Hyperparameters for different ImageNet models

MadryLab / blackbox-bandits

Code for "Prior Convictions: Black-Box Adversarial Attacks with Bandits and Priors"

https://arxiv.org/abs/1807.07978

MIT License

61 stars 14 forks source link

Hyperparameters for different ImageNet models #2

Open suyeecav opened 5 years ago

suyeecav commented 5 years ago

How sensitive are the hyperparameters to different ImageNet models? I currently tested targeted attacks (slight modifications to attack loss and instance update rule) on densenet model, but the failure rate is pretty high (~50%). After careful sanity check, it seems hyperparameters might be be the main reason, especially for targeted attacks. Any guidance on this will be helpful, thanks!

andrewilyas commented 5 years ago

Hi!

We didn't tune hyperparameters too much for Inceptionv3/ResNet/VGG (in fact I think we used the same hyperparameters for all three) but we've never tried DenseNet---in general, we would suggest doing a hyperparameter grid with, e.g. 100 images (should be pretty fast) and then using the best hyperparameters from there.