Federated Learning for Low-Latency Emitter Identification from Space

Emitter identification (ID) is a critical function in many adaptive radio-frequency (RF) systems, informing interference management, communications planning, and anomaly detection. Specifically, RF emitter ID of communication transmitters often facilitates specific emitter geolocation in contested environments, as identification of the signal-of-interest typically needs to be confirmed before executing geolocation routines. Distributed sensing is generally required for acquiring the data needed for emitter ID over wide coverage areas. Proliferated low-Earth orbit (pLEO) satellite constellations provide potential architectures for performing this distributed sensing. In conventional approaches, raw sensor data is aggregated at a centralized node in the sensor network, and emitter identification is executed using the aggregated dataset. However in instances where the data volume required for accurate emitter ID is large, limited communications bandwidth available for aggregating the required data can lead to high latency between data collection and emitter identification. This paper explores the viability of federated learning (FL) as an approach for reducing latency between the distributed sensing of an emitted waveform and its reliable identification. Our approach is validated via simulation in a scenario involving a pLEO satellite constellation which collects snapshots of digitally modulated RF waveforms transmitted from ground-based emitters. Channel impairments are modeled to be representative of those found in typical pLEO ground-to-satellite links. This data is used to train lightweight convolutional neural networks (CNNs) onboard each emulated pLEO satellite for emitter ID. The CNN model weights are shared among satellite nodes, and federated learning (FL) is used to fuse the DNNs into a global consensus model. We quantify the latency improvements gained from sharing model weights during FL relative to a baseline method which aggregates the raw data necessary to classify the transmissions. We find that a federated learning scheme can achieve nearly a 100-fold improvement over the baseline method. These findings indicate that federated learning coupled with satellite-based sensing is a potential enabler of low-latency emitter identification across wide surveillance regions. Ultimately, this capability could assist in the creation of RF spectrum maps using sensor networks limited in communications bandwidth.