trojanNN question about the neuron selection about trojanzoo

[ain-soph]

Definitely a bug. Let me fix it and retest the performance locally.

[ain-soph]

Actually I don't see any clear performance difference. Well, it's now fixed.

Command I use:

CUDA_VISIBLE_DEVICES=0 python examples/backdoor_attack.py --color --verbose 1 --pretrained --validate_interval 1 --epochs 10 --lr 0.01 --mark_random_init --attack trojannn --tqdm

Try the new tqdm I implemented yesterday!

[ain-soph]

Just to clarify, in case you didn't notice.
We have docs at https://ain-soph.github.io/trojanzoo/trojanvision/attacks/backdoor.html#trojanvision.attacks.TrojanNN

[CHR-ray]

Thank you for fast reply. But I still have some question 1. Do you implement the training with reverse engineering data + trigger ? 2. So, the neuron selection does not affect much? And I see that activation in your screenshot almost does not change. Moreover, for Refool, I think your implementation is different from the paper. In reflection selection part: ***@***.*** It takes a set of reflection to update their weight, but you take them one by one. I think this may not affect a lot. But I just want to know whether you test the difference between these two implementation? Thanks. On Mar 4, 2022, at 2:11 AM, Local State ***@***.******@***.***>> wrote: Just to clarify, in case you didn't notice. We have docs at https://ain-soph.github.io/trojanzoo/trojanvision/attacks/backdoor.html#trojanvision.attacks.TrojanNN — Reply to this email directly, view it on GitHub<#105 (comment)>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/ASCAZHT7WSA6HDTIY5X3VTLU6D6GZANCNFSM5PZSZEFQ>. Triage notifications on the go with GitHub Mobile for iOS<https://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675> or Android<https://play.google.com/store/apps/details?id=com.github.android&referrer=utm_campaign%3Dnotification-email%26utm_medium%3Demail%26utm_source%3Dgithub>. You are receiving this because you authored the thread.Message ID: ***@***.***>

[ain-soph]

1. No, we don't implement reverse engineering training data, because we are measuring many attacks using the same BadNet problem setting. (I may implement that in the future.)
2. That might because we are currently using resnet18, which is different from VGGface that has multiple FC layers.

I've consulted the original author Yingqi Liu days ago to confirm:

1. During neuron index selection, conv layer weights (*, *, KH, KW) shall be averaged for the last 2 dimensions.
2. They don't test ResNet architectures.

[ain-soph]

Reflection Backdoor is implemented by others. I'll check it today, update codes and write docs.

And could you illustrate the problem in more details?

Btw, there is another 3rd party implementation from Tsinghua group:
https://github.com/THUYimingLi/BackdoorBox/blob/main/core/attacks/Refool.py

[ain-soph]

@CHR-ray I just checked the original Caffe implementation of TrojanNN.

It seems they don't follow Algorithm 1 to generate their triggers:

1. They don't follow the cost definition to update input trigger. They actually assign one_hot gradient to the intermediate feature map and directly conduct the backward. It's not relevant to target_value (=100) in paper.
2. There is no denoise in paper algorithm 1, but exists in the code. (there is denoise in paper algorithm 2 though)

[CHR-ray]

I mean this part for reflection backdoor: ***@***.*** In line 59, you can see that you use the image one by one to train the model (one at a time) and update the weight, but just as screenshot, the algorithm 1 in Refool actually uses a set of random sampled images in that step, and update them simultaneously. In Refool, it says: ***@***.*** ***@***.*** On Mar 4, 2022, at 2:50 AM, Local State ***@***.******@***.***>> wrote: Reflection Backdoor is implemented by others. I'll check it today, update codes and write docs. And could you illustrate the problem in more details? — Reply to this email directly, view it on GitHub<#105 (comment)>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/ASCAZHWODS2X24SRE5O7BOLU6ECX5ANCNFSM5PZSZEFQ>. Triage notifications on the go with GitHub Mobile for iOS<https://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675> or Android<https://play.google.com/store/apps/details?id=com.github.android&referrer=utm_campaign%3Dnotification-email%26utm_medium%3Demail%26utm_source%3Dgithub>. You are receiving this because you authored the thread.Message ID: ***@***.***>

[ain-soph]

I will deal with Reflection Backdoor later.

TrojanNN issue should have been solved.
I've followed original author's caffe codes to update the preprocess algorithm. They are now almost equivalent, and I have commented all differences in the code.

You may search Original in the file to see things I've abandoned because there is no performance difference.

https://github.com/ain-soph/trojanzoo/blob/main/trojanvision/attacks/backdoor/trojannn.py

[ain-soph]

@CHR-ray Just want to tell you that reflection backdoor is already reimplemented and docs are finished as well.

It is tested and the performance is good,