Federated deep learning for UAV Mobile Edge Computing environments

Unmanned Aerial Vehicles (UAVs) are increasingly used by public and private organizations such as the Politie, Brandweer or Rijkswaterstaat. The data these organizations collect can be used to achieve common goals such as vehicle detection, license plate or obstacle recognition. However, sometimes the data that they collect is private and cannot be shared, limiting the learning ability of the shared deep learning model. Federated deep learning is a technique where multiple mobile edge devices (such as edge-enabled UAVs) train a commonly shared deep learning model using their local data. Afterwards, they submit the model parameters to a shared location where the model finally is harmonized. This way, organizations do not have to share their private data thus ensuring privacy while still being able to learn from others' knowledge. In addition, compared to traditional centralized deep learning frameworks, less data needs to be transferred from the UAV to a remote station, which is valuable in unreliable connectivity situations. Current federated learning frameworks have yet to improve on communication overhead, hardware utilization and robust federated learning [1, 2].

The aim of this MSc topic is to design a federated learning framework for a swarm of (or at least two) UAVs to achieve vehicle speed detection from UAV videos.