Close
I rewrote the post a bit to clarify layer numbers.

kiowa said:
For example, consider there is a white scratch on a black background, if you create a mask from the unfiltered background, the white scratch will be retained because it's white and is masked from filtering. Creating from the filtered layer avoids this problem.
In theory, using an unfiltered layer as a mask defines noise level, using a filtered layer as a mask defines noise reduction amount. but I couldn't find obvious difference between both ways.
it might not matter which I use. but in these kind of case, I'm prefer using noise level.

The darkest areas contain much less noise than the highlight areas in this image. Your assumption is quite reasonable because black ink can easily create a large contrast with other color inks and white paper, introducing high luma channel noise level. But if an area is dark enough, the noise level begins to decline with darkness level on the contrary. I think grayscale masking is a powerful method for preserving bright areas from overfiltering, but users should be aware that this masking method also has some deviation from real noise levels and should limit masking strength and look out for overfiltering in dark areas.
While I'm using only removing CrCb and Darken only mode, darks can avoid overfiltering because the ways hasn't white back.

The problem of overfiltering on darks were maily caused by normal filtering.
Therefore, I'm using a opposite mask to avoid filtering on darks as filtering a whole image. because darks already haven't any noise after removed screening with that method.

for example,

apply one weak filter set => create a new layer with a mask WITHOUT inversion , then up gamma at 6~9 => apply another strong filter set at the new layer