Unimodal Backdoor Defense on ImageNet-1K 1.0 (test)
77.47Clean Accuracy (CA)Platonic Defense
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| Platonic DefenseEvaluation protocol=linear probing, Attack=Average (Backdoor attacks on unimodal encoders)2026.06 | 77.47 | 66.49 | 94 | 3.43 | |
| DecompEvaluation protocol=linear probing, Attack=Average (Backdoor attacks on unimodal encoders)2026.06 | 75.83 | 0.58 | 99 | 99.31 | |
| No DefenseEvaluation protocol=linear probing, Attack=Average (Backdoor attacks on unimodal encoders)2026.06 | 75.63 | 0.6 | — | 99.18 | |
| ZIPEvaluation protocol=linear probing, Attack=Average (Backdoor attacks on unimodal encoders)2026.06 | 65.68 | 17.35 | — | 71.49 | |
| DeDeEvaluation protocol=linear probing, Attack=Average (Backdoor attacks on unimodal encoders)2026.06 | 29.56 | 0.36 | 78 | 33.82 | |
| BeatrixEvaluation protocol=linear probing, Attack=Average (Backdoor attacks on unimodal encoders)2026.06 | — | — | 97 | — | |
| DetectCLIPEvaluation protocol=linear probing, Attack=Average (Backdoor attacks on unimodal encoders)2026.06 | — | — | 31 | — |