So I'll be licensing my deep learning model with pre-trained weights for deployment to a customer. The usage of this deep learning model will be done over this customer's servers. For security reasons, I was wondering what I could do to obfuscate my Python code and model to make it harder if anybody potentially wanted to reverse-engineer it.
So far, I have used Cython to compile my Python script (used to handle data and call the model) into C. I also have followed Making an executable in Cython to make an executable. Is there any more that I could do to obfuscate better or more efficiently?
In addition, I'm more concerned about obfuscating my pre-trained model's weights. I have the model in ONNX and Tensorflow Lite formats. What are the best ways to protect my models (for ONNX and Tensorflow Lite)?
Note, these models are meant to be deployed in a real-time setting — so it wouldn't be good to continually load the model.
Check out https://pypi.org/project/pyarmor/
afaik there isn't a way to decrypt it