一种防御投影梯度下降攻击的图时空注意力网络方法

A Graph Spatial-Temporal Attention Network Method for Defending Against Projected Gradient Descent Attacks

摘要: 图时空网络在交通预测与疫情分析等领域广泛应用，然而其时空耦合结构也使其易受对抗攻击干扰。针对投影梯度下降PGD对抗攻击对模型性能的破坏，文章提出了一种鲁棒图时空自注意力网络RGSTAN，该方法引入交叉时空自注意力强化时空特征建模能力，并结合时空平滑策略在空间与时间维度上对动态节点特征表示进行平滑处理与对抗扰动，降低对抗噪声对关键特征的影响，从而提升模型鲁棒性。此外，还分析了图时空模型面临的安全威胁，揭示了投影梯度下降对抗攻击在时空信息传播过程中的干扰原理。最后，基于疫情与流量动态图数据的实验结果表明，在不同攻击约束的对抗环境下，时空平滑机制能强化时空自注意力网络层的鲁棒性，使RGSTAN能有效缓解投影梯度攻击的干扰。

Abstract: Graph spatial-temporal networks are widely used for traffic forecasting and epidemic analysis, but their spatial-temporal correlated structure makes them susceptible to perturbation from adversarial attacks. In order to limit the performance degradation caused by projected gradient descent attacks, this paper proposes a Robust Graph Spatial-Temporal Self-Attention Network model. The model integrates cross-spatial-temporal self-attentions to enhance the spatial-temporal representation dependencies, and utilizes a spatial-temporal smoothing strategy that smooths and perturbs dynamic node features from spatial and temporal dimensions, thus mitigating the impact of adversarial noise on crucial representations and improving the model’s robustness. Additionally, this paper explores the security threats of graph spatial-temporal models and analyzes the PGD attack mechanism, which interferes with spatial-temporal information flow. Finally, experimental results based on epidemic and traffic dynamic graph data show that the spatiotemporal smoothing mechanism can enhance the robustness of the spatiotemporal self-attention network layer under adversarial environments with different attack constraints, enabling RGSTAN to effectively mitigate the interference of projection gradient attacks.