search

lixucuhk / adversarial-attack-on-GMM-i-vector-based-speaker-verification-systems

Implementation of Adversarial Attacks on GMM i-vector based Speaker Verification Systems (ICASSP2020) https://arxiv.org/abs/1911.03078
Apache License 2.0
33 stars 10 forks source link

Which receipe of kaldi is used ? #4

Closed GolferChen closed 4 years ago

GolferChen commented 4 years ago

Hello, thanks for releasing the code.

I am currently trying to reproduce the white box attack result of iv-mfcc.

My question is which receipe of kaldi is used.

According to my experience on Kaldi, it seems that kaldi_root/egs/voxceleb/v1 is used.

However, I did not see LDA dimenison reduction in your code, which is used by kaldi_root/egs/voxceleb/v1.

Thanks a lot.

lixucuhk commented 4 years ago

Yeah, you are right. Kaldi scripts have LDA dimension reduction, but I skipped that function for simplicity when computing backward gradients. In my latter works, I added it into the system and verified that the LDA dimension reduction still cannot prevent systems from being attacked. Hope this can meet your queries. Thank you!

GolferChen commented 4 years ago

Ok. Thanks for your reply. In fact, I agree with you since LDA dimension reduction is totally differentiable.