Research map / Federated learning and privacy / Privacy attacks and secure learning
Connecting Pixels to Privacy and Utility: Automatic Redaction of Private Information in Images
Read the paperCheck 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 |
|---|---|---|---|
| Use sequence labeling with word and image features for textual attributes instead of generic image segmentation | Textual attribute redaction · Textual attributes (test set) | mAP: 0.5 -> 26.8 | Helped |
| Fine-tune FCIS for privacy-region segmentation | Visual attribute redaction · Visual attributes (test set) | mAP: 13.5 -> 68.3 | Helped |
| Add text features to visual classification (ablation) | WSL:I+T · Multimodal attributes (test set) | mAP: 45.4 -> 56.2 | Helped |
| Use bag-of-words text encoding rather than LSTM or Conv1D encoding (ablation) | WSL:I+T · Multimodal attributes | mAP | Helped |
| Use salient-object localization instead of predicting the entire image (ablation) | WSL:I+T · Multimodal attributes (test set) | mAP: 56.2 -> 36.2 | Hurt |
| Refine text-region hulls with DenseCRF (ablation) | WSL:I+T · Multimodal attributes (test set) | mAP: 56.2 -> 53.6 | Hurt |
| Increase redaction beyond the ground-truth region (ablation) | Ground-truth segmentation redaction · All attributes | privacy | No clear effect |
| Redact fewer pixels than the ground-truth region (ablation) | Ground-truth segmentation redaction · All attributes | privacy | Hurt |
| Over-predict the redacted area for small textual regions (ablation) | Automatic redaction · Textual attributes | privacy-utility score: 81% | Helped |
| Redact ground-truth attribute regions rather than leave images unredacted (ablation) | Ground-truth segmentation redaction · All attributes | privacy: 94% | Helped |
Closest papers by meaning
- PerceptAnon: Exploring the Human Perception of Image Anonymization Beyond Pseudonymization for GDPR (ICML 2024 · similarity 0.84)
- Balancing Privacy and Performance: A Many-in-One Approach for Image Anonymization (AAAI 2025 · similarity 0.84)
- MultiPriv: Benchmarking Individual-Level Privacy Reasoning in Vision-Language Models (ICML 2026 · similarity 0.84)
- Private Attribute Inference from Images with Vision-Language Models (NeurIPS 2024 · similarity 0.84)
- PID: Prompt-Independent Data Protection Against Latent Diffusion Models (ICML 2024 · similarity 0.83)
- MaSS: Multi-attribute Selective Suppression for Utility-preserving Data Transformation from an Information-theoretic Perspective (ICML 2024 · 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.