Research map / Federated learning and privacy / Differential privacy
When Privacy Meets Recovery: The Overlooked Half of Surrogate-Driven Privacy Preservation for MLLM Editing
Siyuan Xu, Yibing Liu, Peilin Chen, Yung-Hui Li, Shiqi Wang, Sam Kwong
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 |
|---|---|---|---|
| train recovery model SOER instead of VISII, Prompt-Diffusion or Edit-Transfer | surrogate-to-original recovery · SPPE test, Textual | SSIM: 0.7483 -> 0.8049 (vs E-tran) | Improved |
| use SOER on unseen images with full-image surrogate | SOER · InstructPix2Pix subset | DirS: 0.1717 -> 0.2077 (vs E-tran) | Improved |
| SOER on source integrity vs baselines | SOER · SPPE Textual source integrity | SSIM: VISII 0.689 vs SOER 0.636 (GT 0.627) | Mixed |
| add privacy mask guidance (ablation) | SOER · SPPE | SSIM: 0.789 -> 0.802 | Improved |
| add edit map guidance (ablation) | SOER · SPPE | SSIM: 0.787 -> 0.802 | Improved |
| add text prompt guidance (ablation) | SOER · SPPE | SSIM: 0.714 -> 0.802 | Improved |
| use VAE-based transformer encoder instead of CNN for mask/edit maps (ablation) | SOER · SPPE | DirS: 0.592 -> 0.666 | Improved |
| use transformer-based spatial encoder instead of single CLS token (ablation) | SOER · SPPE | DirS: 0.620 -> 0.666 | Improved |
Closest papers by meaning
- MultiTrust: A Comprehensive Benchmark Towards Trustworthy Multimodal Large Language Models (NeurIPS 2024 · similarity 0.86)
- Conceal, Reconstruct, Jailbreak: Exploiting the Reconstruction--Concealment Tradeoff in MLLMs (NeurIPS 2026 · similarity 0.86)
- MultiPriv: Benchmarking Individual-Level Privacy Reasoning in Vision-Language Models (ICML 2026 · similarity 0.86)
- MaSS: Multi-attribute Selective Suppression for Utility-preserving Data Transformation from an Information-theoretic Perspective (ICML 2024 · similarity 0.85)
- Open LLMs are Necessary for Current Private Adaptations and Outperform their Closed Alternatives (NeurIPS 2024 · similarity 0.84)
- Blameless Users in a Clean Room: Defining Copyright Protection for Generative Models (NeurIPS 2025 · similarity 0.84)
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.
