IP protection should be top of your agenda in mitigating security vulnerabilities. How does raw data play into such concerns and how do methods for obtaining or harvesting training data come into play? Thought about data augmentation, classification, or modification? Have you considered your model training, like improving training speed or quality of the model? Do we need customised hardware or new architecture? Will integrating into a larger system help?