Summary of Extracting Training Data From Document-based Vqa Models, by Francesco Pinto et al.
Extracting Training Data from Document-Based VQA Models
by Francesco Pinto, Nathalie Rauschmayr, Florian Tramèr, Philip Torr, Federico Tombari
First submitted to arxiv on: 11 Jul 2024
Categories
- Main: Computer Vision and Pattern Recognition (cs.CV)
- Secondary: Machine Learning (cs.LG)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary The proposed study examines the capabilities of Vision-Language Models (VLMs) in Visual Question Answering tasks, focusing on their potential to memorize and regurgitate responses. The researchers demonstrate that these models can recall training samples even when relevant visual information is removed, which raises privacy concerns as they could potentially divulge sensitive Personal Identifiable Information (PII). To quantify the extractability of information, the study conducts controlled experiments and differentiates between memorization and generalization capabilities. Additionally, it investigates factors influencing memorization across various state-of-the-art models and proposes an effective countermeasure to prevent PII extractability. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Vision-Language Models can remember answers from training samples and repeat them even when visual information is removed. This means they could accidentally reveal personal information like names or addresses. The researchers tested this by hiding some details in the training set, then seeing if the models would still recall them. They found that this memory effect comes from both memorization and generalizing what they learned. To prevent this, they came up with a solution to help protect sensitive data. |
Keywords
» Artificial intelligence » Generalization » Question answering » Recall
