Machine Unlearning on CIFAR-100 (test)
92.75Retain AccBase
Evaluation results
| Method | Links | ||||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| BaseClass=Baby, Backbone=ViT2026.06 | 92.75 | — | 87.5 | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| BaseClass=Sea, Backbone=ViT2026.06 | 92.67 | — | 87.51 | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| BaseClass=Rocket, Backbone=ViT2026.06 | 92.66 | — | 93.7 | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| BaseClass=Mountain, Backbone=ViT2026.06 | 92.64 | — | 95.31 | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| BaseClass=Lamp, Backbone=ViT2026.06 | 92.61 | — | 97.66 | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| BaseClass=Apple, Backbone=ViT2026.06 | 92.6 | — | 98.44 | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| BaseClass=Mushroom, Backbone=ViT2026.06 | 92.59 | — | 99.22 | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| MaGAmodel=ViT, class=RKT2026.02 | 92.34 | — | 0 | — | 0 | — | — | — | — | — | — | — | — | — | — | — | |
| MaGAmodel=ViT, class=MR2026.02 | 92.33 | — | 0 | — | 0 | — | — | — | — | — | — | — | — | — | — | — | |
| baselinemodel=ViT, class=RKT2026.02 | 92.27 | — | 93.14 | — | 0.8488 | — | — | — | — | — | — | — | — | — | — | — | |
| baselinemodel=ViT, class=MR2026.02 | 92.2 | — | 98.44 | — | 0.9024 | — | — | — | — | — | — | — | — | — | — | — | |
| retrainmodel=ViT, class=MR2026.02 | 92.18 | — | 0 | — | 0.0086 | — | — | — | — | — | — | — | — | — | — | — | |
| retrainmodel=ViT, class=RKT2026.02 | 92.04 | — | 0 | — | 0.0629 | — | — | — | — | — | — | — | — | — | — | — | |
| SSDmodel=ViT, class=MR2026.02 | 91.78 | — | 0 | — | 0.0145 | — | — | — | — | — | — | — | — | — | — | — | |
| SSDmodel=ViT, class=RKT2026.02 | 91.39 | — | 0 | — | 0.0662 | — | — | — | — | — | — | — | — | — | — | — | |
| UNSIRmodel=ViT, class=RKT2026.02 | 90.83 | — | 24.57 | — | 0.1143 | — | — | — | — | — | — | — | — | — | — | — | |
| AMNCmodel=ViT, class=RKT2026.02 | 90.53 | — | 0 | — | 0.0106 | — | — | — | — | — | — | — | — | — | — | — | |
| UNSIRmodel=ViT, class=MR2026.02 | 89.95 | — | 56.25 | — | 0.0261 | — | — | — | — | — | — | — | — | — | — | — | |
| AMNCmodel=ViT, class=MR2026.02 | 89.95 | — | 0 | — | 0.0088 | — | — | — | — | — | — | — | — | — | — | — | |
| SSDClass=Apple, Backbone=ViT2026.06 | 87.93 | — | 0 | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| DFMUClass=Mushroom, Backbone=ViT2026.06 | 86.75 | — | 0 | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| DFMUClass=Baby, Backbone=ViT2026.06 | 85.89 | — | 0 | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| FTmodel=ViT, class=RKT2026.02 | 84.26 | — | 0 | — | 0.16 | — | — | — | — | — | — | — | — | — | — | — | |
| FTmodel=ViT, class=MR2026.02 | 84.15 | — | 0 | — | 0.0505 | — | — | — | — | — | — | — | — | — | — | — | |
| Retrain2026.03 | 83.92 | — | — | — | — | — | 37 | — | — | — | — | — | — | — | — | — | |
| DFMUClass=Apple, Backbone=ViT2026.06 | 83.71 | — | 0 | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| DFMUClass=Lamp, Backbone=ViT2026.06 | 82.95 | — | 0.78 | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| DFMUClass=Mountain, Backbone=ViT2026.06 | 80.75 | — | 0 | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| DFMUClass=Sea, Backbone=ViT2026.06 | 79.9 | — | 2.34 | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| DFMUClass=Rocket, Backbone=ViT2026.06 | 79.35 | — | 0 | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Two-Stage Machine Unlearning2026.03 | 78.17 | — | — | — | — | — | 2.33 | — | — | — | — | — | — | — | — | — | |
