Thing is, I don't have to to extract your AI. I can just separate your obfuscated AI execution into a DLL. And then I call that DLL with lots of randomly gener

Thing is, I don't have to to extract your AI. I can just separate your obfuscated AI execution into a DLL. And then I call that DLL with lots of randomly generated input data and estimate the numerical gradients from that. And now I have everything I need to copy things over into a similar NN architecture. Yes, it might take a few days to evaluate everything, but CPU time is cheap compared to research and employees needed to reverse engineer your implementation. That said, NN-512 is great because it produces optimized CPU code, thus making deployment cheaper.