PGD Adversarial Training
A classical adversarial training method that uses projected gradient descent (PGD) to generate adversarial examples. Proposed by Madry et al., it is one of the benchmark methods for evaluating adversarial robustness.
Timeline (last 90 days)
Adversarial training (AT) improves model robustness by generating adversarial examples during training and including them in training, but it incurs enormous memory overhead and involves a trade-off between clean accuracy and robust accuracy.
The PGD adversarial training proposed by Madry et al. is a classic adversarial training implementation; it requires additional computation of adversarial examples at each step, making training far more costly than standard training.
Compared with PGD-based adversarial training, this method achieves a relative improvement in robust accuracy of at least 20.58% under PGD attacks and a relative improvement of 10.20% in classification accuracy on clean samples.
All Facts (3)
Adversarial training (AT) improves model robustness by generating adversarial examples during training and including them in training, but it incurs enormous memory overhead and involves a trade-off between clean accuracy and robust accuracy.
50%UnverifiedThe PGD adversarial training proposed by Madry et al. is a classic adversarial training implementation; it requires additional computation of adversarial examples at each step, making training far more costly than standard training.
50%UnverifiedCompared with PGD-based adversarial training, this method achieves a relative improvement in robust accuracy of at least 20.58% under PGD attacks and a relative improvement of 10.20% in classification accuracy on clean samples.
50%