/KWSAttackAdvCGAN

Primary LanguagePython

KWS Adversarial Attack with CGAN

This is the implemention of Fast Speech Adversarial Example Generation for Keyword Spotting System with Conditional GAN

Setup

The pre-trained victim model is available for use and can be downloaded from here

The pre-trained target(generator) model is available for use and can be downloaded from

single model: link:https://pan.baidu.com/s/1xLJqCljv1mjDSiMZB8mrGw fetech code：gk5c

multi-model ensemble: link：https://pan.baidu.com/s/1geJxrXoJxVYgLIyX5c3q0A fetech code：8h7s

Notice: The directory structure of the checkpoints file is as follows:

|-- checkpoints
    |-- vgg19.pth	
    |-- resnet18.pth
    |-- resnext29_8_64.pth
    |-- dpn92.pth
    |-- densenet_bc_250_24.pth
    |-- wideResNet28_10_9414.pth

The folder structure of generated adversarial examples is as follows:

|-- output
    |-- wideresnet
    	|--target_yes
    		|--generated
    			|-- gen
	    			|-- no
    					|-- fake_yes2no_xxxx.wav
    			|-- real
    				|-- no
    					|-- real_yes2no_xxxx.wav
    	|--target_no
    		|--generated
    			|-- gen
    				|-- yes
    					|-- fake_no2yes_xxxx.wav
    			|-- real
				    |-- yes
    					|-- real_no2yes_xxxx.wav

Attack Evaluation

To run:

python test_gan.py --data_dir original_speech.wav --target yes --checkpoint checkpoints