| baselinemodel=RN18, class=MR2026.02 | 76.38 | — | 82.03 | — | 0.9565 | — | — | — | — | — | — | — | — | — | — | — | |
| baselinemodel=RN18, class=RKT2026.02 | 76.3 | — | 82.81 | — | 0.9661 | — | — | — | — | — | — | — | — | — | — | — | |
| MaGAmodel=RN18, class=MR2026.02 | 76.25 | — | 0 | — | 0 | — | — | — | — | — | — | — | — | — | — | — | |
| retrainmodel=RN18, class=MR2026.02 | 76.23 | — | 0 | — | 0.0601 | — | — | — | — | — | — | — | — | — | — | — | |
| SSDmodel=RN18, class=MR2026.02 | 76.2 | — | 0 | — | 0.0025 | — | — | — | — | — | — | — | — | — | — | — | |
| retrainmodel=RN18, class=RKT2026.02 | 76.19 | — | 0 | — | 0.0806 | — | — | — | — | — | — | — | — | — | — | — | |
| SSDmodel=RN18, class=RKT2026.02 | 75.86 | — | 0 | — | 0.0066 | — | — | — | — | — | — | — | — | — | — | — | |
| MaGAmodel=RN18, class=RKT2026.02 | 75.75 | — | 0 | — | 0 | — | — | — | — | — | — | — | — | — | — | — | |
| UNSIRmodel=RN18, class=MR2026.02 | 74.26 | — | 8.07 | — | 0.0168 | — | — | — | — | — | — | — | — | — | — | — | |
| UNSIRmodel=RN18, class=RKT2026.02 | 73.83 | — | 41.15 | — | 0.0308 | — | — | — | — | — | — | — | — | — | — | — | |
| AMNCmodel=RN18, class=RKT2026.02 | 73.59 | — | 0 | — | 0.2862 | — | — | — | — | — | — | — | — | — | — | — | |
| AMNCmodel=RN18, class=MR2026.02 | 73.22 | — | 0 | — | 0.4606 | — | — | — | — | — | — | — | — | — | — | — | |
| FTmodel=RN18, class=RKT2026.02 | 65.43 | — | 0 | — | 0.1004 | — | — | — | — | — | — | — | — | — | — | — | |
| FTmodel=RN18, class=MR2026.02 | 63.9 | — | 0 | — | 0.1222 | — | — | — | — | — | — | — | — | — | — | — | |
| SSDClass=Mushroom, Backbone=ViT2026.06 | 41.11 | — | 0 | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| SSDClass=Mountain, Backbone=ViT2026.06 | 1.42 | — | 0 | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| SSDClass=Lamp, Backbone=ViT2026.06 | 1.29 | — | 0 | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| RetrainBackbone=ResNet-50, Unlearning Protocol=Full Retraining2024.06 | 0.9997 | 1 | 0.5204 | 0.7326 | 0.635 | 0 | — | — | — | — | — | — | — | — | — | — | |
| NegGrad+ RUM (high -> med -> low)Backbone=ResNet-50, Unlearning Protocol=RUM, Subset Ordering=high -> med -> low2024.06 | 0.997 | 0.929 | 0.499 | 0.7108 | 0.657 | 0.022 | — | — | — | — | — | — | — | — | — | — | |
| NegGrad+Backbone=ResNet-50, Unlearning Protocol=One-go2024.06 | 0.9964 | 0.861 | 0.6364 | 0.7097 | 0.477 | 0.159 | — | — | — | — | — | — | — | — | — | — | |
| Fine-tuneBackbone=ResNet-50, Unlearning Protocol=One-go2024.06 | 0.9958 | 0.734 | 0.764 | 0.7074 | 0.496 | 0.139 | — | — | — | — | — | — | — | — | — | — | |
| SSDClass=Baby, Backbone=ViT2026.06 | 0.99 | — | 0 | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| L1-sparse RUM (high -> med -> low)Backbone=ResNet-50, Unlearning Protocol=RUM, Subset Ordering=high -> med -> low2024.06 | 0.9873 | 0.908 | 0.5397 | 0.6696 | 0.591 | 0.044 | — | — | — | — | — | — | — | — | — | — | |
| NegGrad+ RUMFBackbone=ResNet-50, Unlearning Protocol=RUM, Subset Ordering=low -> med -> high2024.06 | 0.9841 | 0.921 | 0.5079 | 0.6882 | 0.576 | 0.059 | — | — | — | — | — | — | — | — | — | — | |
