Machine Unlearning on CIFAR-10 (test)
95.41Test AccSCRUB
Evaluation results
| Method | Links | |||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| SCRUBBackbone=ResNet-18, Number of trials=3, Uses Retain set (Dr)=true, Run Time Efficiency (s)=160.392025.04 | 95.41 | — | — | — | — | 4.69 | — | — | — | — | — | — | — | |
| BadTBackbone=ResNet-18, Number of trials=3, Run Time Efficiency (s)=182.782025.04 | 95.27 | — | — | — | — | 4.69 | — | — | — | — | — | — | — | |
| RetrainBackbone=ResNet-18, Number of trials=3, Run Time Efficiency (s)=2681.722025.04 | 95.23 | — | — | — | — | — | — | — | — | — | — | — | — | |
| SalUnBackbone=ResNet-18, Number of trials=3, Run Time Efficiency (s)=150.512025.04 | 95.18 | — | — | — | — | 3.88 | — | — | — | — | — | — | — | |
| FinetuneBackbone=ResNet-18, Number of trials=3, Run Time Efficiency (s)=132.022025.04 | 95.08 | — | — | — | — | 4.38 | — | — | — | — | — | — | — | |
| ESCBackbone=ResNet-18, Number of trials=3, Uses Retain set (Dr)=false, Run Time Efficiency (s)=2.242025.04 | 95.07 | — | — | — | — | 1.45 | — | — | — | — | — | — | — | |
| SSDBackbone=ResNet-18, Comp. Cost (PFLOPs)=0.062025.09 | 94.8 | — | — | — | — | 3.83 | — | — | — | — | — | — | 0.13 | |
| FTBackbone=ResNet-18, Comp. Cost (PFLOPs)=6.322025.09 | 94.7 | — | — | — | — | 0.65 | — | — | — | — | — | — | 0.03 | |
| RetrainBackbone=ResNet-18, Comp. Cost (PFLOPs)=27.372025.09 | 94.67 | — | — | — | — | 0 | — | — | — | — | — | — | 0 | |
| NoTBackbone=ResNet-18, Comp. Cost (PFLOPs)=7.522025.09 | 94.65 | — | — | — | — | 0.3 | — | — | — | — | — | — | 0.02 | |
| SCRUBBackbone=ResNet-18, Comp. Cost (PFLOPs)=8.932025.09 | 94.61 | — | — | — | — | 0.62 | — | — | — | — | — | — | 0.06 | |
| CoUnBackbone=ResNet-18, Comp. Cost (PFLOPs)=8.022025.09 | 94.57 | — | — | — | — | 0.25 | — | — | — | — | — | — | 0.1 | |
| NegGrad+Backbone=ResNet-18, Comp. Cost (PFLOPs)=6.022025.09 | 94.32 | — | — | — | — | 0.58 | — | — | — | — | — | — | 0.35 | |
| CUBackbone=ResNet-18, Comp. Cost (PFLOPs)=3.362025.09 | 94.18 | — | — | — | — | 0.61 | — | — | — | — | — | — | 0.49 | |
| BadTBackbone=ResNet-18, Comp. Cost (PFLOPs)=9.892025.09 | 94.05 | — | — | — | — | 1.58 | — | — | — | — | — | — | 0.62 | |
| SalUnBackbone=ResNet-18, Comp. Cost (PFLOPs)=8.662025.09 | 93.84 | — | — | — | — | 0.59 | — | — | — | — | — | — | 0.83 | |
| ℓ1-sparseBackbone=ResNet-18, Comp. Cost (PFLOPs)=6.922025.09 | 93.81 | — | — | — | — | 0.38 | — | — | — | — | — | — | 0.86 | |
| RLBackbone=ResNet-18, Number of trials=3, Run Time Efficiency (s)=2.572025.04 | 93.66 | — | — | — | — | 12.33 | — | — | — | — | — | — | — | |
| ESC-TBackbone=ResNet-18, Number of trials=3, Uses Retain set (Dr)=false, Run Time Efficiency (s)=29.472025.04 | 92.73 | — | — | — | — | 2.23 | — | — | — | — | — | — | — | |
