I word as an AIML associate in a startup. A client wants to implment a system where,

Data providers will be giving sensitive data, they will need ml/dl based solutions to be implemented on this data

Since data scientists will be competing with each other to build solutions and they do not have much credibility to them, we need to implement some sort of encryption so that the ML alogrithms will not get affected while still provide some security.

The basic encryption that comes to mind is encoding, which can be said it is not an encryption at all, but it does hide some details.

I know that ml algorithms cannot work/nearly impossible on data that is encrypted with 128/256 bit encryption. I want to implement a system where I can obsure the dataset as much as possible while providing the least amount of performance impact both in terms of computation and accuracy of this model which is built on the encrypted dataset.

I read about order preserving encryption. What are your thoughts on it? Can a ML model be build on a dataset that is encrypted with order preserving encryption (or other forms of encryption)? Please mention any ways with which I can abstract the data?

I think in theory decision trees will work but not linear models. What about SVM and other non-linear models and what about NN-based models?