| NegGrad+ shuffleBackbone=ResNet-50, Unlearning Protocol=Sequential, Subset Selection=Random Shuffle2024.06 | 0.9799 | 0.613 | 0.8801 | 0.7087 | 0.218 | 0.417 | — | — | — | — | — | — | — | — | — | — | |
| L1-sparse RUMFBackbone=ResNet-50, Unlearning Protocol=RUM, Subset Ordering=low -> med -> high2024.06 | 0.9727 | 0.883 | 0.5244 | 0.6432 | 0.602 | 0.033 | — | — | — | — | — | — | — | — | — | — | |
| Fine-tune RUMFBackbone=ResNet-50, Unlearning Protocol=RUM, Subset Ordering=low -> med -> high2024.06 | 0.9671 | 0.784 | 0.6177 | 0.6309 | 0.542 | 0.093 | — | — | — | — | — | — | — | — | — | — | |
| L1-sparse shuffleBackbone=ResNet-50, Unlearning Protocol=Sequential, Subset Selection=Random Shuffle2024.06 | 0.9531 | 0.604 | 0.8211 | 0.6388 | 0.282 | 0.353 | — | — | — | — | — | — | — | — | — | — | |
| SalUnBackbone=ResNet-50, Unlearning Protocol=One-go2024.06 | 0.9421 | 0.545 | 0.8897 | 0.6633 | 0.259 | 0.372 | — | — | — | — | — | — | — | — | — | — | |
| Fine-tune shuffleBackbone=ResNet-50, Unlearning Protocol=Sequential, Subset Selection=Random Shuffle2024.06 | 0.9362 | 0.589 | 0.8169 | 0.6265 | 0.29 | 0.345 | — | — | — | — | — | — | — | — | — | — | |
| L1-sparseBackbone=ResNet-50, Unlearning Protocol=One-go2024.06 | 0.9337 | 0.824 | 0.5408 | 0.6329 | 0.546 | 0.089 | — | — | — | — | — | — | — | — | — | — | |
| SalUn RUMFBackbone=ResNet-50, Unlearning Protocol=RUM, Subset Ordering=low -> med -> high2024.06 | 0.7972 | 0.614 | 0.5849 | 0.5561 | 0.454 | 0.181 | — | — | — | — | — | — | — | — | — | — | |
| SalUn shuffleBackbone=ResNet-50, Unlearning Protocol=Sequential, Subset Selection=Random Shuffle2024.06 | 0.7548 | 0.538 | 0.6339 | 0.5371 | 0.398 | 0.237 | — | — | — | — | — | — | — | — | — | — | |
| SSDClass=Sea, Backbone=ViT2026.06 | 0.74 | — | 0 | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| SSDClass=Rocket, Backbone=ViT2026.06 | 0.72 | — | 0 | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| AMUN2026.05 | — | — | — | 72.7 | — | — | — | — | — | — | — | — | — | — | — | — | |
| ESCBackbone=ViT, KD Data Source=Unseen data (test data)2025.04 | — | — | — | — | — | — | 0.2 | 95.61 | 0 | 90.67 | 97.66 | 95.11 | — | — | — | — | |
| ESC-TBackbone=ViT, KD Data Source=Unseen data (test data)2025.04 | — | — | — | — | — | — | 0.98 | 98.01 | 0.5 | 93.09 | 98.51 | 96.19 | — | — | — | — | |
| FTBackbone=ViT-B2026.03 | — | — | — | — | — | — | 62.33 | — | — | — | — | — | — | — | 77.83 | 83.89 | |
| FT2026.05 | — | — | — | 68 | — | — | — | — | — | — | — | — | — | — | — | — | |
| GABackbone=ViT-B2026.03 | — | — | — | — | — | — | 0 | — | — | — | — | — | — | — | 15.17 | 84.42 | |
| GA2026.05 | — | — | — | 67.8 | — | — | — | — | — | — | — | — | — | — | — | — | |
| IU2026.05 | — | — | — | 72.2 | — | — | — | — | — | — | — | — | — | — | — | — | |
| ℓ1-sparseBackbone=ViT-B2026.03 | — | — | — | — | — | — | 59.33 | — | — | — | — | — | — | — | 85.17 | 89.33 | |
| LTD2026.05 | — | — | — | 71 | — | — | — | — | — | — | — | — | — | — | — | — | |
| NegGrad (SAM, SAM)Pre-training Optimizer (A)=SAM, Unlearning Optimizer (U)=SAM, Forget set difficulty (F)=high, Backbone=ResNet50, Unlearning algorithm=NegGrad2025.06 | — | — | — | — | — | — | — | — | — | — | — | — | 0.095 | 0.3723 | — | — | |
