Research map / Image generation / Autoregressive and latent generation
PerceptAnon: Exploring the Human Perception of Image Anonymization Beyond Pseudonymization for GDPR
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Builds a 4,000-image human-annotated dataset of anonymized images (COCO, VOC, LFW, CelebA-HQ) and trains a CNN-based learned metric, PerceptAnon, to predict human-perceived anonymity of whole images beyond pseudonymization.
Problem
Existing image anonymization evaluation relies on localized detection and re-identification rates, which miss privacy cues in the whole image (locations, objects, backgrounds) and are binary, so they do not reflect human-perceived anonymity under GDPR.
Approach
Collect human anonymity scores on anonymized images (standalone and paired with originals) and train a ResNet-based regressor/classifier on these scores so the metric captures whole-image privacy cues, comparing regression vs. classification and label granularity.
Key results
- PerceptAnon (ResNet50, regression) reaches Spearman 0.8817 with human anonymity scores on HA1 (all setup), above LPIPS (0.7675), SSIM (0.8358 magnitude, as reported), and FID (0.6578).
- On HA2, PerceptAnon attains Spearman 0.8421 on the all setup, exceeding the best traditional metric (LPIPS 0.7622 / SSIM 0.7655).
- Classification with finer granularity (10, 5, 3 classes) correlates better with human judgments than binary classification, and the authors report classification is generally more effective than regression, especially in Kendall's tau.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train PerceptAnon on human annotations rather than use traditional image assessment metrics | HA1 All · HA1 All | Spearman's ρ with human anonymity scores: -0.8358 -> 0.8817 | Helped |
| Train PerceptAnon on human annotations rather than use traditional image assessment metrics | HA2 All · HA2 All | Spearman's ρ with human anonymity scores: -0.7655 -> 0.8421 | Helped |
| Use classification rather than regression to learn human anonymity scores (ablation) | PerceptAnon · All | Kendall's correlation with human scores | Helped |
| Use finer score granularity (10-, 5-, or 3-class) rather than binary labels (ablation) | PerceptAnon · All | Correlation with human perception | Helped |
| Assess anonymized images without an original reference (HA1) rather than compare original-anonymized pairs (HA2) (ablation) | PerceptAnon · Shared blur and inpaint images | Spearman's ρ with human scores: 0.8912; 0.8346; 0.8848 | Mixed |
Surprising findings
- Global images (with background retained) receive consistently lower anonymity scores than local images, showing humans weigh background context heavily.
- Weighted cross-entropy to handle the skew of human scores toward high anonymity gave no substantial improvement.
Limitations
- Reliance on annotated data quality and diversity; limited cultural diversity of perceptions.
- Human scores cluster toward the high-anonymity end, limiting score range; weighted CE did not help.
- Dataset scale is modest (4,000 images), limiting robustness and generalization claims.
Experimental setup
Datasets: MS-COCO PASCAL VOC LFW CelebA-HQ WIDER Face
Benchmarks: PerceptAnon dataset HA1 PerceptAnon dataset HA2
Metrics: Spearman correlation Kendall correlation
Baselines: MSE PSNR SSIM LPIPS FID
Builds on: SemSim DreamSim DeepPrivacy2 LPIPS
Closest papers by meaning
- MultiPriv: Benchmarking Individual-Level Privacy Reasoning in Vision-Language Models (ICML 2026 · similarity 0.88)
- Balancing Privacy and Performance: A Many-in-One Approach for Image Anonymization (AAAI 2025 · similarity 0.86)
- EgoPrivacy: What Your First-Person Camera Says About You? (ICML 2025 · similarity 0.86)
- Private Attribute Inference from Images with Vision-Language Models (NeurIPS 2024 · similarity 0.86)
- PID: Prompt-Independent Data Protection Against Latent Diffusion Models (ICML 2024 · 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.