Hari Bhaskar SFollow


Bundling ML models can compromise model security and tightly couples model execution with application execution. Managing computing resources can be difficult if multiple applications access the same model and maintaining multiple ML models through their life cycle can become increasingly complex as the application grows. This disclosure describes techniques that utilize a function-based paradigm for on-device machine learning models. The functions can be implemented using various ML models, implemented in isolated containers. Per the techniques, on-device applications (e.g., mobile apps) can be defined with IAM/roles to access one or more functions that provide ML functionality. The techniques provide enhanced security by isolating model security from app security. Isolation also allows a model to continue to be available (to other applications) in case a particular application goes down. Further, the techniques enable centralized monitoring (e.g., through endpoint management services) to perform resource management. An on-device model registry is provided to maintain a mapping of functions to models deployed on device. Separating the model from the application also allows easier access control, versioning, and deployment.

Creative Commons License

Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.