Research map / Image generation / Text-to-image diffusion
When Anonymity Breaks: Identifying Models Behind Text-to-Image Leaderboards
Ali Naseh, Anshuman Suri, Yuefeng Peng, Harsh Chaudhari, Alina Oprea, Amir Houmansadr
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 training-free centroid in embedding space instead of trained image classifier | deanonymization attack · 22-way identification | top-1 accuracy (%): 54.86 -> 90.86 | Improved |
| use embedding centroids instead of noise-residual fingerprints (Marra et al.) | deanonymization attack · 22-way identification | top-1 accuracy (%): 24.40 -> 90.86 | Improved |
| add text embedding to image-embedding classifier (ablation) | embedding MLP classifier · 22-way identification | top-1 accuracy (%): 43.00 -> 42.50 | No clear effect |
| increase reference images per model from 1 to 10 (ablation) | centroid attack · 22-way identification | top-1 accuracy: about 62% -> nearly same as 30 samples | Improved |
| increase reference images from 10 to 30 per model (ablation) | centroid attack · 22-way identification | top-1 accuracy | No clear effect |
| use modern image encoder (CLIP-ViT-bigG-14) instead of older encoders (ablation) | centroid attack · 22-way identification | top-1 accuracy | Improved |
| grow candidate pool from 5 to 15 models (ablation) | centroid attack · subsets of models | top-1 accuracy: 96.64% -> 93.86% | Hurt |
| select prompts with high distinguishability score (ablation) | centroid attack · prompt control | top-1 accuracy: 0.90 -> 100% | Improved |
| apply adversarial post-processing, eps=2 (ablation) | defense · 22-way identification | top-1 accuracy: 0.90 -> 0.75 | Improved |
| apply adversarial post-processing, eps=8 (ablation) | defense · 22-way identification | top-1 accuracy: 0.90 -> 0.46 | Improved |
| raise alpha quantile threshold from 0.80 to 0.95 (ablation) | threshold classifier · one-vs-rest, target only | accuracy: 0.926 -> 0.841 | Hurt |
Closest papers by meaning
- Position: Towards Implicit Prompt For Text-To-Image Models (ICML 2024 · similarity 0.85)
- T2ISafety: Benchmark for Assessing Fairness, Toxicity, and Privacy in Image Generation (CVPR 2025 · similarity 0.85)
- Personalization as a Shortcut for Few-Shot Backdoor Attack against Text-to-Image Diffusion Models (AAAI 2024 · similarity 0.84)
- Privacy Attacks on Image AutoRegressive Models (ICML 2025 · similarity 0.84)
- Variance-Based Membership Inference Attacks Against Large-Scale Image Captioning Models (CVPR 2025 · similarity 0.83)
- PerceptAnon: Exploring the Human Perception of Image Anonymization Beyond Pseudonymization for GDPR (ICML 2024 · 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.
