Research map / Federated learning and privacy / Privacy attacks and secure learning
VizWiz-Priv: A Dataset for Recognizing the Presence and Purpose of Private Visual Information in Images Taken by Blind People
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train on uncorrupted VizWiz-Priv images | ResNet-50 privacy recognition · VizWiz-Priv test (uncorrupted) | AP: a 2% boost | Helped |
| Fine-tune the VISPR-trained model on uncorrupted VizWiz-Priv images | ResNet-50 privacy recognition · VizWiz-Priv test (uncorrupted) | AP: a 3% boost | Helped |
| Train directly on VizWiz-Priv with inpainted private-region holes | ResNet-50 privacy recognition · VizWiz-Priv test (uncorrupted) | AP: 77.97% versus 68.47% | Hurt |
| Replace private-region holes with ImageNet mean values during training | ResNet-50 privacy recognition · VizWiz-Priv test (uncorrupted) | AP: 77.97% versus 69.6% | Hurt |
| Fine-tune the VISPR-trained model on VizWiz-Priv with inpainted private-region holes | ResNet-50 privacy recognition · VizWiz-Priv test (uncorrupted) | AP: 77.97% to 75.62% | Hurt |
| Fine-tune the VISPR-trained model on VizWiz-Priv with ImageNet mean values in holes | ResNet-50 privacy recognition · VizWiz-Priv test (uncorrupted) | AP: 77.97% to 71.94% | Hurt |
| Train jointly on VISPR and VizWiz-Priv with inpainted private-region holes | ResNet-50 privacy recognition · VizWiz-Priv test (uncorrupted) | AP: 77.97% to 80.22% | Helped |
| Train jointly on VISPR and VizWiz-Priv with ImageNet mean values in holes | ResNet-50 privacy recognition · VizWiz-Priv test (uncorrupted) | AP: 77.97% to 78.34% | Helped |
| Train on uncorrupted VizWiz-Priv rather than VISPR | ResNet-50 privacy recognition · VISPR test | AP: 96.78% to 88.08% | Hurt |
| Use the question text alone | Question privacy-purpose recognition · VizWiz-Priv-VQA test | AP: 30.25 percentage point boost over the status quo | Helped |
| Use the original image alone | Question privacy-purpose recognition · VizWiz-Priv-VQA test | AP: 24.9 percentage point boost over the status quo | Helped |
| Use question and image features jointly rather than either alone | Question privacy-purpose recognition · VizWiz-Priv-VQA test | AP: a further 3.5 percentage point boost | Helped |
Closest papers by meaning
- Private Attribute Inference from Images with Vision-Language Models (NeurIPS 2024 · similarity 0.86)
- PerceptAnon: Exploring the Human Perception of Image Anonymization Beyond Pseudonymization for GDPR (ICML 2024 · similarity 0.85)
- MultiPriv: Benchmarking Individual-Level Privacy Reasoning in Vision-Language Models (ICML 2026 · similarity 0.85)
- Image Safeguarding: Reasoning with Conditional Vision Language Model and Obfuscating Unsafe Content Counterfactually (AAAI 2024 · similarity 0.84)
- Explaining CLIP's Performance Disparities on Data from Blind/Low Vision Users (CVPR 2024 · similarity 0.84)
- EgoPrivacy: What Your First-Person Camera Says About You? (ICML 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.