WildChat / research hub

Uses WildChat

When Shuffling Fails: Token Recovery from Obfuscated KV Caches

Chengzhi Yi, Xiao Cheng, Wei Song, Haoyu Wang

arXiv preprint · First arXiv submission

Studies how repeated-token calibration defeats shuffled KV-cache obfuscation when secrets are reused across requests. Uses WildChat prompts to test token recovery and evaluates token ordering separately.

Data used

3,000 real user prompts from WildChat, used to evaluate recovery from obfuscated KV caches.

Findings

Reports 100% token-multiset recovery on GPU for the evaluated configurations. Recovering the token multiset does not establish exact original order. Refreshing row-level secrets interrupts transfer of the calibrated attack across requests.

Sections III–V: mechanism, attack assumptions, and WildChat evaluation. The abstract and evaluation section differ in model/configuration counts; this entry reports the shared 3,000-prompt scope. Cites WildChat. Read the source.

Dates and versions

First arXiv version. ArXiv version history.

arXiv preprint.

Cite this paper

Chengzhi Yi, Xiao Cheng, Wei Song, Haoyu Wang. 2026. When Shuffling Fails: Token Recovery from Obfuscated KV Caches. arXiv preprint.

Show BibTeX
@misc{kvcachetokenrecovery2026,
  title = {{When Shuffling Fails: Token Recovery from Obfuscated KV Caches}},
  author = {Yi, Chengzhi and Cheng, Xiao and Song, Wei and Wang, Haoyu},
  year = {2026},
  doi = {10.48550/arXiv.2611.24276},
  eprint = {2611.24276},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.24276}
}

Reuse this entry