Image Classification on Contrastive Learning Pre-training CorruptEncoder+ Attack (test)
62Backdoor AccuracyContrastiveCrop
Evaluation results
| Method | Links | ||
|---|---|---|---|
| ContrastiveCropCropping Strategy=Semantic-aware localization mechanism2022.11 | 62 | 98.5 | |
| No Defense (MoCo-v2)Data Augmentation Settings=Default MoCo-v2 operations2022.11 | 61.7 | 97.8 | |
| CompRess DistillationCompression Ratio=20%, Extra clean pre-training dataset=true2022.11 | 58.6 | 1.1 | |
| Localized CroppingCropping Strategy=Localized (delta = 0.2)2022.11 | 56.1 | 0.8 | |
| CompRess DistillationCompression Ratio=5%, Extra clean pre-training dataset=true2022.11 | 49.9 | 0.9 | |
| No Other Data AugsData Augmentation Settings=Only random cropping used2022.11 | 44.2 | 75.7 | |
| No Random CroppingCropping Strategy=None2022.11 | 31.9 | 1.5 |