Savants: Federated Learning
Federated learning is changing the way machine learning interacts with the network. Even though federated learning addresses privacy by not sharing raw data, it still depends on a proper network infrastructure. Because the network characteristics have a significant impact on the training duration, the construction of federation of user devices and enabling of convergence of federated learning requires reliable, real-time and secure data exchanges. As a result, the data exchanges at the application level across heterogeneous access networks need a new perspective in which the system measures the communication of data and adapts appropriately to a specific environment.
With NEMI, we combine the research on federated learning with the one on reliable data exchange, by proposing and evaluating network extensions. They involve the addition of a data substrate to the end-to-end communication that acts as a middleware for an access network with specific characteristics. NEMI enables the acquisition of real-time performance statistics to create a self-adaptive solution for the federated learning. We have validated the prototype using remote maintenance use case involving federated learning as a reference, thus anchoring the obtained results into the requirements of a real-world use case. This work creates the foundation for a comprehensive and highly adaptive data substrate within the network that addresses the communication needs and improves the network for federated learning.