Algorithm Issue

[astorfi]

Hi,

Thanks for releasing the code of your paper. In the algorithm 1 of your paper, you first add noise to the gradients and then clip the gradient. According to definitions in The Algorithmic Foundations of Differential Privacy and Deep Learning with Differential Privacy we MUST clip the gradients first to limit and control the sensitivity. Then, we can add noise to the gradients. So your algorithm is apparently inconsistent with those definitions.

Is that correct? Thanks

[MiaoXiong2320]

I guess they have proved that we can clip the weight directly instead of clipping the gradients in the paper.

[astorfi]

@Emma0120 Thank you for your clarification. Yes, that what they claim. But even if clipping the weights is the right choice, still, I think it should be done before adding noise to bound sensitivity. Doing it after adding noise means the effect of noise is enforced on the clipped weights which is NOT correct.

[jcForDream]

@Emma0120 Thank you for your clarification. Yes, that what they claim. But even if clipping the weights is the right choice, still, I think it should be done before adding noise to bound sensitivity. Doing it after adding noise means the effect of noise is enforced on the clipped weights which is NOT correct.

Hi. Did you run through the author's code? If it did, how did you do it?