papermsucode / advhat Goto Github PK

When I used your code to experiment with my own photos, I printed out the stickers and pasted them on the hat. After taking photos again, the final similarity was 0.6, which was still very high. Is it just because of the problem with my printer? Or something else?

Can you provide tf models of r50 r34 and mobileface? I try to change from mxnet models in mobile-ZOO you provide to tf models. But I can't gain tf-models having same nodes and names of nodes with yours.

how can i change the placement of the sticker ?

l try to use attack.py to prepare face attack. l followed you steps and when trying to use anchor_face .it shows me that File "attack.py", line 24, in prep
return np.transpose(im, [2, 0, 1]).reshape((1, 3, 112, 112)) * 2 - 1#.reshape( [1,3,120, 120])*2-1
ValueError: cannot reshape array of size 12544 into shape (1,3,112,112). Is there something l need to set

Hello, is it possible to provide proper documentation or guideline to test the whole project?
Suppose, what are the pre-requisites and steps needed to follow?
Thanks.

I had a problem generating the face mask. Can you give a list of what libraries are required for pip?

Line 143:

anch_im = rescale(io.imread(args.anchor_face)/255.,112./600.,order=5,multichannel=True)

Line 149:

anch_im = rescale(io.imread(args.anchor_face)/255.,112./600.,order=5,multichannel=True)

I'm wondering if this method can be applied to do physical target attack since this is more dangerous. I have tried to optimize the adv patch by force the face to target face, but the similarity never goes up to 0.55. Do you have any suggestions?

In the part of Preparation of your own Centroids,how many images of your own that you use to train?

Thanks a lot.

as I know
assume

arc length

It is the same as the formula in the paper
so we get

As described in this figure:

but the equetion in the paper is:

in which the order of x and x' is reversed
So how can I get the formula in the paper？

Hello, thanks for your work.

Where is your pre-trained model from? Is it converted from mxnet model or you trained it from scratch using TensorFlow?

Thanks.

Hi, thank you for your contribution. When I trying to reproduce your work, the tf model can't be loaded with error "InvalidArgumentError (see above for traceback): NodeDef mentions attr 'dilations' not in Op<name=Conv2D; signature=input:T, filter:T -> output:T; attr=T:type,allowed=[DT_HALF, DT_FLOAT]; attr=strides:list(int); attr=use_cudnn_on_gpu:bool,default=true; attr=padding:string,allowed=["SAME", "VALID"]; attr=data_format:string,default="NHWC",allowed=["NHWC", "NCHW"]>; NodeDef: conv0 = Conv2D[T=DT_FLOAT, data_format="NCHW", dilations=[1, 1, 1, 1], padding="SAME", strides=[1, 1, 1, 1], use_cudnn_on_gpu=true, _device="/job:localhost/replica:0/task:0/device:GPU:0"](_arg_image_input_0_0/_77, W_conv0). (Check whether your GraphDef-interpreting binary is up to date with your GraphDef-generating binary.)." I think this is caused by the tf version. So, what's your tensorflow version? My version is tf 1.4.0.

We found that the code for the projector part difficult to understand, there seems a lot cumsum and index operations. We might not be familiar for parabolic transformation in the 3d space. Could you provide more references for understand?

when i run demo.py, error happened: Generic conv implementation only supports NHWC tensor format for now, node conv0 , in line 90 , sess.run(embedding,feed_dict={image_input:batch,keep_prob:1.0,is_train:False}) .
when i hide this line (90) , it can work well, show rectangle face correctly.

what' wrong ? my tensorflow verson is 1.14.0 .,Python 3.6.
thanks a lot.

Hi, I'm correctly working on target attack with your method. I found it always fails when trying to
attack other face recognition models, even though the logo is large. For example, I got 0.97 cos similarity on your model, but when applied to IR50 model in pytorch, the similarity only reached 0.42. So I'm wondering if your method only works for your model, because I did't manage to get a good mask on pytorch IR50 model even for untargeted attack. I'm wondering have you tried your method on other face recognition model.