Federated Learning for Radar Systems - Sharing Knowledge without sharing Data

Federated Learning (FL) is a technique that was introduced by Google in the area of Internet of Things (IoT) as a privacy preserving way of using user data in the training of their neural networks [1]. The core element of FL is a distributed learning algorithm that runs in local clients, e.g. phones or computers, and uses data stored on these devices. After this local training, the models trained on the clients are uploaded onto a server, which performs a model aggregation, e.g. by averaging the model parameters, and creates a new global model that is again distributed to the clients. This process is continuously repeated to improve the model performance over time and in this way the clients benefit from each other without sharing their data. In this paper, this learning scheme is proposed for radar networks and radar systems in general. This principle has some advantages since data, which might be confidential, does not have to be shared. Another possible application is that customers of a company benefit from each other when a common model is aggregated and redistributed among the customers using the learning results of all customers. A proof of concept example is presented using data compression by an autoencoder network.