Slice-Aware Adversarial Vulnerability Analysis of TabNet-Based Intrusion Detection in 5G Network Slicing

Source abstract

A TabNet-based network intrusion detection system can achieve 97% global accuracy while simultaneously exhibiting near-total adversarial collapse in the URLLC slice (PSRI = 0.023)—a critical vulnerability that conventional metrics entirely conceal. This paper introduces the Per-Slice Robustness Index (PSRI) and Slice-Weighted F1 (SW-F1) to expose such hidden vulnerabilities in 5G network slicing environments, and proposes Attention-Guided Evasion (AGE), a transfer-based adversarial attack that exploits TabNet’s native attention mask to concentrate perturbations on the most influential features. Using the CICIoT2023 dataset with 34 attack categories mapped to eMBB, URLLC, and mMTC service slices, we demonstrate that AGE achieves a 50.2% accuracy drop at ε=0.10 while using 17× less L2 perturbation than FGSM and modifying only 8.5% of features. The proposed metrics expose a 0.121-point gap between global weighted F1 (0.928) and slice-balanced SW-F1 (0.807), demonstrating that conventional evaluation systematically overestimates protection for minority slices in heterogeneous 5G environments. These findings highlight the necessity of slice-aware adversarial evaluation frameworks for 5G network security.