Adversarial Detection on CIFAR-10 (Detection Rates)
100PGD Detection Rate (linf=2/255)FEATURE SQUEEZING
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| FEATURE SQUEEZINGBackbone=ResNet-182025.10 | 100 | — | 5 | 14 | |
| KOALA+Backbone=ResNet-18, Embedding Objective=KOALA+2025.10 | 71.26 | 87 | 73.54 | 68.78 | |
| KL+L0+COSINEBackbone=ResNet-18, Embedding Objective=KL+L0+COSINE2025.10 | 65.28 | 79.66 | 67.46 | 65.8 | |
| COSINE+L0Backbone=ResNet-18, Embedding Objective=COSINE+L02025.10 | 64.55 | 81.02 | 67.05 | 63.69 | |
| KL+L0Backbone=ResNet-18, Embedding Objective=KL+L02025.10 | 64.28 | 85.04 | 66.23 | 64.98 | |
| COSINE+KLBackbone=ResNet-18, Embedding Objective=COSINE+KL2025.10 | 52.54 | 82.94 | 54.62 | 54.66 | |
| MAGNETBackbone=ResNet-182025.10 | 0 | — | 55 | 72 |