Skip to content

Self-Supervised Learning of Graph Representations for Network Intrusion Detection

Conference: NeurIPS 2025 arXiv: 2509.16625 Code: N/A Area: Cybersecurity / Graph Learning Keywords: network intrusion detection, graph neural networks, self-supervised learning, masked autoencoder, anomaly detection

TL;DR

This paper proposes GraphIDS, a self-supervised intrusion detection model that unifies graph representation learning and anomaly detection via a masked autoencoder, achieving a PR-AUC of 99.98% and macro F1 of 99.61% on multiple NetFlow benchmarks, surpassing baselines by 5–25 percentage points.

Background & Motivation

Background: Network intrusion detection is highly challenging under limited annotations and continuously evolving attack patterns; graph neural networks (GNNs) have recently been introduced to this domain.

Limitations of Prior Work: Existing GNN-based methods decouple representation learning from anomaly detection, resulting in learned embeddings that are not optimized for identifying attacks.

Key Challenge: Unsupervised/self-supervised approaches are necessary due to the ever-changing nature of attack types and insufficient labeled data, yet the representation learning objective is misaligned with the detection objective.

Key Insight: Unifying graph representation learning and anomaly detection within an end-to-end framework so that embeddings are directly optimized for the downstream detection task.

Method

Overall Architecture

GraphIDS proceeds in three steps: (1) constructing local communication graphs from network traffic; (2) encoding each flow and its topological context with an inductive GNN; and (3) reconstructing embeddings with a Transformer encoder–decoder, where flows with high reconstruction error at inference time are flagged as intrusions.

Key Designs

  1. Local Graph Representation Learning

    • Function: Constructs a local subgraph containing neighboring communications for each network flow.
    • Mechanism: Captures local topological patterns of normal communication.
    • Design Motivation: A global graph is infeasible due to scale; local graphs preserve essential context.

  2. Inductive Graph Neural Network Encoder

    • Function: Embeds each flow together with its local topological context into a vector space.
    • Mechanism: Employs inductive message-passing GNNs to generalize to unseen IP addresses.
    • Design Motivation: Network environments are dynamic and must handle previously unseen nodes.

  3. Transformer Masked Autoencoder

    • Function: Learns global co-occurrence patterns.
    • Mechanism: An encoder–decoder reconstructs masked embeddings; self-attention implicitly models global inter-flow relationships without requiring explicit positional encodings.
    • Design Motivation: Global patterns complement local topology.

  4. Reconstruction Error-Based Anomaly Detection

    • At inference time, normal flows yield low reconstruction error, while attack flows deviate from normal patterns and thus incur high reconstruction error.
    • The detection threshold is determined using normal data from a validation set.

Loss & Training

  • Trained exclusively on normal traffic (no labeled attack samples required).
  • Masking ratio: 30% of embeddings are randomly masked.
  • Loss function: MSE reconstruction loss.

Key Experimental Results

Main Results: Performance on Benchmark Datasets

Dataset Metric Prev. SOTA GraphIDS Gain
NF-CSE-CIC-IDS2018 PR-AUC 91.23% 99.98% +8.75pp
NF-UNSW-NB15 PR-AUC 82.56% 95.42% +12.86pp
NF-ToN-IoT PR-AUC 74.31% 99.52% +25.21pp
NF-CSE-CIC-IDS2018 Macro F1 88.45% 99.61% +11.16pp
NF-UNSW-NB15 Macro F1 79.23% 93.87% +14.64pp
NF-ToN-IoT Macro F1 72.15% 98.94% +26.79pp

Ablation Study

Configuration PR-AUC (CSE-CIC) Macro F1
No local graph (flow features only) 89.34% 86.52%
No Transformer (GNN only) 95.67% 93.21%
No masking (full reconstruction) 97.45% 96.88%
GNN + Transformer (no masked AE) 96.12% 94.53%
GraphIDS (full) 99.98% 99.61%

Key Findings

  • The combination of local graph and global Transformer contributes most significantly to performance.
  • The masking mechanism forces the model to learn richer representations, yielding a 2.5pp improvement over full reconstruction.
  • The model demonstrates strong generalization to unseen attack types.
  • The inductive GNN enables the model to handle dynamic network topologies.

Highlights & Insights

  • End-to-End Unification: Representation learning directly serves the detection objective, eliminating the two-stage disconnect.
  • Self-Supervised Paradigm: Training requires no attack labels, which is well-suited to real-world deployment scenarios.
  • The exceptionally high PR-AUC of 99.98% is notable.

Limitations & Future Work

  • Detection granularity is at the NetFlow level, which may fail to capture application-layer attack details.
  • Threshold selection relies on assumptions about the distribution of normal traffic.
  • Latency and throughput for large-scale real-time deployment are not thoroughly evaluated.
  • Effectiveness on encrypted traffic has not been validated.
  • Mechanisms for adapting to temporal dynamics (concept drift) are absent.
  • MAE (He et al. 2022): masked autoencoder.
  • E-GraphSAGE: GNN for network intrusion detection.
  • DeepSVDD (Ruff et al. 2018): deep anomaly detection.
  • Insight: Self-supervised masked reconstruction shows broad promise in the security domain and is extensible to IoT and industrial control networks.

Rating

  • Novelty: ⭐⭐⭐⭐ — First application of a unified framework with masked AE in NIDS.
  • Experimental Thoroughness: ⭐⭐⭐⭐⭐ — Multiple datasets with comprehensive ablations.
  • Writing Quality: ⭐⭐⭐⭐ — Clear and coherent presentation.
  • Value: ⭐⭐⭐⭐⭐ — High practical deployment value with outstanding performance.