Research map / Federated learning and privacy / Privacy attacks and secure learning
Vulnerability of Privacy-Preserving Visual Localization against Diffusion-based Attacks
Maxime Pietrantoni, Torsten Sattler, Gabriela Csurka
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 |
|---|---|---|---|
| invert with diffusion model instead of FFConv feed-forward network (high sparsity) | inversion attack · Cambridge Landmarks, SIFT 500 kpts | LPIPS: 0.44 -> 0.28 | Improved |
| invert with diffusion model instead of FFConv (medium sparsity) | inversion attack · Cambridge Landmarks, SIFT 1K kpts | FID: 198 -> 97 | Improved |
| invert with diffusion model instead of FFConv (low sparsity) | inversion attack · Cambridge Landmarks, SIFT 2K kpts | FID: 248 -> 130 | Improved |
| invert with diffusion model instead of FFUnetSA (SegLoc) | inversion attack · Cambridge Landmarks | FID: 189 -> 91 | Improved |
| invert with diffusion model instead of FFConv (RandLine SIFT) | inversion attack · Cambridge Landmarks | FID: 158 -> 73 | Improved |
| invert with diffusion model instead of FFTrans (CoordPerm SIFT) | inversion attack · Cambridge Landmarks | FID: 220 -> 76 | Improved |
| increase keypoint density from 500 to 2K per image (ablation) | FFConv inversion · Cambridge Landmarks | SSIM: 0.45 -> 0.38 | Hurt |
| increase keypoint density from 500 to 2K per image (ablation) | diffusion inversion · Cambridge Landmarks | SSIM: 0.49 -> 0.41 | Hurt |
| condition on DPT segmentations instead of SuperPoint descriptors | inversion attack · 7-Scenes | FID: 96.8 -> 222.5 | Improved |
| condition on DPT segmentations instead of SuperPoint descriptors | inversion attack · 7-Scenes | RD(SSIM): 0.71 -> 0.26 | Improved |
| use finer segmentation granularity (SAM2 instead of DPT) | inversion attack · 7-Scenes | LPIPS: 0.59 -> 0.50 | Hurt |
| use SegLoc segmentation instead of RandLine (SIFT) for localization | PPVL method · Cambridge Landmarks | median translation error (m): 0.07 -> 0.22 | Hurt |
Closest papers by meaning
- Privacy-Preserving Representations Are Not Enough: Recovering Scene Content From Camera Poses (CVPR 2023 · similarity 0.92)
- Towards Privacy-Preserving Visual Recognition via Adversarial Training: A Pilot Study (ECCV 2018 · similarity 0.88)
- Seeing Through the Weights: Privacy Leakage in Scene Coordinate Regression (ECCV 2026 · similarity 0.88)
- VPR-Cloak: A First Look at Privacy Cloak Against Visual Place Recognition (ICCV 2025 · similarity 0.87)
- Privacy Preserving Image-Based Localization (CVPR 2019 · similarity 0.87)
- Efficient Privacy-Preserving Visual Localization Using 3D Ray Clouds (CVPR 2024 · similarity 0.85)
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.
