speagleyao/cost-sensitive-adv

Combination of cost-sensitive learning and adversarial learning

Cost-sensitive Adversarial Learning

Combination of cost-sensitive learning and adversarial learning

News

Ready to do some fundamental experiments on basic settings!

Doing experiments to see how one specific attack from one kind to another is influenced.

Settings

• Normal settings: normal adversarial training
• Specific settings: only generate adversarial examples between given classes
• Test: check accuracy of natural, adversarial, and specific attack results

Expeiments

• Do experiments between two specific classes, e.g. frog-truck.
• Do experiments between two types of classes, e.g. animal-vehicle.

What I've done

In the first experiments, I now have trained 5 models as follows:

• res18_natural: Train the model with natural data
• res18_normal : Train the model with natural data + all untargeted pgd data
• res18_trades : Train the model with TRADES loss
• res18_alltar : Train the model with natural data + all targeted pgd data
• res18_sp_35 : Train the model with natural data + all targeted pgd data between class 3&5
• res18_ow_35 : Train the model with natural data + all targeted pgd data from class 3 to class 5

All these models are tested under natural testset, untargeted pgd testset, and targeted pgd testset between two specific classes.

Then, I ran test under several SOTA models, and output their test results under targeted attack and confusion matrixes.

Targeted attack results can be seen in here.

Confusuin matrixes can be seen in here.