Disclosed are a neural network model update method and device, and a computer storage medium. The method includes: randomly generating a preset number of sub-neural network models as nodes in a blockchain; using a ring signature to share a local data set in the blockchain, and uniformly dividing a data set in the blockchain to generate a training set and a test set; training each node through the training set to generate a trained model, packaging the trained model as a model transaction and sharing the model transaction in the blockchain; using the test set selected by voting to test the model transaction and generating a test result; when the test result is greater than a benchmark evaluation, taking the sub-neural network model corresponding to the test result as a valid vote; and voting a previous block corresponding to the valid vote, selecting a consistent block, and updating all nodes.