Noise communication for federated learning

A server system for training a federated machine learning model based on received prediction outputs provided by client computing devices, the system comprising:

• one or more processors and one or more non-transitory media storing program instructions that, when executed by the one or more processors, causes the one or more processors to perform operations comprising:
• sending a distributed model instance of a federated machine learning model to a client computing device, wherein an updated model instance of the distributed model instance on the client computing device is derived from the client computing device training the distributed model instance with labeled images accessible to the client computing device;
• sending noise images to the client computing device, wherein a set of prediction outputs related to the noise images is derived from the client computing device executing the updated model instance on the noise images;
• and
• in response to the set of prediction outputs being obtained from the client computing device, training a server model instance of the federated machine learning model by using the noise images as server-side training inputs and the set of prediction outputs as reference feedback,
• wherein a set of objective function values is included as part of the server model instance based on the training of the server model instance with the noise images and the set of prediction outputs.