Compare methods.
See the variation across seeds.
Explore the paper's reported means and standard deviations on Pins Face Recognition. Choose a model, scenario and metric, then download the values.
Explore the published tables
Dots show means. Whiskers show one standard deviation across seeds, not confidence intervals. Axis limits stay fixed when switching models or scenarios for the same metric.
| Method | Mean | Standard deviation |
|---|
How to read these results
Accuracy differences are unlearned minus retrained, measured in percentage points. A negative value means lower accuracy than the retrained reference. Closeness to zero describes agreement on that metric.
Full-class results average over five forget classes within each seed, then report the mean and standard deviation across ten seeds. The random-sample scenario uses a 0.1% forget set. Its forget evaluation uses the training forget samples; retain evaluation uses the test set.
Seeds 260–269 use the matched protocol (J = K = 1), so the spread combines stage effects rather than separating training, unlearning and evaluation variance. The paper's experiments ran on one NVIDIA L40S GPU. These are rounded aggregate summaries; individual seed values, checkpoint files and execution times are not included. A small metric difference alone does not establish complete removal of training-data influence.
Take the existing results with you
- Main table: accuracy differences and weight distances
- Appendix: activation distances and MIA differences
- Appendix: raw summaries and retrained reference
- All results and metadata (JSON)
- Source and measurement definitions
Values are transcribed from the existing paper. Standard deviations of differences are preserved from the paper, not estimated from rounded raw values.
Use SUPREME in your research
Run the published experiment configuration or register your own method, model, dataset or metric through the Python API.
Add your method · Discuss an experiment
Experiment reproduction requires the documented environment, data and compute. This viewer runs entirely in your browser.
Complete main table (available without JavaScript)
| Model | Scenario | Method | Forget difference | Retain difference | Layer distance |
|---|---|---|---|---|---|
| ResNet18 | fullclass | FT | 29.22 ± 5.21 | 2.52 ± 0.11 | 31.52 ± 0.22 |
| ResNet18 | fullclass | BadT | 0.26 ± 0.17 | -2.35 ± 0.27 | 31.72 ± 0.34 |
| ResNet18 | fullclass | UNSIR | 89.44 ± 1.98 | 2.51 ± 0.12 | 32.25 ± 0.34 |
| ResNet18 | fullclass | RL | 0.00 ± 0.00 | 2.58 ± 0.12 | 31.98 ± 0.34 |
| ResNet18 | fullclass | SSD | 1.97 ± 6.22 | -9.14 ± 7.43 | 31.55 ± 0.36 |
| ResNet18 | fullclass | LFSSD | 0.00 ± 0.00 | -3.66 ± 1.53 | 31.56 ± 0.34 |
| ResNet18 | random | FT | 2.78 ± 24.32 | -2.60 ± 15.25 | 37.70 ± 7.91 |
| ResNet18 | random | BadT | -35.00 ± 25.93 | -32.47 ± 29.83 | 41.99 ± 8.78 |
| ResNet18 | random | RL | -48.89 ± 34.03 | -4.59 ± 20.12 | 42.11 ± 8.72 |
| ResNet18 | random | SSD | -75.00 ± 19.47 | -79.97 ± 23.79 | 40.25 ± 9.06 |
| ResNet18 | random | LFSSD | -68.33 ± 15.28 | -58.61 ± 19.20 | 40.82 ± 8.72 |
| ViT | fullclass | FT | 0.03 ± 0.06 | -22.66 ± 23.47 | 105.29 ± 1.31 |
| ViT | fullclass | BadT | 17.90 ± 5.18 | -0.88 ± 0.10 | 33.97 ± 0.12 |
| ViT | fullclass | UNSIR | 35.52 ± 4.92 | -0.31 ± 0.13 | 39.38 ± 0.18 |
| ViT | fullclass | RL | 0.00 ± 0.00 | 0.20 ± 0.05 | 36.92 ± 0.18 |
| ViT | fullclass | SSD | 0.00 ± 0.00 | -2.40 ± 0.99 | 63.38 ± 2.58 |
| ViT | fullclass | LFSSD | 0.00 ± 0.00 | -3.88 ± 2.11 | 86.77 ± 6.07 |
| ViT | random | FT | 8.33 ± 5.40 | 1.42 ± 0.12 | 60.40 ± 0.19 |
| ViT | random | BadT | -8.33 ± 7.97 | 0.21 ± 0.52 | 32.76 ± 0.18 |
| ViT | random | RL | -76.67 ± 11.94 | 1.33 ± 0.12 | 35.71 ± 0.20 |
| ViT | random | SSD | -55.00 ± 37.99 | -54.09 ± 45.60 | 171.80 ± 84.31 |
| ViT | random | LFSSD | -89.44 ± 6.11 | -90.70 ± 6.59 | 231.91 ± 15.22 |