attack on deepspeech chinese version

rtaori / Black-Box-Audio

Targeted Adversarial Examples for Black Box Audio Systems

MIT License

67 stars 30 forks source link

attack on deepspeech chinese version #20

Open jaeholo opened 3 years ago

jaeholo commented 3 years ago

Hi! I just wonder if it is possible to use this algorithm on DeepSpeech's Chinese version. The Chinese version is just a pbmm file, do you think we can use this project to attack with pbmm file?

rtaori commented 3 years ago

Hi @jaeholo,

Sorry, I am not sure about that at all. You're more than welcome to try it out - let me know if you do!

jaeholo commented 3 years ago

@rtaori Thanks and I would have a try~

jaeholo commented 3 years ago

@rtaori Hi, I think your algorithm need to use tf.ctc_loss when attacking, and the ctc loss need the alphabet for all the possible label . For example, for English, the alphabet should contain a-z and blank. However, if we want to attack a Chinese model, we have no such alphabet, so I think I can't use this algo on Chinese model...

Is there anything wrong in my opinion? especially the ctc loss part, looking forward to your answer ,thanks