Image Classification on CIFAR-10 down-sampled to 32x32 (test)
96.6Median AccuracyAll layers
Evaluation results
| Method | Links | ||
|---|---|---|---|
| All layersepsilon=8, Fine-tuning algorithm=DP-SGD, Backbone=28-10 Wide-ResNet, Pre-training=ImageNet2022.04 | 96.6 | 0.08 | |
| All layersepsilon=4, Fine-tuning algorithm=DP-SGD, Backbone=28-10 Wide-ResNet, Pre-training=ImageNet2022.04 | 96.1 | 0.06 | |
| All layersepsilon=2, Fine-tuning algorithm=DP-SGD, Backbone=28-10 Wide-ResNet, Pre-training=ImageNet2022.04 | 95.4 | 0.15 | |
| Yu et al. (2021b)epsilon=2, Backbone=28-10 Wide-ResNet, Pre-training=ImageNet2022.04 | 94.8 | — | |
| All layersepsilon=1, Fine-tuning algorithm=DP-SGD, Backbone=28-10 Wide-ResNet, Pre-training=ImageNet2022.04 | 94.8 | 0.08 | |
| Yu et al. (2021b)epsilon=1, Backbone=28-10 Wide-ResNet, Pre-training=ImageNet2022.04 | 94.3 | — | |
| Classifier layerepsilon=8, Fine-tuning algorithm=DP-SGD, Backbone=28-10 Wide-ResNet, Pre-training=ImageNet2022.04 | 94.2 | 0.07 | |
| Classifier layerepsilon=4, Fine-tuning algorithm=DP-SGD, Backbone=28-10 Wide-ResNet, Pre-training=ImageNet2022.04 | 94 | 0.08 | |
| Classifier layerepsilon=2, Fine-tuning algorithm=DP-SGD, Backbone=28-10 Wide-ResNet, Pre-training=ImageNet2022.04 | 93.6 | 0.05 | |
| Classifier layerepsilon=1, Fine-tuning algorithm=DP-SGD, Backbone=28-10 Wide-ResNet, Pre-training=ImageNet2022.04 | 93.1 | 0.03 | |
| Tramèr and Boneh (2021)epsilon=2, Backbone=28-10 Wide-ResNet, Pre-training=ImageNet2022.04 | 92.7 | — |