Private model syncing between vehicles based on isometric context transformations

Abstract

Methods, computing systems, and technology for remotely training a machine-learned model using masked inputs is presented. For example, a computing system may be configured to obtain vehicle operations data associated with a user and a vehicle. The computing system may be configured to transform the vehicle operations data based on one or more isometric transformations to obtain a masked input, the one or more isometric transformations associated with a feature masking key. The computing system may be configured to input the masked input to a machine-learned model to obtain a masked output, wherein the machine-learned model was trained using previous masked inputs. The computing system may be configured to determine, based on an action masking key, an action associated with the masked output. The computing system may be configured to control a component of the vehicle based on the action.