search

xuanqing94 / RobustNet

Robust neural network
MIT License
7 stars 4 forks source link

The RSE defense: low accuracy after the C&W attack #1

Open adam-dziedzic opened 5 years ago

adam-dziedzic commented 5 years ago

Hello Xuanquing,

I trained the VGG16 model with init noise set to 0.2 and the noise in other layers set to 0.1. The accuracy of the model is roughly about 92% on clean data. When I run your attack code in attack.py, the accuracy after the attack drops to about 10%, which is equivalent to a random classifier. I implemented the ensemble method since I could not have found it in your code but the accuracy didn't increase, my code is here: https://github.com/adam-dziedzic/RobustNet-1/commit/4349023cafb438fdf068b1e1eb0901f3aaa5b228. Could you please let me know how to use your ensemble model so that the accuracy after the attack is higher? Maybe, you could provide the full code for the attack and the ensemble method?

Adam

xuanqing94 commented 5 years ago

Can you try the code in dev branch? https://github.com/xuanqing94/RobustNet/tree/dev

adam-dziedzic commented 5 years ago

Thank you for the quick answer. I'll look into the dev branch.

zhyhy commented 4 years ago

Hello Xuanquing, Thanks for your contribution. I try to train the VGG16, but I find that in the code, the noise layer is after the layer of convolution? And in the paper, the noise layer is before the layer of convolution? And I'm confused, maybe can you tell me which one I should follow? I'm sorry that I'm not good at English...

xuanqing94 commented 4 years ago

Hello Xuanquing, Thanks for your contribution. I try to train the VGG16, but I find that in the code, the noise layer is after the layer of convolution? And in the paper, the noise layer is before the layer of convolution? And I'm confused, maybe can you tell me which one I should follow? I'm sorry that I'm not good at English...

https://github.com/xuanqing94/RobustNet/blob/dev/models/vgg_rse.py#L43

zhyhy commented 4 years ago

Thank you .