Method for privacy-preserving federated learning

Abstract

A computer implemented method performed by a user device suitable for training image generation machine learning models. The method receives sets of trainable parameters from a server, uses local training data on the user device to determine the most suitable sets, applies the parameters to an image generation model and calculates a loss for the parameters, selects the parameters that provide the model with the lowest calculated loss, uses local training data to train the set of trainable parameters and extracts the updated parameters from the training, adds noise to the extracted set of parameters (noisy trainable model parameters) and finally transmits the noisy trainable model parameters to the server for aggregation with other similar parameters received from other user devices to generate multiple models. In this way, federated learning of machine learning models may create user personalised models preserving also the privacy of the users. A model may convert text to image, text to music, image to music. [Fig 4]