An Efficient Edge-Cloud Partitioning of Random Forests for Distributed Sensor Networks

Intelligent edge sensors that augment legacy 'unintelligent' manufacturing systems provides cost-effective functional upgrades. However, the limited computing at these edge devices requires tradeoffs in efficient edge-cloud partitioning and raises data privacy issues. This work explores policies for partitioning random forest approaches, which are widely used for inference tasks in smart manufacturing, among sets of devices with different resources and data visibility. We demonstrate, using both publicly available datasets and a real-world grinding machine deployment, that our privacy-preserving approach to partitioning and training offers superior latency-accuracy tradeoffs to purely on-edge computation while still achieving much of the benefits from data-sharing cloud offload strategies.