Semantic Enhanced Hypergraph Attack Method

Hypergraph Neural Networks (HGNNs) have demonstrated exceptional capability in modeling high-order correlations; however, their vulnerability to adversarial attacks remains inadequately addressed due to the limited scope of existing security investigations. The prevailing white-box structural attack, HyperAttack, relies exclusively on gradient-derived information and overlooks the semantic affinities between nodes and hyperedges. This oversight limits attack efficacy because gradient signals can be noisy or ambiguous under certain conditions (e.g., saturated regions or local optima), whereas semantic similarities provide complementary cues that help identify hyperedges whose perturbation more reliably alters the target node’s representation. To mitigate this limitation, this paper introduces a semantic enhanced adversarial attack framework for hypergraph neural networks, termed SE-HyperAttack. Specifically, hyperedge features are first aggregated, and semantic similarity scores are computed based on the feature similarity between target nodes and their incident hyperedges to capture latent semantic correlations. These semantic similarity scores are subsequently integrated with integrated gradient scores via a weighted summation scheme, refining the precision of hyperedge selection. Extensive experiments on two datasets demonstrate that the proposed SE-HyperAttack achieves an optimal average attack success rate (ASR) of 79.4%, showing an improvement of 2.6% over HyperAttack. Ablation studies further ascertain that a semantic weight of 30% yields peak performance, beyond which degradation is observed. Notably, the proposed approach preserves computational efficiency commensurate with HyperAttack, incurring negligible additional overhead. These findings substantiate that the integration of semantic information effectively enhances adversarial attack effectiveness on hypergraph neural networks without compromising efficiency.