Research map / Image generation / Autoregressive and latent generation
Privacy Assessment on Reconstructed Images: Are Existing Evaluation Metrics Faithful to Human Perception?
Xiaoxiao Sun, Nidham Gazagnadou, Vivek Sharma, Lingjuan Lyu, Hongdong Li, Liang Zheng
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 |
|---|---|---|---|
| use SemSim instead of PSNR | privacy metric · CIFAR-100 | Kendall tau with human ranking: 0.4725 -> -0.7143 | Improved |
| use SemSim instead of best baseline FID | privacy metric · CIFAR-100 | Kendall tau: -0.5604 -> -0.7143 | Improved |
| use SemSim instead of best baseline MSE | privacy metric · Caltech-101 | Kendall tau: -0.5525 -> -0.6889 | Improved |
| use SemSim instead of best baseline LPIPS | privacy metric · Imagenette | Kendall tau: -0.5922 -> -0.7012 | Improved |
| use SemSim instead of best baseline PSNR | privacy metric · CelebA | Kendall tau: 0.5604 -> -0.6923 | Improved |
| use SemSim instead of best baseline SSIM on fine-grained dataset | privacy metric · Stanford Dogs | Kendall tau: 0.3048 -> -0.5938 | Improved |
| evaluate SemSim trained on InvGrad against unseen DLG attack | SemSim · CIFAR-100 | Kendall tau: -0.5237 -> -0.7342 | Improved |
| evaluate SemSim on unseen CAFE attack | SemSim · CIFAR-100 | Kendall tau: 0.5392 -> -0.7510 | Improved |
| evaluate SemSim on unseen GradAttack | SemSim · CIFAR-100 | Kendall tau: -0.4819 -> -0.7260 | Improved |
| reduce number of human-annotated training samples down to 50 (ablation) | SemSim · CIFAR-100 | correlation with human: 50 samples | Mixed |
| use simple LeNet backbone instead of ResNet50 (ablation) | SemSim · CIFAR-100 | Kendall tau: 0.65 vs 0.5604 | Mixed |
| triplet loss instead of contrastive loss (ablation) | SemSim · CIFAR-100 | correlation with human | No clear effect |
Closest papers by meaning
- PerceptAnon: Exploring the Human Perception of Image Anonymization Beyond Pseudonymization for GDPR (ICML 2024 · similarity 0.89)
- MultiPriv: Benchmarking Individual-Level Privacy Reasoning in Vision-Language Models (ICML 2026 · similarity 0.87)
- Private Attribute Inference from Images with Vision-Language Models (NeurIPS 2024 · similarity 0.87)
- PID: Prompt-Independent Data Protection Against Latent Diffusion Models (ICML 2024 · similarity 0.85)
- EgoPrivacy: What Your First-Person Camera Says About You? (ICML 2025 · similarity 0.85)
- How Private is Private? A Comparative Study for Face De-Identification (NeurIPS 2026 · 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.
