Machine Unlearning on AGNews unlearning set Du
99.89AccuracyPre-unlearn
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| Pre-unlearn2026.06 | 99.89 | — | — | |
| Certified Hessian Forgettingthreshold=0.32026.06 | 99.88 | 88.06 | 0.52 | |
| Fisher Forgettingthreshold=0.352026.06 | 99.79 | 89.1 | 0.46 | |
| Relabel with Fine-tunethreshold=0.32026.06 | 91.42 | 90.87 | 0.3 | |
| Gradient Ascent with Fine-tunethreshold=0.32026.06 | 91.1 | 90.79 | 0.3 | |
| Fine-tunethreshold=0.22026.06 | 91.06 | 89.17 | 0.35 | |
| Adversarial Retrainthreshold=0.32026.06 | 89.68 | 90.86 | 0.28 | |
| Retrainthreshold=0.32026.06 | 89.44 | 90.88 | 0.29 | |
| Hessian Forgettingthreshold=0.32026.06 | 88.45 | 82.99 | 0.88 | |
| Gradient Ascentthreshold=0.052026.06 | 67.53 | 60.11 | 1.53 | |
| Relabelthreshold=0.052026.06 | 63.1 | 52.54 | 2.34 |