New Blackknife framework enables black-box attacks on graph neural networks

Researchers have developed Blackknife, a novel framework designed to perform black-box adversarial attacks on heterogeneous graph neural networks (HGNNs). This attack method operates under strict limitations, requiring only one-hop neighborhood information and a small number of hard-label queries, without needing access to the victim model's architecture, parameters, gradients, or full graph structure. Blackknife constructs a surrogate model from observable neighborhoods, optimizes perturbations using continuous soft weights, and then discretizes these into relation-preserving structural rewiring operations. Experiments on benchmark datasets like ACM, DBLP, and IMDB show that Blackknife achieves significant attack success rates against various HGNN models, even when faced with topology-based defense strategies. AI

IMPACT Highlights vulnerabilities in graph neural networks, potentially spurring development of new defense mechanisms against adversarial attacks.

RANK_REASON This is a research paper detailing a new attack framework for HGNNs. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• ACM
• Blackknife
• DBLP
• Heterogeneous Graph Neural Networks
• IMDB

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →