Data-driven Internet of Things: A Holistic Paradigm of Data Generation, Dissemination and Application in Autonomous Cars and RF Fingerprinting

Abstrat

This two-part talk describes a data- and systems-centric design to enable emerging applications for the Internet of Things (IoT) paradigm.  In the first part of the talk, we explore how contextual knowledge of the environment, obtained from vehicle-mounted LiDAR, camera and GPS sensors, can expedite the creation of high bandwidth millimeter wave links in autonomous cars. We describe fusion models that combine these diverse data sources as well as multi-modal federated learning to lower millimeter wave beam selection time compared to what is possible via the standard today. The second part of the talk addresses the challenge of authentication in dense deployments, where hundreds or thousands of wireless devices may operate in untrusted environments. This is accomplished at the physical layer by learning subtle but discriminative distortions present in the transmitted signal, also called as RF fingerprints. Deep convolutional neural networks (CNNs) combined with data augmentation, both with and without receiver feedback, are used to demonstrate 99% identification accuracy in these large radio populations. For each of these two use cases, we also describe our experiences with the cross cutting component of data collection, involving large, community-scale NSF platforms for advanced wireless research (PAWR) to the world largest RF emulator- Colosseum, and finally, real-world autonomous vehicle testbeds. In closing, we summarize ongoing efforts to democratize access to wireless datasets, create tools and software APIs to create such datasets, and design learning resources at all levels to advance the use and sharing of RF data.