| NGBackbone=ResNet-18, Number of trials=3, Run Time Efficiency (s)=26.792025.04 | 89.67 | — | — | — | — | 3.65 | — | — | — | — | — | — | — | |
| RetrainBackbone=ResNet-182024.06 | 83.917 | 1 | 64.156 | 100 | 0.549 | 0 | — | — | — | — | — | — | — | |
| NegGrad+ shuffleBackbone=ResNet-182024.06 | 83.7 | 0.632 | 99.6 | 99.929 | 0.029 | 0.52 | — | — | — | — | — | — | — | |
| Fine-tune RUM (high → med → low)Backbone=ResNet-18, Strategy=Reverse Order2024.06 | 83.587 | 0.942 | 68.222 | 99.515 | 0.478 | 0.071 | — | — | — | — | — | — | — | |
| Fine-tune RUMFBackbone=ResNet-18, Strategy=RUM (low -> med -> high)2024.06 | 83.23 | 0.937 | 69.133 | 99.664 | 0.45 | 0.099 | — | — | — | — | — | — | — | |
| nothing → Fine-tune → SalUnBackbone=ResNet-18, Strategy=Algorithm Refinement2024.06 | 83.007 | 0.965 | 66.011 | 99.205 | 0.515 | 0.034 | — | — | — | — | — | — | — | |
| Fine-tune shuffleBackbone=ResNet-182024.06 | 82.037 | 0.712 | 88.389 | 96.578 | 0.451 | 0.098 | — | — | — | — | — | — | — | |
| L1-sparse shuffleBackbone=ResNet-182024.06 | 81.047 | 0.716 | 88.244 | 97.152 | 0.211 | 0.338 | — | — | — | — | — | — | — | |
| L1-sparse RUMFBackbone=ResNet-18, Strategy=RUM (low -> med -> high)2024.06 | 81.047 | 0.9 | 69.967 | 98.458 | 0.477 | 0.072 | — | — | — | — | — | — | — | |
| SalUnBackbone=ResNet-182024.06 | 80.37 | 0.731 | 40.056 | 99.796 | 0.923 | 0.374 | — | — | — | — | — | — | — | |
| L1-sparseBackbone=ResNet-182024.06 | 79.373 | 0.794 | 79.3 | 98.187 | 0.374 | 0.175 | — | — | — | — | — | — | — | |
| Fine-tuneBackbone=ResNet-182024.06 | 79.277 | 0.849 | 73.067 | 97.923 | 0.429 | 0.12 | — | — | — | — | — | — | — | |
| SalUn RUMFBackbone=ResNet-18, Strategy=RUM (low -> med -> high)2024.06 | 77.857 | 0.887 | 62.878 | 96.457 | 0.518 | 0.031 | — | — | — | — | — | — | — | |
| NegGrad+Backbone=ResNet-182024.06 | 77.337 | 0.802 | 76.867 | 98.415 | 0.319 | 0.23 | — | — | — | — | — | — | — | |
| SalUn shuffleBackbone=ResNet-182024.06 | 77.27 | 0.727 | 81.889 | 94.693 | 0.315 | 0.234 | — | — | — | — | — | — | — | |
| NegGrad+ RUMFBackbone=ResNet-18, Strategy=RUM (low -> med -> high)2024.06 | 77.063 | 0.879 | 61.744 | 96.655 | 0.415 | 0.134 | — | — | — | — | — | — | — | |
| ESCBackbone=ViT, KD Data Source=Unseen data (test data)2025.04 | — | — | — | — | — | — | 1.92 | 99.56 | 2.3 | 98.36 | 98.81 | 98.03 | — | |
| ESC-TBackbone=ViT, KD Data Source=Unseen data (test data)2025.04 | — | — | — | — | — | — | 0.86 | 99.92 | 0.5 | 99.06 | 99.53 | 99.28 | — | |
| OriginalBackbone=ViT, KD Data Source=Unseen data (test data)2025.04 | — | — | — | — | — | — | 99.96 | 99.91 | 99.2 | 99.02 | 0.08 | 1.59 | — |