A method of aggregating models

Computer-implemented federated learning method comprising: for each of a number of clients 40: determining a diversity score D of a dataset 41 corresponding to that client for training a machine learning model W, wherein the diversity score is a measure of dataset variability; aggregating, weighted by the respective diversity score, models 43 corresponding to each of the clients; and sending the aggregated model 46 to at least one receiving client 47. This may result in a less biased model. Models trained in this way may be used in applications such as detecting shadows in images. Aggregating the models may include assigning each client to a cluster based on dataset attributes, generating aggregated cluster weights for each cluster, and aggregating the aggregated cluster weights. The aggregation may be performed on one of the clients, or on a central server. The method may further comprise, in response to a trigger condition, for each data sample determining a confidence score of the data sample and adding the data sample to the dataset if the score is above a threshold, or discarding it if the score is below the threshold.