Open ClickAB opened 11 months ago
Thanks for your attention, our approach can identify adversarial data within an unknown dataset, for instance, CIFAR-10, by employing a pre-trained diffusion model initially trained on a known dataset like ImageNet. In contrast to the typical procedure, it's necessary to resize all images to dimensions compatible with the input requirements of the pre-trained diffusion model, e.g., 224x224x3. Additionally, you'd better employ the EPSs calculated on a set of natural samples (with size=500) from your preferred dataset as the reference.
Thank you very much for answering my question! I generally understand what you mean. I need some similar clean samples for comparison, comparing their EPS with the EPS of the samples to be detected. For example, if I want to detect adversarial samples on an aerial image dataset named A, I need to first find a certain number of clean samples (with size around 500 ) of aerial images as clean datasets named B to calculate their EPS-B, and then input the aerial image dataset A to calculate their EPS-A, and calculate MMD between EPS-A and EPS-B,is this right?
Yes, you are right!
I input an unknown dataset, and I don't know which ones are clean samples and which ones are adversarial samples. How can I calculate the EPS of the unknown sample and compare it with the EPS of the clean sample through MMD