Please check my work "The Normal Distributions Indistinguishability Spectrum and its Application to Privacy-Preserving Machine Learning" on arXiv.

I updated my work The Normal Distributions Indistinguishability Spectrum and its Application to Privacy-Preserving Machine Learning on arXiv.


In this work, I described a theory what we called NDIS theorem. It computes a closed-form analytic computation for DP parameters of mechanisms with arbitrary Gaussian outputs. Ideally, for any mechanism with (asymptotically) Gaussian outputs, the NDIS theorem can determine the tightest DP spectrum. In the paper, I demonstrate its applicability to random projection and its downstream application in approximate least squares (ALS). Please check my paper for more details!

One point to note is that applying the NDIS theorem to obtain a clear DP statement requires additional effort in calculations, which might still pose challenges for non-expert users. I am working on mitigating this weakness and broadening its applications. If you have an interesting complex ML task and would like to explore it with the NDIS theorem, please do not hesitate to send me a message if you’re interested in working with me :smile: .