Research in GNN and Foundational Models
Discover the science behind our cutting-edge technology
What is a Graph Neural Network (GNN)?
Graph Neural Networks (GNNs) are a type of artificial neural network specifically designed to work with data structured as graphs. Unlike traditional neural networks that process data in regular formats like images or sequences, GNNs can handle complex relationships between entities.
In the context of cybersecurity, GNNs are especially powerful because they can naturally model the relationships between network devices, traffic patterns, and anomalous behaviors, capturing the inherent structure of communication networks.
Key characteristics of GNNs:
- Processing of complex relational data
- Ability to learn patterns in graphs of any size
- Invariance to node permutations
- Computational efficiency on sparse data
What is a Foundational Model?
A foundational model is an artificial intelligence model trained on large volumes of diverse data that can be adapted for a wide range of specific tasks. These models form the 'foundation' upon which specialized applications are built.
Hypergraph GNN® is our foundational model specialized in cybersecurity, trained with petabytes of real network traffic and billions of documented attack patterns. This massive knowledge base allows it to detect never-before-seen threats through transfer learning.
Advantages of our foundational model:
- Training with real-world data at a massive scale
- Generalization capability for new threats
- Rapid adaptation to specific environments
- Continuous improvement through incremental learning
Our Scientific Publications
"PPT-GNN: A Practical Pre-Trained Spatio-Temporal Graph Neural Network for Network Security"
This work introduces PPT-GNN, a model based on Graph Neural Networks designed for near real-time network intrusion detection. Unlike previous proposals, it allows for faster attack detection and generalizes better across different networks. A self-supervised pre-training method is also introduced, reducing the need for labeled data. The results show significant improvements over existing models on several public datasets.
"Towards a graph-based foundation model for network traffic analysis"
This work proposes a foundational model for network traffic analysis based on dynamic graphs. It uses self-supervised pre-training at the flow level to learn the spatial and temporal dynamics of traffic. It demonstrates improvements in tasks such as intrusion detection, traffic classification, and botnet detection, with a performance 6.87% higher than training from scratch.