Metrics
Metrics evaluate how well a technique has erased a concept and how much it has degraded general image quality.
Erasure metrics
| Metric | Description |
|---|---|
| ASR I2P | Attack success rate on the I2P benchmark using a nudity/violence classifier |
| ASR P4D | Attack success rate using P4D adversarial prompts |
| ASR Ring-a-Bell | Attack success rate using Ring-a-Bell concept vectors |
| ASR MMA-Diffusion | Attack success rate using MMA-Diffusion adversarial prompts |
| ERR | Erasure recall rate — fraction of concept prompts successfully suppressed |
Quality metrics
| Metric | Description |
|---|---|
| FID | Fréchet Inception Distance — distributional similarity to a reference set |
| CLIP Score | Prompt–image alignment measured via CLIP embeddings |
| TIFA | Text-image faithfulness via VQA |
| UA-IRA | Unlearning accuracy vs. image retention accuracy trade-off |