Machine Unlearning on CIFAR-10 single-class forgetting
90.5Accuracy FOriginal Model
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| Original ModelBackbone=ViT-S2025.03 | 90.5 | 90.57 | 72.7 | 75.73 | — | 84.6 | |
| Boundary ExpandBackbone=ViT-S2025.03 | 8.02 | 75.86 | 7.8 | 67.42 | 66.14 | 9.24 | |
| Saliency UnlearnBackbone=ViT-S2025.03 | 6.22 | 77.54 | 5.9 | 68.46 | 67.62 | 7.14 | |
| Bad TeacherBackbone=ViT-S2025.03 | 6.02 | 74.66 | 5.4 | 67.27 | 67.28 | 7.32 | |
| Boundary ShrinkBackbone=ViT-S2025.03 | 1.84 | 73.16 | 1.7 | 65.59 | 68.19 | 1.66 | |
| Random LabelBackbone=ViT-S2025.03 | 1.32 | 77.38 | 1.8 | 68.51 | 69.68 | 2.24 | |
| Influence UnlearnBackbone=ViT-S2025.03 | 0.98 | 87.09 | 0.9 | 75.2 | 73.46 | 0.78 | |
| Learn to UnlearnBackbone=ViT-S2025.03 | 0.46 | 85.11 | 0.3 | 73.2 | 72.8 | 0.36 | |
| Negative GradientBackbone=ViT-S2025.03 | 0.38 | 85.25 | 0.5 | 73.34 | 72.77 | 0.34 | |
| Retrain ModelBackbone=ViT-S2025.03 | 0 | 92.42 | 0 | 76.44 | 74.52 | 0 | |
| OursBackbone=ViT-S2025.03 | 0 | 90.74 | 0 | 77.21 | 74.89 | 0 |