cyhuang-tw / attack-vc Goto Github PK

Dear authors,

Thanks for providing the high quality codes and the nice paper. I am trying to generate the adversarial samples in the demo. Since it is not a targeted attack(no y in the demo), I could not use the targeted attack code you provide to generate the adversarial samples directly. Could you please provide the untargeted end-to-end attack code (for equation (1) in the paper )for generating the demo samples?

My attempted modification of the e2e_attack function for untargeted attack is:

def e2e_attack(
model: nn.Module,
vc_src: Tensor,
vc_tgt: Tensor,
adv_tgt: Tensor,
eps: float,
n_iters,
) -> Tensor:
ptb = torch.zeros_like(vc_tgt).normal_(0, 1).requires_grad_(True)
opt = torch.optim.Adam([ptb])
criterion = nn.MSELoss()
pbar = trange(n_iters)

with torch.no_grad():
    org_out = model.inference(vc_src, vc_tgt)
  
losses = []
ptbs=[]

for _ in pbar:
    adv_inp = vc_tgt + ptb
    adv_out = model.inference(vc_src, adv_inp)
    loss =  -  criterion(adv_out, org_out)
    opt.zero_grad()
    loss.backward()
    opt.step()
    ptb.data.clamp_(-eps, eps)

return vc_tgt + ptb

However, the adversarial samples I generate(.e.g epsilon = 0.1 ) is different from yours in the demo.
Any advise or reply will be highly appreciated!

Zihao

Hello,Thanks for your code and the pretrained-model. However, the GoogleDrive link become invalid these days. Would you please update this link? Thanks you very much!

Hi, dear author

I learn a lot from this work, it's a great work.

Currently, I‘m working on a project, I want to do normalization on my data (mel-spectrogram ) for a better training performance.

I noticed that the std and mean of the data is precomputed, would you mind to tell me how to compute the std and mean of the data.

Thanks.