scope:

This standard defines a framework and architectures for machine learning in which a model is trained using encrypted data that has been aggregated from multiple sources and is processed by a third party trusted execution environment (TEE). A distinctive feature of this technique is the essential use of a third-party TEE for computations. The standard specifies functional components, workflows, security requirements, technical requirements, and protocols.

Purpose

There are many use cases in industries ranging from finance to healthcare to education where practitioners wish to apply machine learning to data sets that are aggregated from sources that cannot or should not be combined due to regulatory, competitive, or ethical considerations. Two fundamentally different approaches exist for addressing this: federated machine learning and shared machine learning (SML). In federated machine learning, models are constructed by training local models on local data samples and exchanging intermediate parameters (e.g., the weights generated for a neural network or bases for a vector space that defines an embedding) among multiple parties to generate a global model shared by the participants. In trusted execution environment (TEE) based SML, the data are shared but are encrypted and given to a trusted third party to construct a model that is then shared. This standard will provide a verifiable basis for trust and security.