Assessment of Defense Methodologies in DDC-AT Using Adversarial Training Techniques

[YanghuiMo]

B.3. Evaluating Defense Methods Table 6: Comparing the "Robust" PSPNet from Xu et al., 2021 against white-box adversarial attacks.

For the DDC-AT, did you conduct adversarial training using SegPGD or CosPGD attack samples on this DDCAT model by yourself?

Or did you simply use the pretrained model from the "Robust" PSPNet by Xu et al., 2021, which was trained only with BIM attack samples, and then test its defensive effectiveness against SegPGD and CosPGD attacks, as shown in the results in Table 6?

[shashankskagnihotri]

Hi, in Table 6 we imply use the pretrained model from the "Robust" PSPNet by Xu et al., 2021, which was trained only with BIM attack samples, and then test its defensive effectiveness against SegPGD and CosPGD attacks.

We did not train DDC-AT using SegPGD or CosPGD attack samples.