search

inspire-group / DetectorGuard

Code for "DetectorGuard: Provably Securing Object Detectors against Localized Patch Hiding Attacks"
MIT License
14 stars 0 forks source link

Bagnet question #2

Closed skimtea closed 2 years ago

skimtea commented 2 years ago

Hello,I appreciate that you guys could open your paper source code. There are some questions at below puzzled me when i used your algorithms to defend adversarial patch in some datasets. The officially designed coco dataset cannot be used for image classification tasks; However, you perform image classification of the coco dataset on bagnet, get local_features, and then get the objectness map; So, How do you implement this image classification task firstly? I hope you could help me solve this question. Thank you so much! !

xiangchong1 commented 2 years ago

See Section 5.1 -> Objectness Predictor -> BagNet-33 for the BagNet training strategy. See https://github.com/inspire-group/DetectorGuard/tree/main/checkpoints for ad-hoc BagNet training scripts.