search

QQiuyp / FIN

This is the source code of paper FIN: Flow-based Robust Watermarking with Invertible Noise Layer for Black-box Distortions, which is received by AAAI' 23.
48 stars 5 forks source link

whether the forward-processing of the INL would be used in the training process or not #4

Open youyiketing opened 1 year ago

youyiketing commented 1 year ago

The code you gave in the train.py shows that you didn't distort the watermarked image into a black-box version for decoder training, and the noise layer for training only contains white-box distortion, which is different from the structure you gave in the paper. So I wonder whether the forward-processing of the INL would be used in the training process or not? `#################

forward:

    #################

    stego_img, left_noise = fed(input_data)
    stego_noise_img = noise_layer(stego_img.clone())

    #################
    #   backward:   #
    ################

    guass_noise = torch.zeros(left_noise.shape).cuda()
    output_data = [stego_noise_img, guass_noise]
    re_img, re_message = fed(output_data, rev=True)`