access Control (R-AEIDS-TLDSAC), integrating traffic monitoring, recoverable trustworthiness tracking, and slice access control to block malicious users. AEIDS achieves accuracies of 96.46% and 96.83% on two 5G testbed datasets. In the unknown attack scenarios, it maintains consistently high per-class ROC-AUC values, approximately 0.95 to 1.00, across all attack types without using any labeled attack data. In contrast, supervised learning methods perform well on attack types observed during training, but show unstable performance on unknown attack types. Recovery-enabled TLDSAC reduces 50th/95th/average response times and failure rate by up to 83.95%, 71.15%, 77.19%, and 99.67%, respectively, relative to the no-protection baseline. Compared with TLDSAC without recovery, the recovery mechanism maintains protection performance while lowering false positive rate by 67.62%.