Adversarial Attacks on Knowledge Graph Embeddings via Instance Attribution Methods

Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 8225–8239 November 7–11, 2021. c⃝2021 Association for Computational Linguistics 8225 Adversarial Attacks on Knowledge Graph Embeddings via Instance Attribution Methods Peru Bhardwaj1 John Kelleher2∗ Luca Costabello3∗ Declan O’Sullivan1∗ 1 ADAPT Centre, Trinity College Dublin, Ireland 2 ADAPT Centre, TU Dublin, Ireland 3 Accenture Labs, Ireland peru.bhardwaj@adaptcentre.ie Abstract Despite the widespread use of Knowledge Graph Embeddings (KGE), little is known about the security vulnerabilities that might disrupt their intended behaviour. We study data poisoning attacks against KGE models for link prediction. These attacks craft adversar- ial additions or deletions at training time to cause model failure at test time. To select adversarial deletions, we propose to use the model-agnostic instance attribution methods from Interpretable Machine Learning, which identify the training instances that are most influential to a neural model’s predictions on test instances. We use these influential triples as adversarial deletions. We further propose a heuristic method to replace one of the two entities in each influential triple to generate ad- versarial additions. Our experiments show that the proposed strategies outperform the state-of- art data poisoning attacks on KGE models and improve the MRR degradation due to the at- tacks by up to 62% over the baselines. 1 Introduction Knowledge Graph Embeddings (KGE) are the state- of-art models for relational learning on large scale Knowledge Graphs (KG). They drive enterprise products ranging from search engines to social net- works to e-commerce (Noy et al., 2019). However, the analysis of their security vulnerabilities has re- ceived little attention. Identifying these vulnerabil- ities is especially important for high-stake domains like healthcare and finance that employ KGE mod- els to make critical decisions (Hogan et al., 2020; Bendtsen and Petrovski, 2019). We study the se- curity vulnerabilities of KGE models through data poisoning attacks (Biggio and Roli, 2018; Joseph et al., 2019) that aim to degrade the predictive per- formance of learned KGE models by adding or removing triples to the input training graph. ∗Equal contribution by last authors. deposits owns Sam Joe allied_with deposits owns Sam Joe deposits owns Bob Target triple becomes False by adversarial (a) deletion (b) addition (a) (b) allied_with Figure 1: Adversarial attacks against KGE models for fraud detection. The knowledge graph consists of two types of entities - Person and BankAccount. The missing target triple to predict is (Sam, allied_with, Joe). Original KGE model predicts this triple as True. But a malicious attacker uses the instance attribution methods to either (a) delete an adversarial triple or (b) add an adversarial triple. Now, the KGE model predicts the missing target triple as False. Designing data poisoning attacks against KGE models poses two main challenges. First, to se- lect adversarial deletions or additions, we need to measure the impact of a candidate perturbation on the model’s predictions. But the naive approach of re-training a new KGE model for each candi- date perturbation is computationally prohibitive. Second, while the search space for adversarial dele- tions is limited to …

Content truncated to 500 words