Abstract: A personalized eye model is used to generate synthetic gaze features at ground-truth eye poses Gg. Corresponding synthetic gaze poses Gs are estimated from the synthetic gaze features using an average eye model. A linear regression is applied between Gg and Gs to generate a gaze correction function. The gaze correction function represents differences between the synthetic gaze Gs of the subject eye at the display and that of the average eye model Gg at the display, but does not contain security- or privacy-sensitive information. Further, the personalized eye model cannot be recovered from the gaze correction function, and thus the gaze correction function can be stored unencrypted and available for use during a cold boot of a device prior to login. On a cold boot of the device, the gaze correction function may be accessed and used with an average eye model to improve gaze-based interactions.