Necessary steps to run Scheme 2?

[hanglearning]

Dear VITA-Group,

I have read your awesome paper You are caught stealing my winning lottery ticket! Making a lottery ticket claim its ownership and I am interested in tuning the idea of QR Code embedding and verification, and leverage it into our own project. Currently, I wish to debug the entire code process of Scheme 2 to understand the nitty gritties.

I could successfully run the code main_imp_new.py, but I have difficulties running Scheme 2. The documentation only has this one line of command

Scheme 2

python embed_res20s_cifar10.py

The code couldn't run by just this command. I have also tried to run main_imp_new.py to create a model for embed_res20s_cifar10.py to load, but then there would be more exceptions.

May I ask how I can execute Scheme 2?

Thanks for your help!

[xxchenxx]

Thank you for your interest; I will update a more detailed steps to help run the codes. Maybe later this week/weekend and thank you for your patience.

[hanglearning]

Thank you so much for your prompt response! Your help is much appreciated and will help big on our project! 👍

[xxchenxx]

btw, could you please share me the errorlog? Thanks!

[xxchenxx]

Hi Hang,

I have ran with our previous codes and I got the following outputs:

0.19369237907088122
layer2.1.conv1.weight_mask
0.5267538644470868
dict_keys(['layer1.0.conv1.weight_mask', 'layer1.0.conv2.weight_mask', 'layer1.1.conv1.weight_mask', 'layer1.1.conv2.weight_mask', 'layer1.2.conv1.weight_mask', 'layer1.2.conv2.weight_mask', 'layer2.0.conv1.weight_mask', 'layer2.0.conv2.weight_mask', 'layer2.1.conv1.weight_mask', 'layer2.1.conv2.weight_mask', 'layer2.2.conv1.weight_mask', 'layer2.2.conv2.weight_mask', 'layer3.0.conv1.weight_mask', 'layer3.0.conv2.weight_mask', 'layer3.1.conv1.weight_mask', 'layer3.1.conv2.weight_mask', 'layer3.2.conv1.weight_mask', 'layer3.2.conv2.weight_mask'])
layer2.1.conv1.weight_mask
2 3
(29, 29)
-481
(4, 2254)
(4, 481)
0.19369237907088122

It seems fine at my side. So if you could please share your errorlog with me, I can debug with more efficiently. Thanks!

[hanglearning]

Hi Xuxi,

Sorry for the late reply! I should have replied to you in the afternoon as I actually thoroughly went through the main_imp_new.py today and generated a toy res20s_cifar10 ticket model (with 3 epochs per pruning state, total 3 pruning states as well), and I directed embed_res20s_cifar10.py to read that model at line 4, and removed assert False at line 44, now the code can run successfully! I then went to go through extract_qrcode_*.py and was planning to follow up for extract_qrcode_*.py if I have any questions. I apologize for the delay.

I do still have one question about embed_*.py. I see there is a pdf generated to visualize the final QR code. If I understand it correctly, this QR code is the final mask with the signature embedded. But the QR Code shown on my pdf does not contain those 3 big corners, therefore it cannot be scanned. Is this the correct behavior of the code? I thought the output should be like figure 3 or 4 in the paper.

Now I have difficulties running extract_qrcode_res20s_cifar10.py. I used my toy ticket model 2model_SA_best.pth.tar as its checkpoint. The code caught an exception in prune.CustomFromMask.apply(m, 'weight', mask=mask_dict[name+'.weight_mask']), prune_model_custom() in pruning_utils_2.py. . Unfortunately the error message does not seem to tell anything. Here is a screenshot.

Please advise, thanks!

Btw, should I open a new issue for questions about extract_qrcode_res20s_cifar10.py?

[xxchenxx]

Yes we remove the 3 big corners because they can be added back (and also hiding the big corners will make the attack more difficult because the location is not directly shown). In Figure 3 or 4 we actually add the corners back for better visualization.

You don't need to open a new issue - we can continue to discuss here. From my perspective of view I think this might be related to pytorch version problem. Can you share me the version of pytorch you used? Thanks!

[xxchenxx]

I hope you are not in some emergency need - I can update a more detailed instruction later in May and make everything clear

[hanglearning]

Hi Xuxi,

I appreciate your quick response always! Got it. If it won't take you too much time, could you briefly explain how I may add the corners back? Or any useful resource would be highly appreciated!

I am using Pytorch 1.6.0 as directed in readme, with CUDA 10.2. I guess the issue may also be caused by the fact that I used too few epochs and pruning states. Could you share your checkpoint with me so I can verify if the issue is indeed caused by the Pytorch version?

I will be having a presentation in the coming week about the possibility of integrating the qrcode embedding into our project. I just have to prove that the concept is working. If it won't take you much time, I think once I have the extract code working with a checkpoint, I should be good for a while 😊

[xxchenxx]

The pruning states and few epochs won't cause this problem since the problem now is about pruning ... Let me work out a way to share files with you.