Answers are grounded in this paper and cite their sources.
The rapid growth of encrypted network traffic challenges intrusion detection systems that rely on payload inspection. We propose a topology-aware unsupervised intrusion detection framework that models network flows as interaction graphs and applies a graph neural autoencoder to identify structural anomalies from flow metadata alone, without inspecting packet payloads. The autoencoder is trained with a topology-aware augmentation strategy (Gaussian feature noise, edge feature masking, node feature masking) together with edge-level contrastive learning. We evaluate the framework on six datasets organized into eight conditions, spanning enterprise, cloud, industrial IoT and testbed networks as well as end-to-end encrypted DNS-over-HTTPS (DoH) tunneling and command-and-control (C2) traffic over TLS and QUIC. Against six unsupervised baselines, including GraphIDS, Anomal-E, KitNET, and a variational autoencoder, the framework attains the highest mean ROC-AUC on both the general benchmarks (0.986) and the encrypted conditions (0.936). A benign-vs-benign origin control reveals a dataset-origin confound in cross-corpus DoH evaluation (origin-AUC 0.942), so the encrypted-DoH claim rests on a same-origin setting, where the proposed model reaches ROC-AUC 0.797. Without recalibration on the target, the model loses 3.0% (relative) of its mean ROC-AUC under cross-environment transfer and attains a mean cross-domain ROC-AUC of 0.911, compared with 0.415–0.609 for the baselines, and it processes a two-minute window in under 18 ms on a CPU, the highest throughput among the timed models.
No takes yet. Share an insight, caveat, or question.
Lee et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: