(НИЯУ МИФИ, 2025) ANAEDEVHA, R. N.; TROFIMOV, A. G.; Трофимов, Александр Геннадьевич
We propose a unified uncertainty quantification framework for network intrusion detection systems (NIDS) across heterogeneous environments. The system employs (i) hierarchical Gaussian processes (HGP) with adversarially-robust inducing points for cloud datasets and (ii) stochastic Probably Approximately Correct (PAC)-Bayesian transformers for general NIDS datasets. Both pipelines sustain >94% accuracy under adversarial attacks, and provide epistemic uncertainty for risk-based triage.