Why Torch and Scikit-Learn, and not Keras?

Zama works on a specific product vision and executes deliverables on a quarterly basis. We have restricted ourselves to scikit-learn (for machine learning) and torch (for deep learning) for the moment.

So, at this time, per se, you can’t natively compile a keras model in FHE but you can however turn the forward function into numpy, and the rest should follow successfully, with Concrete Numpu.

But if you are willing to contribute, please help! Please have a look here in Concrete Numpy or [here in Concrete ML(https://docs.zama.ai/concrete-ml/developer-guide/workflow/contributing).

New version of Concrete-ML (v0.3): now, for custom models (ie, Deep Learning), we support ONNX, and so, virtually anything which could be converted to ONNX (including keras, tensorflow)