| NegGrad (SAM, SAM)Pre-training Optimizer (A)=SAM, Unlearning Optimizer (U)=SAM, Forget set difficulty (F)=mid, Backbone=ResNet50, Unlearning algorithm=NegGrad2025.06 | — | — | — | — | — | — | — | — | — | — | — | — | 0.0913 | 0.3397 | — | — | |
| NegGrad (SAM, SAM)Pre-training Optimizer (A)=SAM, Unlearning Optimizer (U)=SAM, Forget set difficulty (F)=low, Backbone=ResNet50, Unlearning algorithm=NegGrad2025.06 | — | — | — | — | — | — | — | — | — | — | — | — | 0.1008 | 0.4901 | — | — | |
| NegGrad (SAM, SGD)Pre-training Optimizer (A)=SAM, Unlearning Optimizer (U)=SGD, Forget set difficulty (F)=high, Backbone=ResNet50, Unlearning algorithm=NegGrad2025.06 | — | — | — | — | — | — | — | — | — | — | — | — | 0.126 | 0.5212 | — | — | |
| NegGrad (SAM, SGD)Pre-training Optimizer (A)=SAM, Unlearning Optimizer (U)=SGD, Forget set difficulty (F)=mid, Backbone=ResNet50, Unlearning algorithm=NegGrad2025.06 | — | — | — | — | — | — | — | — | — | — | — | — | 0.1246 | 0.4959 | — | — | |
| NegGrad (SAM, SGD)Pre-training Optimizer (A)=SAM, Unlearning Optimizer (U)=SGD, Forget set difficulty (F)=low, Backbone=ResNet50, Unlearning algorithm=NegGrad2025.06 | — | — | — | — | — | — | — | — | — | — | — | — | 0.1248 | 0.5212 | — | — | |
| NegGrad (SGD, SAM)Pre-training Optimizer (A)=SGD, Unlearning Optimizer (U)=SAM, Forget set difficulty (F)=high, Backbone=ResNet50, Unlearning algorithm=NegGrad2025.06 | — | — | — | — | — | — | — | — | — | — | — | — | 0.1076 | 0.4901 | — | — | |
| NegGrad (SGD, SAM)Pre-training Optimizer (A)=SGD, Unlearning Optimizer (U)=SAM, Forget set difficulty (F)=mid, Backbone=ResNet50, Unlearning algorithm=NegGrad2025.06 | — | — | — | — | — | — | — | — | — | — | — | — | 0.0983 | 0.3723 | — | — | |
| NegGrad (SGD, SAM)Pre-training Optimizer (A)=SGD, Unlearning Optimizer (U)=SAM, Forget set difficulty (F)=low, Backbone=ResNet50, Unlearning algorithm=NegGrad2025.06 | — | — | — | — | — | — | — | — | — | — | — | — | 0.1093 | 0.4901 | — | — | |
| NegGrad (SGD, SGD)Pre-training Optimizer (A)=SGD, Unlearning Optimizer (U)=SGD, Forget set difficulty (F)=high, Backbone=ResNet50, Unlearning algorithm=NegGrad2025.06 | — | — | — | — | — | — | — | — | — | — | — | — | 0.1293 | 0.5212 | — | — | |
| NegGrad (SGD, SGD)Pre-training Optimizer (A)=SGD, Unlearning Optimizer (U)=SGD, Forget set difficulty (F)=mid, Backbone=ResNet50, Unlearning algorithm=NegGrad2025.06 | — | — | — | — | — | — | — | — | — | — | — | — | 0.1259 | 0.4901 | — | — | |
| NegGrad (SGD, SGD)Pre-training Optimizer (A)=SGD, Unlearning Optimizer (U)=SGD, Forget set difficulty (F)=low, Backbone=ResNet50, Unlearning algorithm=NegGrad2025.06 | — | — | — | — | — | — | — | — | — | — | — | — | 0.1271 | 0.4959 | — | — | |
| OriginalBackbone=ViT, KD Data Source=Unseen data (test data)2025.04 | — | — | — | — | — | — | 98.42 | 98.1 | 93.3 | 93.12 | 3.11 | 12.5 | — | — | — | — | |
| OriginalBackbone=ViT-B2026.03 | — | — | — | — | — | — | 95 | — | — | — | — | — | — | — | 91.75 | 98 | |
| Our methodBackbone=ViT-B2026.03 | — | — | — | — | — | — | 0.67 | — | — | — | — | — | — | — | 89.5 | 93.22 | |
| Retrain2026.05 | — | — | — | 71.7 | — | — | — | — | — | — | — | — | — | — | — | — | |
| RL2026.05 | — | — | — | 67.5 | — | — | — | — | — | — | — | — | — | — | — | — | |
| SalUn2026.05 | — | — | — | 67.5 | — | — | — | — | — | — | — | — | — | — | — | — | |
| SCRUBBackbone=ViT-B2026.03 | — | — | — | — | — | — | 3 | — | — | — | — | — | — | — | 89.58 | 94.69 |