Mismatch of comment and data type in secure_mode

[heyday98]

Hi team,

I have a question about _generate_noise function in optimizer.py. From the comment in _generate_noise function, it references p=53 (as though using float64), but I don’t see any explicit dtype set. By default, PyTorch generates float32 noise (with p=23). I run few test and found that generated noise has dtype of torch.float32. If that’s correct, we might need 6 samples (n=3) to achieve 128-bit security. Is the code/comment correct, or am I missing something?

Thanks a lot!

[HuanyuZhang]

Let me check and get back to you.

[HuanyuZhang]

I think you are correct. The comment implicitly assumes torch.float64, which is not very common (by default it is 32 as you mentioned). Will change the comment to clarify.