Federated learning (FL)
Google introduced FL in 2016 in response to the increasing concerns about the misuse of personal data. Since then, it has emerged as a standard approach to processing data at source and reducing its exposure, thus protecting its privacy and confidentiality.
The approach utilizes a shared pre-trained foundation model. Collaborating parties download and train the shared model using their own private data and utilize encryption to secure the new model and upload it to the central location. This includes only the trained model or the weights. The split of training data minimizes data exposure and allows regulatory compliance. The following diagram illustrates a typical federated learning configuration:
Figure 10.7 – Typical federated learning system
TensorFlow has introduced some support for FL, and you can find an example of implementing FL with the MNIST dataset using Keras at https://www.tensorflow.org/federated/tutorials...
