Workshop: 13th Annual International Workshop on Innovating the Network for Data Intensive Science — INDIS
Authors: Carlyn Lee (California Institute of Technology, Oligo Space); Harvey Newman (California Institute of Technology); Yifan Wu (California State Polytechnic University, Pomona); and Spencer Axani (University of Delaware)
Abstract: Scientific instruments are increasingly built from large numbers of small, heterogeneous edge devices that operators cannot inspect or trust. Citizen-science cosmic-ray networks, in which commodity smartphones act as particle detectors, are an extreme example. Centralizing raw observations from such networks is the default architecture, but its wide-area traffic grows with every observation, and it requires participants to surrender raw data. We present an end-to-end federated alternative and attempt to measure its network behavior on smart-phone data. Our system trains a compact artifact-rejection classifier for smartphone cosmic-ray detections entirely in the browser. Hits never leave the phone; clients exchange small weight payloads with a FedAvg coordinator on the National Research Platform. At matched accuracy, federation cuts WAN traffic 3.2–11.4 x, governed by per-client data volume, not client count, locating the crossover. We also address aggregator hardening and validation against a low-cost calibrated muon telescope.
Back to 13th Annual International Workshop on Innovating the Network for Data Intensive Science — INDIS Archive Listing
Back to Full Workshop Archive Listing