GAP calculation function return zero array

[verazuo]

https://github.com/JunfengGo/AEVA-Blackbox-Backdoor-Detection-main/blob/9e2fc44573f665097bd555abbc3ffb1ef06fee0c/outlier.py#L18-L38

Hi,

Thanks for sharing this wonderful work!

I am trying to reproduce the results. And I found the function def get_s(num_labels) which is used to calculate GAP value only returns a zero array because it has never been used after initialization. This would then causes the failure of the analomy(array) function.

Would you please check the code?

Best wishes, Vera

[JunfengGo]

Can you plz follow the implementation procedure described in the README.md. You should first generate the adversarial perturbations and then run analomy(array) function.

[JunfengGo]

python -u detect_main.py --sp 0 --ep 10 --cuda 0 --task cifar10

[verazuo]

Hi Junfeng,

Thank you for your response. Yes, I follow the instruction of the README.md and run the following commands.

python -u detect_main.py --sp 0 --ep 10 --cuda 0 --task cifar10

The adversarial perturbations are also generated and successfully loaded in outlier.py. What confuses me is that, if you look at the get_s() function I mentioned above, the return variable s has never been used after initialization (line 28). Though there is a loop that does some calculations, but the results are stored in the matrix variable coll. I am wondering if any lines are accidentally missed. Maybe a formula that would transfer the variable coll to s?

Best wishes, Vera

[JunfengGo]

Sorry for confusing you. You are right. I will update the corrected Code ASAP

[JunfengGo]

Thanks for your correction, i have updated by : adding s[i]+=v below. Plz see my updated code