Research map / Federated learning and privacy / Privacy attacks and secure learning
Towards a Visual Privacy Advisor: Understanding and Predicting Privacy Risks in Images
Tribhuvanesh Orekondy, Bernt Schiele, Mario Fritz
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 the CNN end-to-end instead of training an SVM on frozen features (ResNet-50) | privacy attribute prediction · VISPR test | C-MAP: 40.50 -> 47.45 | Improved |
| fine-tune the CNN end-to-end instead of SVM (CaffeNet) | privacy attribute prediction · VISPR test | C-MAP: 37.93 -> 42.99 | Improved |
| fine-tune the CNN end-to-end instead of SVM (GoogleNet) | privacy attribute prediction · VISPR test | C-MAP: 39.88 -> 43.29 | Improved |
| use ResNet-50 instead of CaffeNet backbone (end-to-end) | privacy attribute prediction · VISPR test | C-MAP: 42.99 -> 47.45 | Improved |
| predict privacy risk directly with extra FC layers (PR-CNN) instead of from predicted attributes (AP-PR) | privacy risk prediction · VISPR personalized risk | L1-error: 0.656 -> 0.637 | Improved |
| PR-CNN instead of AP-PR at risk 4+ | privacy risk prediction · VISPR personalized risk (4+) | MAP: 77.89 -> 78.38 | Improved |
| PR-CNN instead of AP-PR at risk 1+ | privacy risk prediction · VISPR personalized risk (1+) | MAP: 94.94 -> 94.35 | Hurt |
| PR-CNN instead of AP-PR on safe images | privacy risk prediction · VISPR safe images | L1-error: 0.34 -> 0.58 | Hurt |
| PR-CNN instead of AP-PR on private images | privacy risk prediction · VISPR private images | L1-error: 0.84 -> 0.67 | Improved |
| use PR-CNN instead of users' own image-based judgment | privacy risk estimation · Humans vs machine | L1-error: 1.1 -> 1.03 | Improved |
| use AP-PR instead of users' own image-based judgment | privacy risk estimation · Humans vs machine | L1-error: 1.1 -> 1.27 | Hurt |
Closest papers by meaning
- PerceptAnon: Exploring the Human Perception of Image Anonymization Beyond Pseudonymization for GDPR (ICML 2024 · similarity 0.87)
- Private Attribute Inference from Images with Vision-Language Models (NeurIPS 2024 · similarity 0.86)
- MultiPriv: Benchmarking Individual-Level Privacy Reasoning in Vision-Language Models (ICML 2026 · similarity 0.85)
- EgoPrivacy: What Your First-Person Camera Says About You? (ICML 2025 · similarity 0.84)
- PID: Prompt-Independent Data Protection Against Latent Diffusion Models (ICML 2024 · similarity 0.84)
- Enhancing Privacy-Utility Trade-offs to Mitigate Memorization in Diffusion Models (CVPR 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.
