Reproduce “Surrogate Representation Learning with Isometric Mapping for Gray-box Graph Adversarial Attacks” on OpenResearch

About this paper

Zihan Liu, Yun Luo, Zelin Zang, Stan Z. Li

Surrogate Representation Learning with Isometric Mapping (SRLIM) enhances gray-box graph adversarial attacks by training surrogate Graph Neural Networks to preserve input graph topological structure in their embeddings, resulting in more transferable gradients that increase target GNN misclassification rates.

Key insights

• Preserving the non-Euclidean topological structure of the input graph in a surrogate model's embedding space is crucial for generating more reliable and transferable adversarial gradients.
• Standard GNNs, primarily optimized for node classification, often implicitly sacrifice topological fidelity in their embedding layers, hindering the transferability of derived adversarial gradients.
• Even marginal improvements in misclassification rates for gray-box attacks, particularly when the victim model's architecture is unknown, signify a meaningful advancement in attack efficacy and transferability.

Summary by alphaXiv.