Machine Unlearning on CIFAR-10 Standard (train val)
94.97Accuracy Retention (Accr)POUR-P
Evaluation results
| Method | Links | ||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| POUR-PBackbone=ResNet18, Access to retain set=No2025.11 | 94.97 | 0 | 99.99 | 0 | 1.01 | 56.67 | — | — | — | — | — | — | |
| FCSBackbone=ResNet18, Access to retain set=Yes2025.11 | 94.89 | 0.67 | 100 | 0.79 | 1 | 56 | 0.51 | 1 | 0.66 | 0.45 | 0.98 | 0.62 | |
| Retrained ModelBackbone=ResNet18, Access to retain set=Reference2025.11 | 94.68 | 0 | 99.98 | 0 | 1 | — | 0.26 | 0.97 | 0.84 | 1 | 1 | 1 | |
| Original ModelBackbone=ResNet18, Access to retain set=Reference2025.11 | 94.47 | 95.03 | 99.99 | 99.99 | 0.51 | 56.7 | 1 | 1 | 0 | 0.26 | 0.98 | 0.42 | |
| FinetuneBackbone=ResNet18, Access to retain set=Yes2025.11 | 93.96 | 0 | 100 | 0 | 0.99 | 54.6 | 0.32 | 0.97 | 0.8 | 0.63 | 0.97 | 0.76 | |
| POUR-DBackbone=ResNet18, Access to retain set=No2025.11 | 92.86 | 0.37 | 99.74 | 0.43 | 0.97 | 51.8 | 0.23 | 0.95 | 0.85 | 0.31 | 0.94 | 0.47 | |
| DELETEBackbone=ResNet18, Access to retain set=No2025.11 | 88.73 | 2.43 | 95.43 | 2.93 | 0.92 | 53.43 | 0.37 | 0.82 | 0.71 | 0.26 | 0.78 | 0.39 | |
| Random LabelBackbone=ResNet18, Access to retain set=No2025.11 | 87.42 | 23.2 | 93.13 | 25.13 | 0.75 | 54.07 | 0.29 | 0.86 | 0.78 | 0.24 | 0.84 | 0.37 | |
| Gradient AscentBackbone=ResNet18, Access to retain set=No2025.11 | 86.71 | 15.37 | 93.51 | 16.29 | 0.8 | 50.4 | 0.21 | 0.8 | 0.79 | 0.18 | 0.77 | 0.29 | |
| Boundary ExpandBackbone=ResNet18, Access to retain set=No2025.11 | 85.74 | 14.63 | 91.21 | 16.66 | 0.8 | 53 | 0.25 | 0.85 | 0.8 | 0.28 | 0.83 | 0.42 | |
| Boundary ShrinkBackbone=ResNet18, Access to retain set=No2025.11 | 85.3 | 12.33 | 90.81 | 13.96 | 0.81 | 53.07 | 0.25 | 0.85 | 0.8 | 0.28 | 0.84 | 0.42 |