Machine Unlearning on SVHN unlearning set Du
99.75AccuracyCertified Hessian
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| Certified HessianUnlearning Approach=Cert.-Hess., Proportion of remaining data Dr used (@)=0.352026.06 | 99.75 | 88.45 | 0.47 | |
| Pre-unlearnUnlearning Approach=Pre-unlearn2026.06 | 99.49 | — | — | |
| HessianUnlearning Approach=Hessian, Proportion of remaining data Dr used (@)=0.32026.06 | 99.24 | 89.35 | 0.42 | |
| FisherUnlearning Approach=Fisher, Proportion of remaining data Dr used (@)=0.352026.06 | 96.78 | 90.33 | 0.35 | |
| Relabel with Fine-tuneUnlearning Approach=Relab.+FT, Proportion of remaining data Dr used (@)=0.352026.06 | 92.87 | 89.21 | 0.44 | |
| Fine-tuneUnlearning Approach=Fine-tune, Proportion of remaining data Dr used (@)=0.32026.06 | 91.68 | 89.14 | 0.36 | |
| Adversarial RetrainUnlearning Approach=Adv. Retr., Proportion of remaining data Dr used (@)=0.32026.06 | 91.09 | 88.75 | 0.44 | |
| Gradient Ascent with Fine-tuneUnlearning Approach=GA+FT, Proportion of remaining data Dr used (@)=0.32026.06 | 90.26 | 89.35 | 0.42 | |
| RetrainUnlearning Approach=Retrain, Proportion of remaining data Dr used (@)=0.32026.06 | 89.06 | 88.81 | 0.53 | |
| Gradient AscentUnlearning Approach=GA, Proportion of remaining data Dr used (@)=0.052026.06 | 44.26 | 42.57 | 8.68 | |
| RelabelUnlearning Approach=Relabel, Proportion of remaining data Dr used (@)=0.052026.06 | 42.77 | 38.01 | 9.61 |