Federated Analytics with Flower and Pandas
With Flower 1.2, you might have noticed that we have added a new example to the repo. If you were curious enough to check out the code, you might have been left a bit straddled… Indeed, this example, unlike the others, is for doing Federated Analytics instead of Federated Learning. In this post, we'll briefly go through what Federated Analytics consists of, what this example is about and, how it is built.
The setting in which Federated Analytics is performed is analogous to the usual FL setting, here we will take the example of 100 clients, each having a subset of the data (not shared with anyone). Instead of training a model, our goal will be to compute statistics of the distributed dataset without allowing data sharing. For instance we might be interested in the mean of the distributed dataset, which can be calculated by asking each client the mean and the length of their dataset and then aggregated the results to obtain the global mean. This is the gist of Federated Analytics, to learn more you can check about this article by Google.
Fun fact: Federated Learning is actually a subset of Federated Analytics, with the Machine Learning model as the statistics being computed. It's just a more complex statistic than a simple mean…
