johndpope / dftnd

attack demo

To generate attacked images, use the following example. Make sure to specify the necessary commands.

python main.py --area C --attack_type PGD_l2 --attack_goal N --model_type resnet18 --strategy WB 

And then the following commands will be executed.

python PGD.py --area C --attack_type PGD_l2 --attack_goal N --model_type resnet18 --strategy WB

The arguments behind are the ones in PGD.py file, which are be used to set the conresponding parameters.

To submit your attacks, name your main attack file with the attack name such as 'PGD.py'. Originize your source files and make sure the attack can run such as the following example.

python PGD.py --area C --attack_type PGD_l2 --attack_goal N --model_type resnet18 --strategy WB

How to add your own attacks?
In main.py, make sure to keep attack_type same with attack_type in your own attack files. For example, 'PGD_linf' and 'PGD_l2' are the attack_types used in PGD.py. So the codes in main.py shoule be like:

if attack_type == 'PGD_linf' or 'PGD_l2':
    attack = 'PGD'

Here attack is the name of your attack file.
Another example is adversarial patch attack.

if attack_type == 'Patch':
    attack = 'Patch_Attack'

The codebook is here:

Codebook:

head		details
area	attack type	attack goal	attacker knowledge
			model type	strategy
C -- Classification	0 -- non-attack	T -- target	V16 -- VGG16	WA -- White Box
D -- Detection	1 -- lp attack	N -- non-target	V19 -- VGG19	TB -- Transfer based Black Box
F -- Face recognition	2 -- patch attack		R50 -- Resnet50	QB -- Query based Black Box
	3 -- trojan attack		R18 -- Resnet18
			L04 – Lenet with 4 convs
			I03 -- inception3