Machine Unlearning on CIFAR-100 (10% random data removal)
99.98Retain Accuracy (RA)Retrain
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| RetrainBackbone=ResNet-182025.09 | 99.98 | 24.26 | 75.56 | 48.44 | 0 | 27.37 | |
| FTBackbone=ResNet-182025.09 | 99.97 | 16.39 | 76.75 | 44.06 | 3.36 | 7.22 | |
| NoTBackbone=ResNet-182025.09 | 99.97 | 17.99 | 76.27 | 44.28 | 2.79 | 7.22 | |
| CoUnBackbone=ResNet-182025.09 | 99.97 | 22.01 | 72.88 | 47.82 | 1.39 | 9.63 | |
| RetrainBackbone=ViT2025.09 | 99.97 | 38.73 | 61.89 | 61.75 | 0 | 86.83 | |
| NegGrad+Backbone=ResNet-182025.09 | 99.96 | 30.09 | 75.46 | 47.72 | 1.67 | 7.62 | |
| l1-sparseBackbone=ResNet-182025.09 | 99.95 | 23.94 | 74.95 | 42.81 | 1.65 | 7.22 | |
| CoUnBackbone=ViT2025.09 | 99.91 | 36.81 | 56.49 | 53.92 | 3.8 | 19.29 | |
| NoTBackbone=ViT2025.09 | 99.89 | 20.29 | 61.82 | 43.55 | 9.2 | 8.68 | |
| NegGrad+Backbone=ViT2025.09 | 99.88 | 45.26 | 59.33 | 55 | 3.98 | 11.58 | |
| CoUnBackbone=VGG-162025.09 | 99.82 | 32.37 | 63.8 | 39.64 | 1.31 | 5.71 | |
| FTBackbone=ViT2025.09 | 99.78 | 10.83 | 61.12 | 31.5 | 14.78 | 5.79 | |
| RetrainBackbone=VGG-162025.09 | 99.75 | 33.23 | 67.07 | 40.69 | 0 | 15.58 | |
| l1-sparseBackbone=ViT2025.09 | 99.32 | 31.71 | 63.33 | 46.49 | 6.09 | 14.47 | |
| l1-sparseBackbone=VGG-162025.09 | 99.27 | 26.96 | 68.01 | 35.31 | 3.27 | 3.42 | |
| FTBackbone=VGG-162025.09 | 99.26 | 26.02 | 68.42 | 35.51 | 3.56 | 3.42 | |
| SalUnBackbone=ViT2025.09 | 99.18 | 38.01 | 54.78 | 69.24 | 4.03 | 5.1 | |
| SalUnBackbone=ResNet-182025.09 | 98.55 | 20.35 | 72.02 | 52.37 | 2.95 | 5.69 | |
| NoTBackbone=VGG-162025.09 | 96.17 | 30.11 | 66.75 | 36.47 | 2.81 | 4.28 | |
| NegGrad+Backbone=VGG-162025.09 | 94.92 | 35.44 | 65.54 | 40.67 | 2.15 | 3.42 | |
| SalUnBackbone=VGG-162025.09 | 92.65 | 33 | 64.04 | 42.85 | 3.13 | 2.79 |