Image Classification (Class Unlearning) on Fashion-MNIST
2.2Activation Distance (Dr)Retrain
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| RetrainBackbone=ResNet182024.11 | 2.2 | 2.39 | 2.23 | 2.2 | |
| NegGradBackbone=ResNet182024.11 | 1.65 | 1.8 | 1.56 | 1.55 | |
| Residual Feature AlignmentBackbone=ResNet182024.11 | 1.36 | 1.37 | 1.36 | 1.37 | |
| SCRUBBackbone=ResNet182024.11 | 1.35 | 1.36 | 1.32 | 1.43 | |
| NegGradBackbone=ViT2024.11 | 0.99 | 0.4 | 0.99 | 0.41 | |
| BadTBackbone=ResNet182024.11 | 0.75 | 0.77 | 0.78 | 2.7 | |
| FinetuneBackbone=ResNet182024.11 | 0.64 | 0.65 | 0.62 | 1.63 | |
| NegGradBackbone=ResNet182024.11 | 0.3 | 1.02 | 0.31 | 1.05 | |
| SCRUBBackbone=ResNet182024.11 | 0.01 | 0.42 | 0.03 | 0.4 | |
| BadTBackbone=ViT2024.11 | 0.01 | 0.67 | 0.03 | 0.62 | |
| FinetuneBackbone=ResNet182024.11 | 0 | 1.7 | 0.03 | 1.4 | |
| BadTBackbone=ResNet182024.11 | 0 | 1.04 | 0.03 | 0.89 | |
| Residual Feature AlignmentBackbone=ResNet182024.11 | 0 | 0.7 | 0.03 | 0.6 | |
| FinetuneBackbone=ViT2024.11 | 0 | 1.18 | 0.02 | 1.11 | |
| SCRUBBackbone=ViT2024.11 | 0 | 0.44 | 0.03 | 0.4 | |
| Residual Feature AlignmentBackbone=ViT2024.11 | 0 | 0.76 | 0.03 | 0.68 |