Research map / Image generation / Fast diffusion sampling
Data Circuit Breaker: Identifying Training, Test, and Generated Data in Image Generative Models
Bihe Zhao, Michel Meintz, Juangui Xu, Franziska Boenisch, Adam Dziedzic
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 |
|---|---|---|---|
| add autoencoder filtering and cross-signal stages (DCB) over ICAS | IAR MIA · MGI direct training IARs | TPR@1%FPR: 52.6 -> 87.4 | Improved |
| add DCB over CLiD for diffusion models | DM MIA · MGI direct training DMs | TPR@1%FPR: 39.7 -> 77.8 | Improved |
| add DCB cross-generator attribution over ICAS | VAR · MGI model derivative | TPR@1%FPR: 81.3 -> 95.8 | Improved |
| add DCB over ICAS | Stable Diffusion 2.1 · MGI model derivative | TPR@1%FPR: 64.4 -> 83.7 | Improved |
| add DCB over Delta for memorized samples | RAR-XXL · Memorized training vs regenerated | AUC: 61.8 -> 97.5 | Improved |
| add DCB over RMIA with shadow models | RAR-XXL · MGI model derivative | TPR@1%FPR: 82.7 -> 96.7 | Improved |
| add DCB over RMIA with shadow models | VAR-d30 · MGI model derivative | TPR@1%FPR: 76.4 -> 95.8 | Improved |
| add augmentation to encoder fine-tuning (ablation) | DCB Stage 1 · Adversarial attack eps=1 | TPR@1%FPR: 68.7 -> 97.4 | Improved |
| add augmentation to encoder fine-tuning (ablation) | DCB Stage 1 · JPEG (60) and Resize (0.5) | TPR@1%FPR: JPEG 91.7 -> 96.1; Resize 88.5 -> 98.4 | Improved |
| replace ground-truth prompts with BLIP2 or LLaVA captions (ablation) | DCB · SD1.4 derivative | TPR@1%FPR: 82.6 -> 81.5; 82.6 -> 78.7 | No clear effect |
| replace ground-truth prompts with BLIP2 or LLaVA captions (ablation) | ICAS · SD1.4 derivative | TPR@1%FPR: 63.5 -> 46.2; 63.5 -> 33.6 | Hurt |
| DCB versus ICAS on natural members vs non-members | RAR · MGI direct training | TPR@1%FPR: 72.5 -> 72.5 | No clear effect |
Closest papers by meaning
- On the Stability of Iterative Retraining of Generative Models on their own Data (ICLR 2024 · similarity 0.88)
- Self-Consuming Generative Models with Curated Data Provably Optimize Human Preferences (NeurIPS 2024 · similarity 0.87)
- The Unmet Promise of Synthetic Training Images: Using Retrieved Real Images Performs Better (NeurIPS 2024 · similarity 0.87)
- Intriguing Properties of Generative Classifiers (ICLR 2024 · similarity 0.87)
- Machine Unlearning for Image-to-Image Generative Models (ICLR 2024 · similarity 0.87)
- ImagenHub: Standardizing the evaluation of conditional image generation models (ICLR 2024 · similarity 0.87)
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.
