EDLNets replace convolutions with unrolled elastic dictionary-learning layers and report improved AutoAttack robustness, e.g., 59.07% AA (l-infinity 8/255) vs 53.16% for HAT on CIFAR-10 ResNet-18.
During the 100th to 150th epochs, the model experiences a catastrophic robust overfitting problem
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Boosting Adversarial Robustness and Generalization with Structural Prior
EDLNets replace convolutions with unrolled elastic dictionary-learning layers and report improved AutoAttack robustness, e.g., 59.07% AA (l-infinity 8/255) vs 53.16% for HAT on CIFAR-10 ResNet-18.