Can You Trust the Vectors in Your Vector Database? Black-Hole Attack from Embedding Space Defects
Summary: Introduces the Black-Hole Attack: poisoning vector databases with a few centroid-adjacent vectors exploits centrality-driven hubness to dominate top-k results (up to 94.4% of queries). Existing hubness mitigation trades off accuracy, while detection defenses are path-dependent, leaving robust protection open.
(summarized by gpt-6-luna on Oct 08 2026)
@article{li_vldb27,
title = {{Can You Trust the Vectors in Your Vector Database? Black-Hole Attack from Embedding Space Defects}},
author = {Li, Hanxi and Zhou, Jianan and Lao, Jiale and Wang, Yibo and Ye, Zhengmao and Cao, Yang and Wang, Junfen and Tang, Mingjie},
journal = {PVLDB},
series = {{VLDB} '27},
volume = {20},
number = {1},
pages = {31--45},
doi = {10.14778/3845598.3845601},
url = {https://doi.org/10.14778/3845598.3845601},
year = {2027}
}
Incoming Citations (Sorted by Pagerank)
Showing 0 of 0 citing papers.
Rank
Citing Paper
Year
Venue
Pagerank
PreviousPage 1 / 1Next
Outgoing Citations (Sorted by Pagerank)
Showing 11 of 11 cited papers.
Citations counted here include only citations to other VLDB/SIGMOD/CIDR/PODS papers in this database.