Yushun Xie, Haiyan Wang, Willy Susilo · ACM SIGSAC Conference on Computer and Communications Security (CCS) 2026 · 2026
DOI: 10.1145/3830454.3846447
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).
Knowledge graph embedding (KGE) models are vulnerable to targeted data poisoning through the injection of a small number of training triples. Existing poisoning attacks typically employ a fixed attack mechanism for all targets, overlooking the heterogeneous vulnerabilities of individual target triples. We propose TAR-KGE, a target-level attack routing framework that adaptively selects attack mechanisms for each target. TAR-KGE first evaluates heterogeneous poison candidates using a unified target-relative virtual score derived from one-step virtual updates, and then constructs complementary poison sets via marginal greedy selection. Experiments on two benchmark datasets with four representative KGE models demonstrate that TAR-KGE consistently outperforms existing fixed-mechanism attacks and, relative to the corresponding clean models, reduces MRR by up to 41.9% and Hits@1 by up to 76.5%. These results validate the effectiveness of target-level attack routing and reveal the benefit of exploiting cross-attack complementarity.
No comments yet — start the discussion below.