融合子图结构的知识图谱嵌入对抗性攻击方法

Adversarial Attack Against Knowledge Graph Embedding Based on Subgraph Structure

知识图谱嵌入(Knowledge Graph Embedding,KGE)技术的高速发展极大提高了人类对于结构化知识的利用效率,该技术也为人工智能的相关应用提供了有利的支撑.但是知识图谱嵌入方法的脆弱性(vulnerability)给知识图谱的应用带来了巨大的挑战,近期的一些研究表明,在训练数据中添加微小的扰动便能对训练后的机器学习模型造成巨大的影响,甚至导致错误的预测结果.目前针对可能破坏知识图谱嵌入模型的安全漏洞的研究大多关注嵌入模型的损失函数而忽略图结构信息的作用,因此本文提出了一种融合子图结构深度学习的攻击方法DLOSSAA(Deep Learning of Subgraph Structure Adversarial Attack),对知识图谱嵌入的健壮性进行研究.DLOSSAA方法首先通过对子图结构的深度学习捕获相关子图的结构信息,然后通过修正的余弦相似度(Adjusted Cosine Similarity)筛选出最佳的攻击样本,最后将攻击样本添加到训练数据中进行攻击.实验结果表明,该方法能够有效降低攻击后的知识图谱嵌入模型的性能,攻击效果优于大部分已有的对抗性攻击方法.

The rapid development of knowledge graph embedding(KGE)technology has greatly improved the utilization efficiency of structured human knowledge,which provides favorable support for the development of artificial intelligence.However,the vulnerability of knowledge graph embedding brings great challenges to the application of knowledge graph.The key problem is that machine learning methods do not perform well under malicious attacks.Recently,many studies show that adding small perturbation into the training data can have a great impact on the well-trained model,and even lead to wrong prediction results.Therefore,this paper proposes a deep learning adversarial attack method DLOSSAA to study the robustness of knowledge graph embedding.The subgraph structure deep learning can capture the structural information of relevant subgraph of knowledge graph,and then adversarial perturbations with best attack effect are selected through the adjusted cosine similarity.Finally,the adversarial perturbations are injected into the training data.Experimental results show that this method can effectively reduce the performance of knowledge graph embedding models,and the attack effect is better than most of existing adversarial attack methods.