sclbd / blackboxbench Goto Github PK

Would you like to ask what your CUDA version is

Hi, I thank you for your contributions in robustness study. When I run the black-box attacks, if the total failure rate is 0.8, clean acc is 0.9, what is the robust acc under black-box attacks?

1. Misclassified samples in normal testing is thrown. Then robust acc is 0.9*0.8=0.72.
2. All samples are tested, thus, robust acc is 0.8.
   I am not sure which one is correct?
   Could the dear colleagues from BlackBoxBench team help me?
   Thanks!
   I wish you are all good!
   Momo

Kindly change pip install to (add s to requirement.txt):
pip install -r requirements.txt

Hey,

I have been using this library for quite a while for my research ( this is the time to mention that I had a great experience doing so).
I noticed that you have changed the library completely in your last update. There are no attacks implementations, and basically all query-based attacks files are no longer available, thus cannot be performed and used.

I would love to get some help, since my research is based on the ability to run those attacks.

Thank you.

File "attack_imagenet.py", line 51, in
from attacks.score.nes_attack import NESAttack
ModuleNotFoundError: No module named 'attacks.score'

8 decision-based attacks: Boundary attack, OPT attack, Sign-OPT, Evoluationary attack, GeoDA, HSJA, Sign Flip, RayS.

Hi, dear team-members of SCLBD,

I am not sure which methods take less time to attack the same DNN models (e.g., ResNet) with the complete deceit.

If the statics result of running time is available, I am sincerely grateful.

Best Regards!
Momo