EMNIST images mirrored and rotated

[hiddely]

Hello,

When trying to generate the adversarial dataset for the EMNIST+ARDIS setting I noticed that the honest (EMNIST) images and the ARDIS images have a different orientation. Specifically, the images from torchvision EMNIST dataset are 90 degrees rotated and mirrored (in generating_poisoned_DA.py), whereas the ARDIS-7's have a normal orientation. The images are then saved together into one attacker dataset. Therefore, it seems that you are training a model on 90deg+mirrored EMNIST while inserting an edge-case ARDIS backdoor that is not rotated/flipped.

Am I missing something here (are you correcting these EMNIST images later somewhere)?

Thanks,
Hidde

[ksreenivasan]

That is possible. I didn't explicitly try to visualize the images after loading them into pytorch. If anything, that is hurting our results since the 7's will stand out in the clean data. Can you check and see if that is indeed the case? If not, I will get to it soon :)

[ksreenivasan]

Thank you for raising the issue!