基于局部影响分析模型的图神经网络对抗攻击

图神经网络(GNN)容易受到对抗攻击安全威胁。现有研究未注意到图神经网络对抗攻击与统计学经典分支统计诊断之间的联系。该文分析了二者理论本质的一致性，将统计诊断的重要成果局部影响分析模型引入图神经网络对抗攻击。首先建立局部影响分析模型，提出并证明针对图神经网络攻击的扰动筛选公式，得出该式的物理意义为扰动对模型训练参数影响的度量。其次为降低计算复杂度，根据扰动筛选公式的物理意义得出扰动筛选近似公式。最后引入投影梯度下降算法实施扰动筛选。实验结果表明，将局部影响分析模型引入图神经网络对抗攻击领域具有合理性；与现有攻击方法相比，所提方法具有有效性。

Adversarial Attacks on Graph Neural Network Based on Local Influence Analysis Model

Graph Neural Networks (GNNs) are vulnerable to adversarial attacks. Existing papers do not pay attention to the relationship between adversarial attacks and statistical diagnosis, a classical branch of statistics. In this paper, the consistency of the two theories is analyzed, and the local influence analysis model, an important achievement of statistical diagnosis, is introduced into adversarial attack on GNNs. Firstly, the local influence analysis model is established to derive the equation of perturbation selecting of attacks, and the physical meaning of this equation is a measurement of the influence of perturbation on model training parameters. Secondly, to reduce the computational complexity, according to the physical meaning of the perturbation selecting equation, the approximate equation is obtained. Finally, the projected gradient descent algorithm is introduced to implement disturbance selecting. Experimental results show that it is reasonable to introduce the local influence analysis model into the field of adversarial attacks on graph neural network; Compared with the existing attack methods, the proposed method is more effective.