Machine Learning and Neural Model encryption

Machine Learning and Neural Models exhibit characteristics of data and take enormous amounts of time, computing resources, and human expertise. It is an intellectual property of the organization. Thus, it is of utmost importance that it is well protected to curb illegitimate reproduction and distribution, which can lead to intellectual property abuse and loss of revenue.

Knox for Model Protection

Knox offers a Neural Model Encryption feature for customers to easily deploy their AI or ML Models on the mobile device while leveraging Samsung Knox’s defense grade security. This feature ensures that plain ML model is never stored on the device. The processing for an encrypted model is separate to that of an unencrypted model, and this separation provides model isolation for customers of Knox for Model Protection.

Along with the encryption of ML files, Knox for Model Protection provides access control over the application packages that can load the encrypted model.

The previous diagram highlights two functions of the Knox for Model Protection service:

• Knox ML Encryption Tool — The tool takes the plain ML model and encrypts it or converts it to ONNX. ONNX conversion depends on the model format.

• Knox APIs — The APIs to create ML Model Protection session, load-execute-close the ML model, and finally destroy the session.

Knox ML Encryption Tool

Knox ML Encryption Tool is a comprehensive software program providing you the Knox Neural Model Encryption service. The Interactive UI of the tool helps you with the process of creating encrypted model packages from selecting a plain ML model file to defining the application access control for the selected ML model. To learn more about the tool, see Knox ML Encryption Tool.

Knox APIs

Knox provides APIs to load and run your encrypted ML model. Since the encrypted ML model is treated differently than a non-encrypted model for security reasons, the APIs help you create a service session, invoke the encrypted ML model loading, and get the model inference. To learn about the ML Model Protection API implementation, see Knox for Model Protection — Knox SDK APIs.

Where to start