Interpretable IoT Network Monitoring via Graph Neural Embeddings
A new research paper introduces an interpretable pipeline for monitoring Internet of Things (IoT) traffic flows using graph neural networks (GNNs). The work addresses the challenge of opaque GNN representations by generating low-dimensional, visualizable embeddings mapped onto a latent manifold. This projection allows for interpretable monitoring of evolving network states, while feature attribution techniques decode the characteristics shaping the manifold structure. The approach aims to improve security-critical operations in complex IoT ecosystems.
Key facts
- The paper is published on arXiv with ID 2602.05817.
- It addresses the complexity and heterogeneity of IoT network topologies.
- Traditional monitoring tools fail to capture evolving relationships between devices.
- GNNs are used to learn from relational data but their internal representations are opaque.
- The proposed pipeline generates directly visualizable low-dimensional representations.
- High-dimensional embeddings are mapped onto a latent manifold.
- Feature attribution techniques decode the characteristics shaping the manifold structure.
- The work enables interpretable monitoring and interoperability of network states.
Entities
Institutions
- arXiv