An edge sensitivity based gradient attack on graph isomorphic networks for graph classification problems.

Area of Science:

• Machine Learning
• Graph Neural Networks
• Adversarial Attacks

Background:

• Graph Neural Networks (GNNs) are powerful for modeling relationships but lack ideal Euclidean space representations.
• GNNs are vulnerable to adversarial attacks, yet gradient-based attacks on latent space embeddings are underexplored.

Purpose of the Study:

• To propose a novel white box gradient-based adversarial attack for GNNs using contrastive latent space representations.
• To develop a robust GNN model that learns spectral and spatial graph properties, considering isomorphic properties.

Main Methods:

• Developed a white box gradient-based attack targeting GNN latent space embeddings.
• Constructed a strong base GNN model incorporating spectral, spatial, and isomorphic properties.
• Validated the GNN model on four molecular property prediction datasets.

Main Results:

• The proposed GNN model outperformed over 75% of evaluated LLM-based architectures.
• The adversarial attack reduced the GNN model's performance by an average of 25% on graph classification tasks.

Conclusions:

• The study introduces a new adversarial attack strategy for GNNs, addressing gaps in existing literature.
• The findings demonstrate the susceptibility of robust GNNs to latent space attacks and inform the development of more resilient models.