My objective is to obfuscate my model so that the clients can’t look into model and the inference code. If I package my model and code to standalone executables using Glow, will this objective be achieved? I did see that Glow supports AOT compilation. Will this solve the problem?
It should do a decent job, you’ll end up with a binary for running the model itself along with binary weights to load. Assuming you strip out any debug info then they’d have to understand the model based on machine code.