Intermediate hidden states transmitted during collaborative LVLM inference leak enough information to reconstruct input images and text with near-perfect token accuracy.
Unveiling the vulnerability of private fine-tuning in split-based frameworks for large language models: A bidirectionally enhanced attack,
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
citation-role summary
background 1
citation-polarity summary
fields
cs.CR 1years
2026 1verdicts
CONDITIONAL 1roles
background 1polarities
unclear 1representative citing papers
citing papers explorer
-
Inverting the Hidden: Unveiling Multimodal Privacy Leakage in Collaborative LVLM Inference
Intermediate hidden states transmitted during collaborative LVLM inference leak enough information to reconstruct input images and text with near-perfect token accuracy.