search

evtimovi / robust_physical_perturbations

Public release of code for Robust Physical-World Attacks on Deep Learning Visual Classification (Eykholt et al., CVPR 2018)
https://iotsecurity.eecs.umich.edu/#roadsigns
MIT License
102 stars 35 forks source link

alignment function #9

Open sli057 opened 4 years ago

sli057 commented 4 years ago

Hi, thanks for the awesome work. I look through the code and there seem to be no codes accounting for the alignment function shown in equation 3 in the paper. Alignment function is used to map transformations on the object to transformations on the perturbation (e.g. if the object is rotated, the perturbation is rotated as well.) Could you clarify this?

evtimovi commented 4 years ago

I believe this is what you need: https://github.com/evtimovi/robust_physical_perturbations/blob/master/imagenet-attack/attack_util.py#L198

Note that we did not use these transformations for the gtsrb and the lisa-cnn.

jinyugy21 commented 4 years ago

Hi, thanks for your wonderful work and the previous answers! But I still have a problem, that is, if the alignment function Ti described in equation 3 in the paper is not used in Lisa-CNN, how does the algorithm ensure that a specific perturbation can adapt to the scene with different distances and angles in Table 1? Thank you for any help you can offer.

zdaiot commented 3 years ago

Hi, thanks for your wonderful work and the previous answers! But I still have a problem, that is, if the alignment function Ti described in equation 3 in the paper is not used in Lisa-CNN, how does the algorithm ensure that a specific perturbation can adapt to the scene with different distances and angles in Table 1? Thank you for any help you can offer.

Hello, Have you solved this problem? I have the same doubts. @jinyugy21