Machine Unlearning Utility and Membership Inference on CIFAR-10 (test)
100Unlearning Accuracy (UA)Retrain
Evaluation results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| Retrainforget_class=02026.06 | 100 | 100 | 92.5 | 50 | — | — | 0 | |
| l1-sparseforget_class=02026.06 | 100 | 97.9 | 92.3 | — | — | — | 0 | |
| GAforget_class=02026.06 | 99.9 | 38.9 | 38.2 | — | — | — | 0.1 | |
| SalUnforget_class=02026.06 | 99.3 | 99.4 | 93.5 | — | — | — | 0.7 | |
| IUforget_class=02026.06 | 97 | 94.8 | 89.1 | — | — | — | 3 | |
| PURGEforget_class=0, objective=kl_retain2026.06 | 91.6 | 96.3 | 84.5 | 49.6 | — | — | 8.4 | |
| RLforget_class=02026.06 | 89.3 | 99.9 | 94.5 | — | — | — | 10.7 | |
| FTforget_class=02026.06 | 31.7 | 99.9 | 94.8 | — | — | — | 68.3 | |
| SFRonBackbone=PreResNet-110, Forgetting Ratio=20%2025.05 | 14.7 | 89.08 | 85.65 | 69.28 | 16.99 | 13.87 | — | |
| RTBackbone=PreResNet-110, Forgetting Ratio=20%2025.05 | 11.17 | 99.97 | 88.92 | 19.03 | 0 | 93.75 | — | |
| MCUβBackbone=PreResNet-110, Forgetting Ratio=20%2025.05 | 8.04 | 97.9 | 88.68 | 13.42 | 2.78 | 6.92 | — | |
| MCUBackbone=PreResNet-110, Forgetting Ratio=20%2025.05 | 7.21 | 98.2 | 88.23 | 13.64 | 2.95 | 6.88 | — | |
| NegGrad+Backbone=PreResNet-110, Forgetting Ratio=20%2025.05 | 5.22 | 98.51 | 89.32 | 10.03 | 4.2 | 2.96 | — | |
| SalUnBackbone=PreResNet-110, Forgetting Ratio=20%2025.05 | 3.87 | 98.76 | 89.95 | 24.94 | 3.86 | 7.96 | — | |
| NegTVBackbone=PreResNet-110, Forgetting Ratio=20%2025.05 | 3.33 | 98.27 | 86.86 | 6.83 | 5.95 | 1.28 | — | |
| RLBackbone=PreResNet-110, Forgetting Ratio=20%2025.05 | 3.24 | 99.34 | 90.24 | 23.64 | 3.62 | 7.83 | — | |
| FTBackbone=PreResNet-110, Forgetting Ratio=20%2025.05 | 0.34 | 99.94 | 90.89 | 3.43 | 7.11 | 4.72 | — | |
| GABackbone=PreResNet-110, Forgetting Ratio=20%2025.05 | 0.03 | 99.98 | 90.86 | 0.8 | 7.83 | 0.65 | — |