Machine Unlearning on CIFAR-100 50% deletion for class 68 (road)
99Utility Accuracy (UA)Inference Unlearning (IU)
Evaluation results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| Inference Unlearning (IU)RTE=27.4 ± 0.32026.05 | 99 | 99.8 | 72.1 | 12.5 | 6.7 | |
| Fine-Tuning (FT)RTE=156.0 ± 0.62026.05 | 94.4 | 94.6 | 68 | 8.6 | 8.8 | |
| AMUNRTE=171.1 ± 0.52026.05 | 92.8 | 94.3 | 72.8 | 11.1 | 7 | |
| Random Labels (RL)RTE=157.8 ± 0.62026.05 | 91.4 | 94.6 | 68 | 85.7 | 18.5 | |
| RetrainRTE=1557.8 ± 12.12026.05 | 86.1 | 99.8 | 72 | 26.3 | 0 | |
| SalUnRTE=347.4 ± 3.92026.05 | 84.2 | 93.2 | 67.4 | 77.8 | 16.2 | |
| Local Teacher Distillation (LTD)RTE=202.6 ± 3.42026.05 | 80.6 | 99.5 | 70.8 | 33.1 | 3.5 | |
| Gradient Ascent (GA)RTE=1.7 ± 0.02026.05 | 66.3 | 97 | 68.4 | 50.2 | 12.6 |