ttbrunner / biased_boundary_attack Goto Github PK

Dear sir:
I am preparing a paper, I want to use your code to conduct experiments of both the original boundary attack and biased boundary attack. If I set USE_PERLIN_BIAS = False, USE_MASK_BIAS = False, USE_SURROGATE_BIAS = False in bench_settings.py.
Does this code produce the results exactly the same with the original boundary attack?

Hi,

Thank you for the interesting paper and releasing the code.

I wonder if you can provide the indices of 1000 sampled images you use in the paper for all those methods? I would like to reproduce the results and make some fair comparison for my modification.

Thanks in advance!

I found it is too slow when I attack a large network in CIFAR-10 dataset. Because my experiments need attacking 1000 images. It will take 1 month!
I found the reason is that you feed only 1 image to network every time. (in def _eval_sample)
Can this code modify to batch image feed version to boost speed?
We first collect a batch of candidate, stats = candidate_future.result() into a block-queue, and then batch feed to network?
What about this proposal? What do you suggest?

In line 231 in generate_boundary_sample of biased_boundary_attack.py, does "vdot" in "dot = np.vdot(sampling_dir, source_direction)" computes between two images of 3 channels? That's to say, if the shape of image is (224, 224, 3), the shape of sampling_dir here is also (224, 224, 3), and we get one numpy value after np.vdot. Is it right? Thanks.