Class Unlearning on CIFAR-10
100MIA Accuracy (Mean)Original
Evaluation results
| Method | Links | ||
|---|---|---|---|
| OriginalBackbone=ResNet182024.11 | 100 | — | |
| NegGradBackbone=ResNet182024.11 | 100 | — | |
| OriginalBackbone=ViT2024.11 | 100 | — | |
| NegGradBackbone=ViT2024.11 | 100 | — | |
| FinetuneBackbone=ViT2024.11 | 100 | — | |
| SCRUBBackbone=ViT2024.11 | 86 | — | |
| CF-kBackbone=ResNet2023.02 | 75.73 | 0.34 | |
| FinetuneBackbone=ResNet2023.02 | 75.57 | 0.69 | |
| FinetuneBackbone=ResNet182024.11 | 73 | — | |
| NegGrad+Backbone=All-CNN2023.02 | 72 | 0 | |
| OriginalBackbone=ResNet2023.02 | 71.1 | 0.67 | |
| NegGrad+Backbone=ResNet2023.02 | 69.57 | 1.19 | |
| CF-kBackbone=All-CNN2023.02 | 69 | 2 | |
| FinetuneBackbone=All-CNN2023.02 | 68 | 1 | |
| OriginalBackbone=All-CNN2023.02 | 66.5 | 0.5 | |
| EU-kBackbone=All-CNN2023.02 | 66.5 | 3.5 | |
| Bad-TBackbone=All-CNN2023.02 | 63.4 | 1.2 | |
| SCRUBBackbone=ResNet182024.11 | 62 | — | |
| RetrainBackbone=All-CNN2023.02 | 55 | 4 | |
| EU-kBackbone=ResNet2023.02 | 54.2 | 2.27 | |
| Bad-TBackbone=ResNet2023.02 | 54 | 1.1 | |
| RetrainBackbone=ViT2024.11 | 53 | — | |
| SCRUBBackbone=ResNet2023.02 | 52.2 | 1.71 | |
| SCRUB+RBackbone=ResNet2023.02 | 52.2 | 1.71 | |
| SCRUBBackbone=All-CNN2023.02 | 52 | 0 | |
| SCRUB+RBackbone=All-CNN2023.02 | 52 | 0 | |
| RetrainBackbone=ResNet2023.02 | 49.33 | 1.67 | |
| RetrainBackbone=ResNet182024.11 | 33 | — | |
| BadTBackbone=ResNet182024.11 | 0 | — | |
| Residual Feature Alignment UnlearningBackbone=ResNet182024.11 | 0 | — | |
| BadTBackbone=ViT2024.11 | 0 | — | |
| Residual Feature Alignment UnlearningBackbone=ViT2024.11 | 0 | — |