HyperNear: Unnoticeable Node Injection Attacks on Hypergraph Neural Networks
Published in ICML, 2025
In this work, we explore adversarial attacks on hypergraph-based models, revealing their structural vulner abilities. We introduce HyperNear, the first node injection attack framework for HNNs, leveraging homophily constraints for stealth. Experiments demonstrate its high attack efficacy, transferability, and unnoticeability. Our analysis of unnoticeability metrics deepens the understanding of hypergraph structure and model robustness, paving the way for future work on strengthening hypergraph-based models.
Recommended citation: Cai T, Jiang Y, Li M, et al. HyperNear: Unnoticeable Node Injection Attacks on Hypergraph Neural Networks[C]//Forty-second International Conference on Machine Learning.
Download Paper