Preventing Catastrophic Overfitting in Fast Adversarial Training

Wang Z, Wang H, Tian C, et al. Preventing Catastrophic Overfitting in Fast Adversarial Training: A Bi-level Optimization Perspective[C]//European Conference on Computer Vision. Springer, Cham, 2025: 144-160.

---

Introduction:

• Adversarial training (AT)

• adversarial examples (AEs)

• fast AT: a single-step attack strategy

---

Problems:

• catastrophic overfitting

• on complex tasks or with large-parameter models

• FGSM-PCO

---

Methods:

• an adaptive mechanism

• loss function

FGSM:

Experiments:

• CIFAR10, CIFAR100 and Tiny-ImageNet

• ResNet18, PreActResNet18, WideResNet34-10

• FGSMRS, FGSM-GA, FreeAT, TRADES, and FGSM-PGI

---