Adel Memariani, Arnab Sharma, Michael Röder, Axel-Cyrille Ngonga Ngomo · ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 2026
DOI: 10.1145/3799682.3840996
Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).
Adversarial attacks pose a major threat to link prediction models. Carefully chosen triple-level perturbations can exploit weaknesses in the learned representations and change the model's ranking decisions. Despite their importance, adversarial attacks on knowledge graph embeddings (KGE) models remain less studied than their counterparts in areas such as image classification. One reason is that KGE inputs are discrete relational triples, not continuous pixel values. Moreover, KGE scores are computed within multi-relational prediction settings, where each relation can produce a different score distribution. As a result, score-based attack strategies cannot be directly transferred to KGE models. Studying training-data perturbations therefore both exposes vulnerabilities and identifies informative graph edits that can improve performance. In this work, we use adversarial triple modifications as controlled probes of KGE decision behavior. We introduce a decision-margin perspective that measures how clearly each training triple is ranked above its strongest negative alternatives. Our results reveal systematic vulnerability patterns, as well as improvement opportunities, that are shared across different KGE models. At the same time, these effects are not determined by the embedding model alone. The impact of training data perturbations also depends on the structure of the underlying knowledge graph. Overall, our study reframes adversarial probes as diagnostic instruments for KGE models.
ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 0 citations
No comments yet — start the discussion below.