Backdoor Attack on SVHN
100Attack Success RateInkDrop
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| InkDropCondensation Method=IDM, Condensation Backbone=ConvNet2026.03 | 100 | 81.87 | — | — | |
| InkDropCondensation Method=DAM, Condensation Backbone=ConvNet2026.03 | 100 | 75.7 | — | — | |
| DoorpingCondensation Method=DAM, Condensation Backbone=ConvNet2026.03 | 100 | 72.09 | — | — | |
| RelaxCondensation Method=DM, Condensation Backbone=ConvNet2026.03 | 100 | 74.73 | — | — | |
| RelaxCondensation Method=DAM, Condensation Backbone=ConvNet2026.03 | 100 | 74.5 | — | — | |
| DoorpingCondensation Method=DM, Condensation Backbone=ConvNet2026.03 | 99.96 | 77.97 | — | — | |
| INACTIVEPre-training Dataset=STL10, Trigger=Invisible (Ours)2024.05 | 99.76 | — | 58.62 | — | |
| InkDropCondensation Method=DM, Condensation Backbone=ConvNet2026.03 | 99.76 | 77.52 | — | — | |
| RelaxCondensation Method=IDM, Condensation Backbone=ConvNet2026.03 | 99.2 | 83.36 | — | — | |
| INACTIVEPre-training Dataset=CIFAR10, Trigger=Invisible (Ours)2024.05 | 98.79 | — | 63.67 | — | |
| DRUPEPre-training Dataset=STL10, Trigger=Patch2024.05 | 96.68 | — | 75.64 | — | |
| DRUPEPre-training Dataset=CIFAR10, Trigger=Patch2024.05 | 96.23 | — | 76.02 | — | |
| BadEncoderPre-training Dataset=CIFAR10, Trigger=CTRL2024.05 | 40.91 | — | 66.33 | — | |
| BadEncoderPre-training Dataset=STL10, Trigger=Ins-Kelvin2024.05 | 38.03 | — | 56.67 | — | |
| BadEncoderPre-training Dataset=CIFAR10, Trigger=Ins-Xpro22024.05 | 30.91 | — | 68.95 | — | |
| BadEncoderPre-training Dataset=CIFAR10, Trigger=Ins-Kelvin2024.05 | 22.13 | — | 68.49 | — | |
| BadEncoderPre-training Dataset=STL10, Trigger=Ins-Xpro22024.05 | 18.68 | — | 58.42 | — | |
| BadEncoderPre-training Dataset=CIFAR10, Trigger=WaNet2024.05 | 17.99 | — | 54.79 | — | |
| BadEncoderPre-training Dataset=STL10, Trigger=WaNet2024.05 | 16.83 | — | 58.29 | — | |
| NaiveCondensation Method=IDM, Condensation Backbone=ConvNet2026.03 | 12.2 | 84.01 | — | — | |
| SimpleCondensation Method=IDM, Condensation Backbone=ConvNet2026.03 | 11.4 | 84.15 | — | — | |
| SimpleCondensation Method=DAM, Condensation Backbone=ConvNet2026.03 | 11.4 | 75.89 | — | — | |
| NaiveCondensation Method=DAM, Condensation Backbone=ConvNet2026.03 | 11.16 | 77.03 | — | — | |
| NaiveCondensation Method=DM, Condensation Backbone=ConvNet2026.03 | 11.08 | 79.85 | — | — | |
| SimpleCondensation Method=DM, Condensation Backbone=ConvNet2026.03 | 11.04 | 74.79 | — | — | |
| DoorpingCondensation Method=IDM, Condensation Backbone=ConvNet2026.03 | 6.08 | 83.92 | — | — | |
| BadEncoderPre-training Dataset=STL10, Trigger=CTRL2024.05 | 3.32 | — | 54.29 | — | |
| BADBONEModel=ResNet182026.05 | — | 66.19 | 92.44 | 96.16 | |
| BADBONEModel=ResNet502026.05 | — | 71.91 | 94.1 | 97.3 | |
| BADBONEModel=BiT-M-RN502026.05 | — | 71.95 | 94.12 | 97.8 | |
| Clean BaselineModel=ResNet182026.05 | — | 69.76 | 61.39 | 39.77 | |
| Clean BaselineModel=ResNet502026.05 | — | 76.15 | 57.91 | 42.51 | |
| Clean BaselineModel=BiT-M-RN502026.05 | — | 74.02 | 73.15 | 29.24 | |
| Clean ModelPre-training Dataset=STL10, Trigger=None2024.05 | — | 55.35 | — | — | |
| Clean ModelPre-training Dataset=CIFAR10, Trigger=None2024.05 | — | 61.52 | — | — |