winterwindwang / SpeechAdvGan

Speech Adversarial Examples with GAN

This is the implemention of Speech Adversarial Examples Attack with Generative Adversarial Networks

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 here

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

|-- checkpoints
    |-- music_genres_generator
        |-- target_blues
        |-- target classical
        |-- ...
    |-- speech_common_generator
        |-- target yes
        |-- target no
        |-- ...
    |-- sampleCNN.pth	
    |-- wideResNet28_10.pth

Attack Evaluation

To run:

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