Attacking Google Cloud Vision API

[Eric-mingjie]

I am trying to attack the google cloud vision api using your code and i was checking on the video you posted about the attacking process. I have some questions about it:

There seems to be some images where the label dog are not among the predicted labels, for examples https://youtu.be/1h9bU7WBTUg?t=43. But I think the partial information attack depends on the target class staying in the predicted labels. So there seems to be contradiction here? Is the video showing the generated images at every iteration?

Also, what is the k used in the attacking the google cloud vision api?

[andrewilyas]

Hi, the only reason the dog doesn't appear in the video is because the number of labels exceeds the size of the scroll bar (so the dog label is just cut off). The dog label is actually maintained throughout classification.

[Eric-mingjie]

Thanks for the timely reply. But what about 9.png and 19.png? Clearly for these two images the label dog does not appear in the predicted labels.