Paper 2026/2941

Recovery Attacks on Distance-Preserving Encrypted Vector Databases

Chen Gong, Xidian University
Xiangyu Wang, Xidian University
Dan Zhu, Northwestern Polytechnical University
Bowen Zhao, Guangzhou University
Jie Li, Huawei Technology
Xinghua Li, Xidian University
Jianfeng Ma, Xidian University
Abstract

Private retrieval-augmented generation (RAG) systems based on distance-comparison-preserving encryption (DCPE) are widely adopted in industry to protect sensitive embeddings while maintaining low-latency search. However, DCPE inherently leaks the relational geometry of stored embeddings by preserving relative distance comparisons across ciphertexts, making its security in practical deployments questionable. This paper demonstrates that the geometric structure DCPE preserves to enable nearest-neighbor search directly exposes the underlying data manifold to full database recovery from snapshots. We introduce two practical recovery attacks: (1) Attack A, which requires only a leaked snapshot, the public encoder, and an auxiliary candidate corpus, achieving 100\% record re-identification; and (2) Attack B, which requires no candidate corpus and uses a single known plaintext-ciphertext pair to resolve the secret scale in DCPE, reconstructing embeddings with 0.997 mean cosine similarity and restoring a functional search index. Furthermore, our evaluation indicates that existing defensive countermeasures offer minimal protection. Crucially, full-database reconstruction is computationally efficient, incurring an execution overhead that is even lower than that of the original encryption pass. Therefore, leaked DCPE snapshots must be treated as plaintext disclosures, underscoring the urgent need to explore and adopt truly secure cryptographic alternatives for private RAG.

Metadata
Available format(s)
PDF
Category
Attacks and cryptanalysis
Publication info
Preprint.
Keywords
Encrypted vector databasesDistance-preserving encryptiondatabase recovery attacksretrieval-augmented generation
Contact author(s)
cgong_11 @ stu xidian edu cn
wangxiangyu01 @ xidian edu cn
zhudan @ nwpu edu cn
bwinzhao @ gmail com
lijie303 @ huawei com
xhli1 @ mail xidian edu cn
jfma @ mail xidian edu cn
History
2026-11-22: approved
2026-11-19: received
See all versions
Short URL
https://ia.cr/2026/2941
License
Creative Commons Attribution-ShareAlike
CC BY-SA

BibTeX

@misc{cryptoeprint:2026/2941,
      author = {Chen Gong and Xiangyu Wang and Dan Zhu and Bowen Zhao and Jie Li and Xinghua Li and Jianfeng Ma},
      title = {Recovery Attacks on Distance-Preserving Encrypted Vector Databases},
      howpublished = {Cryptology {ePrint} Archive, Paper 2026/2941},
      year = {2026},
      url = {https://eprint.iacr.org/2026/2941}
}
Note: In order to protect the privacy of readers, eprint.iacr.org does not use cookies or embedded third party content.