Training the boundary

[SuchitReddi]

In the paper, it's mentioned that:

The key idea of one-class classification methods is to capture the target class distribution and set a tight classification boundary around it, so that all non-target data would be placed outside the boundary.

What I understood was that we take bona fide speech and to train and create a tight boundary around it using oc-softmax loss function.

But when I saw the ASVspoof dataset, in LA cm protocols, the train file has bona fide and also spoof audio files.

So, what I can't conclude is, does one class classification mean we only take the bona fide speech for making a classified boundary?

I am confused about how we train the boundary. Using only bona fide speech data or using both bona fide and spoof data.

Can you clarify this for me @yzyouzhang ?

[yzyouzhang]

We use both classes to train. The idea of one-class learning is to make the bonafide representation compact and separate the spoofing attacks far away from the boundary. This is consistent with [1] in their third case of definition, where the negative class is not statistically representative in the training data.

[1] Khan, S., & Madden, M. (2014). One-class classification: Taxonomy of study and review of techniques. The Knowledge Engineering Review, 29(3), 345-374. doi:10.1017/S026988891300043X

[yzyouzhang]

Yes.

[SuchitReddi]

Thank you very much for replying this quick!

[SuchitReddi]

It would be helpful if you can provide the presentation ppt you used for making your youtube video. Thank you.

[yzyouzhang]

Here is an updated version of my slides. https://labsites.rochester.edu/air/publications/ICASSP2022Slides_Neil.pdf

[SuchitReddi]

Thanks