Inverting Multi-Vector Visual Document Indices
The paper says stored patch-vector indices can leak enough information to reconstruct the original document page.
The authors treat index inversion as conditional document image generation and test it on ViDoRe v3. Pages rebuilt from raw indices recover 47% of words and 45% of sensitive tokens, and retrieve their source page first 98.4% of the time. Token pooling and shuffling reduce word recall to about 8%, but a learned order-restoration attack brings shuffled-index source-page retrieval back to 93.5%. The paper argues these indices should be protected like the documents they encode. HF Daily Papers' note
The authors treat index inversion as conditional document image generation and test it on ViDoRe v3. Pages rebuilt from raw indices recover 47% of words and 45% of sensitive tokens, and retrieve their source page first 98.4% of the time. Token pooling and shuffling reduce word recall to about 8%, but a learned order-restoration attack brings shuffled-index source-page retrieval back to 93.5%. The paper argues these indices should be protected like the documents they encode. HF Daily Papers' note
score 5