Model Distinguishability on CIFAR-10 (test)
0.253Directional Sharpness RatioResNet-BD
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| ResNet-BDBackdoor Type=BadNets, Backbone=ResNet-18, Poison Rate=1%2026.06 | 0.253 | 0.972 | 100 | |
| NoiseNoise Level=1%2026.06 | 0.122 | 0.676 | 99.5 | |
| NoiseNoise Level=2.5%2026.06 | 0.086 | 0.833 | 99.4 | |
| NoiseNoise Level=7.5%2026.06 | 0.081 | 0.418 | 99.8 | |
| NoiseNoise Level=5%2026.06 | 0.079 | 0.546 | 99.8 | |
| SAPAExperiment=Sharpness-aware poisoning attack, Poison Rate=1%, Perturbation budget=16/255, Sharpness radius=0.052026.06 | 0.016 | 0.217 | 100 |