Adversarial Attack against Cross-lingual Knowledge Graph Alignment

Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 5320–5337 November 7–11, 2021. c⃝2021 Association for Computational Linguistics 5320 Adversarial Attack against Cross-lingual Knowledge Graph Alignment Zeru Zhang1, Zijie Zhang1, Yang Zhou1, Lingfei Wu2, Sixing Wu3, Xiaoying Han1, Dejing Dou4,5, Tianshi Che1, Da Yan6 1Auburn University, 2JD.COM Silicon Valley Research Center, 3Peking University, 4University of Oregon, 5Baidu Research, 6University of Alabama at Birmingham {zeruzhang, zijiezhang, yangzhou, xzh0003, tianshiche}@auburn.edu, lwu@email.wm.edu, wusixing@pku.edu.cn, dou@cs.uoregon.edu, yanda@uab.edu Abstract Recent literatures have shown that knowledge graph (KG) learning models are highly vul- nerable to adversarial attacks. However, there is still a paucity of vulnerability analyses of cross-lingual entity alignment under adversar- ial attacks. This paper proposes an adversarial attack model with two novel attack techniques to perturb the KG structure and degrade the quality of deep cross-lingual entity alignment. First, an entity density maximization method is employed to hide the attacked entities in dense regions in two KGs, such that the de- rived perturbations are unnoticeable. Second, an attack signal amplification method is devel- oped to reduce the gradient vanishing issues in the process of adversarial attacks for further improving the attack effectiveness. 1 Introduction Today, multilingual knowledge graphs (KGs), such as WordNet (Miller, 1992), DBpedia (Auer et al., 2007), YAGO (Hoffart et al., 2011), and Concept- Net (Speer et al., 2017), are becoming essential sources of knowledge for various AI-related appli- cations, e.g., personal assistants, medical diagnosis, and online question answering. Cross-lingual entity alignment between multilingual KGs is a powerful tool that align the same entities in different mono- lingual KGs together, automatically synchronize different language-specific KGs and revolutionize the understanding of these ubiquitous multilingual KGs in a transformative manner (Xu et al., 2020b; Sun et al., 2020a; Berrendorf et al., 2021b,a). Unfortunately, real-world KGs are typically noisy due to two main reasons: (1) massive fake information injected by malicious parties and users on online encyclopedia websites (e.g., Wikipedia (Wik) and Answers.com (Ans)), social networks (e.g., Twitter and Facebook), online com- munities (e.g., Reddit and Yahoo Answers), news websites, and search engines that usually serve as data sources of the KGs; and (2) direct adversar- ial attacks on the KGs. Google Knowledge Graph has been criticized for providing answers without source attribution or citation, and thus undermines people’s ability to verify information and to de- velop well-informed opinions (Dewey, 2016). Recent studies have shown that KG learning models remain highly sensitive to adversarial at- tacks, i.e., carefully designed small perturbations in KGs can cause the models to produce wrong prediction results, including knowledge graph em- bedding (Minervini et al., 2017; Pujara et al., 2017; Pezeshkpour et al., 2019; Zhang et al., 2019; Baner- jee et al., 2021) and knowledge graph-based dia- logue generation (Xu et al., 2020a). However, exist- ing techniques focus on the adversarial attacks on single KG learning tasks. These techniques cannot be directly utilized to attack the cross-lingual entity alignment models, as they have to analyze relations within and across KGs. Two critical questions still keep unsolved: (1) …

Content truncated to 500 words