Research map / Graph learning / Graph self-supervised learning

Self-Supervised Learning of Graph Representations for Network Intrusion Detection

NeurIPS 2025 · Graph learning › Graph self-supervised learning · approach: Semi-supervised node representation learning · 4 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes GraphIDS, a self-supervised network intrusion detector that jointly trains an E-GraphSAGE flow encoder and a Transformer masked autoencoder on benign flows, flagging flows by reconstruction error.

Problem

Network intrusion detection under limited labels and evolving attacks, where graph representation learning is usually decoupled from anomaly detection and pretext tasks often need negative samples.

Approach

Jointly train a GNN encoder (E-GraphSAGE, 1-hop) and a Transformer-based masked autoencoder end-to-end on benign flow embeddings, so the GNN embeddings are optimized directly for reconstruction-based anomaly scoring; attention masking is used as structured regularization.

Key results

• On NF-UNSW-NB15-v3, GraphIDS reaches PR-AUC 0.9998 and macro F1 0.9961 across seeds.
• On NF-CSE-CIC-IDS2018-v3, GraphIDS reaches PR-AUC 0.8819 and macro F1 0.9447, versus 0.2555 PR-AUC for the best Anomal-E variant.
• On NF-UNSW-NB15-v2, GraphIDS achieves PR-AUC 0.8116, above Anomal-E's 0.7489, while Anomal-E retains higher macro F1 on CSE-CIC-IDS2018-v2 (0.9410 vs 0.9431 for GraphIDS is comparable).

Measured effects reported in the paper

Surprising findings

• A SimpleAE ablation (MLP instead of Transformer) matches or beats GraphIDS on several PR-AUC numbers, suggesting the autoencoding objective drives much of the gain.
• Positional encodings had little effect, and the T-MAE ablation without graph context performs comparably to GraphIDS on NF-UNSW-NB15-v3.

Limitations

• Performance may degrade under abrupt distribution shifts and needs online adaptation.
• Limited topological context in single-host monitoring scenarios.
• Infiltration attacks resembling benign traffic are often missed on CSE-CIC-IDS2018; NF-UNSW-NB15-v2 has more false positives.

Experimental setup

Datasets: NF-UNSW-NB15-v2 NF-UNSW-NB15-v3 NF-CSE-CIC-IDS2018-v2 NF-CSE-CIC-IDS2018-v3

Benchmarks: NF-UNSW-NB15-v2 NF-UNSW-NB15-v3 NF-CSE-CIC-IDS2018-v2 NF-CSE-CIC-IDS2018-v3

Metrics: PR-AUC macro F1

Baselines: Anomal-E CBLOF SAFE T-MAE SimpleAE HBOS Isolation Forest PCA

Builds on: GraphSAGE E-GraphSAGE BERT GraphMAE Anomal-E

Closest papers by meaning

• Self-Discriminative Modeling for Anomalous Graph Detection (ICML 2025 · similarity 0.85)
• UniFORM: Towards Unified Framework for Anomaly Detection on Graphs (AAAI 2025 · similarity 0.84)
• Stealthy Yet Effective: Distribution-Preserving Backdoor Attacks on Graph Classification (NeurIPS 2025 · similarity 0.84)
• Node Identifiers: Compact, Discrete Representations for Efficient Graph Learning (ICLR 2025 · similarity 0.84)
• Clean-Label Graph Backdoor Attack in the Node Classification Task (AAAI 2025 · similarity 0.84)
• DiffGAD: A Diffusion-based Unsupervised Graph Anomaly Detector (ICLR 2025 · similarity 0.83)