Research map / Federated learning and privacy / Differential privacy
Unsafe2Safe: Controllable Image Anonymization for Downstream Utility
Minh Dinh, SouYoung Jin
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| fine-tune InstructPix2Pix on the anonymization dataset (c_edit prior) (ablation) | U2S InstructPix2Pix · Caltech101 accuracy | top-1 accuracy: 94.315 -> 95.116 | Improved |
| fine-tune InstructPix2Pix on the anonymization dataset (c_edit prior) (ablation) | U2S InstructPix2Pix · MIT Indoor67 | FaceSim: 0.2909 -> 0.2735 | Improved |
| add Safe Attention to the fine-tuned InstructPix2Pix (ablation) | U2S InstructPix2Pix fine-tuned · MIT Indoor67 | FaceSim: 0.2735 -> 0.2465 | Improved |
| add Safe Attention to the fine-tuned InstructPix2Pix (ablation) | U2S InstructPix2Pix fine-tuned · Caltech101 | race entropy: 0.7997 -> 0.8306 | Improved |
| use the edit instruction c_edit instead of public caption as FlowEdit prior (ablation) | U2S FlowEdit · MIT Indoor67 | top-1 accuracy: 80.746 -> 77.090 | Hurt |
| use LLM-composed prior instead of c_public with FlowEdit (ablation) | U2S FlowEdit · Caltech101 | FaceSim: 0.4881 -> 0.3428 | Improved |
| use LLM-composed prior instead of class prior with FlowEdit (ablation) | U2S FlowEdit · Caltech101 | race entropy: 0.4051 -> 0.8751 | Improved |
| use LLM-composed prior instead of c_edit with FlowEdit (ablation) | U2S FlowEdit · Caltech101 | top-1 accuracy: 94.792 -> 92.884 | Hurt |
| train on only the safe subset (discard flagged images) instead of anonymizing them (ablation) | ImageMAE classifier · Caltech101 | top-1 accuracy: 94.27 -> 83.49 | Hurt |
| train on only the safe subset (discard flagged images) instead of anonymizing them (ablation) | ImageMAE classifier · MIT Indoor67 | top-1 accuracy: 83.88 -> 51.12 | Hurt |
| train VQA model on U2S (FE, LLM) anonymized images instead of raw | Qwen3VL-2B on OK-VQA · OK-VQA | VQA accuracy: 0.6064 -> 0.7093 | Improved |
| train captioner on U2S (FE, LLM) anonymized images instead of raw | BLIP-2 on MS-COCO · MS-COCO captioning | CIDEr: 1.413 -> 1.363 | No clear effect |
Closest papers by meaning
- Enhancing Privacy-Utility Trade-offs to Mitigate Memorization in Diffusion Models (CVPR 2025 · similarity 0.87)
- T2ISafety: Benchmark for Assessing Fairness, Toxicity, and Privacy in Image Generation (CVPR 2025 · similarity 0.87)
- MultiPriv: Benchmarking Individual-Level Privacy Reasoning in Vision-Language Models (ICML 2026 · similarity 0.86)
- PID: Prompt-Independent Data Protection Against Latent Diffusion Models (ICML 2024 · similarity 0.86)
- Perturb a Model, Not an Image: Towards Robust Privacy Protection via Anti-Personalized Diffusion Models (NeurIPS 2025 · similarity 0.86)
- Balancing Privacy and Performance: A Many-in-One Approach for Image Anonymization (AAAI 2025 · similarity 0.86)
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.
