AutoGraphAD: Unsupervised Network Anomaly Detection Using Variational Graph Autoencoders
AutoGraphAD is a novel unsupervised anomaly detection method for network intrusion detection, proposed in arXiv:2511.17113. It uses a Heterogeneous Variational Graph Autoencoder operating on graphs built from connection and IP nodes. The model is trained via unsupervised and contrastive learning without labeled data, combining loss weights into an anomaly score. This approach addresses the high cost of labeled datasets and issues with outdated or mislabeled public data.
Key facts
- AutoGraphAD uses a Heterogeneous Variational Graph Autoencoder.
- It operates on heterogeneous graphs from connection and IP nodes.
- Training uses unsupervised and contrastive learning without labeled data.
- Losses are weighted and combined into an anomaly score.
- Aims to reduce reliance on costly labeled datasets.
- Addresses limitations of existing public datasets with outdated or mislabeled attacks.
- Published on arXiv with ID 2511.17113.
- The method is for network intrusion detection systems (NIDS).
Entities
Institutions
- arXiv