Reproduce “Reconstructing training data from document understanding models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jérémie Dentan, Arnaud Paran, Aymen Shabou
This research develops and evaluates the first reconstruction attack, named CDMI, specifically designed for state-of-the-art layout-aware document understanding models. The work demonstrates that these models are vulnerable to data leakage, showing perfect reconstruction of up to 4.1% of sensitive fields in a one-shot attack and up to 22.5% in multi-shot scenarios with membership inference.
Key insights
- Memorization in DU models occurs early in training, even before overfitting, and is not primarily driven by data duplication.
- Both layout information and visual modalities significantly contribute to data memorization in document understanding models, making them more vulnerable than purely text-based models.
- The attack is generalizable across different transformer-based DU architectures like LayoutLM and BROS, highlighting a systemic vulnerability.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.