| Abstract: |
Graph Neural Networks (GNNs) achieve strong performance on a wide range of graph learning tasks but are known to be highly vulnerable to adversarial perturbations of graph structure and node features. While recent work has explored quantum GNNs and attention mechanisms, their robustness under adversarial settings remains largely unexplored. In this work, we present a systematic robustness study of a hybrid classical–quantum Graph Attention Network (QuGAT), which integrates a previously proposed quantum attention module into a scalable message-passing architecture by extending the classical graph attention network (GAT). The quantum component is used exclusively to compute local attention coefficients, while feature aggregation and propagation remain classical, ensuring practical scalability. We evaluate QuGAT under a broad spectrum of adversarial threat models, including poisoning and node injection attacks, across standard node classification benchmarks. Our experiments show that quantum-enhanced attention consistently improves robustness compared to classical GNN baselines, including GAT, without sacrificing predictive performance. To better understand this behavior, we conduct extensive ablation studies over attention depth and number of heads, and provide a spectral analysis linking architectural choices to stability under perturbations. Our results suggest that hybrid quantum attention can act as an implicit regularizer, enhancing robustness in graph learning, and highlight the potential of quantum–classical models for reliable learning under adversarial conditions. |