DP-SGD implemented in Tabula

[Zehao-Zhang]

Hi Zilong, do you think it's possible to add DP-SGD into the tabula due to the privacy concern? I tried to add what you did in CTABGAN+ into the tabulatrainer. The fit process runs with large training loss. The problem is the sample method cannot generate any result and it runs infinitely. Will appreciate it a lot if you can give some ideas about this.

[zhao-zilong]

Hi @Zehao-Zhang sorry for the long wait. To implement DP in tabula can cause several problem. Because in tabula, you first generate a sentence and then sentence translated to tabular data. If DP is added, the sentence can be invalid to translate to tabular data. But in GAN, DP will only make the generation worse, it can still output the valid format data. So the conclusion is as long as there is SGD, you can add DP, but for Tabula, it will indeed make the generation fail a lot.

[Zehao-Zhang]

Hi @zhao-zilong , thank you so much for the reply, it makes a lot of sense. If you don't mind, can I ask that what kind of privacy protection can be integrated into your model?