SkyLLM: Enabling Trustworthy UAV RF Surveillance With Foundation Models for Open-World Signal Recognition

UAV RF surveillance is becoming a critical layer of low-altitude airspace security, yet most fingerprinting pipelines still assume that every signal belongs to a known class. In open-world deployments, where new platforms, protocol updates, and changing propagation conditions are routine, that assumption makes conventional classifiers brittle and often overconfident on unseen emitters. Foundation models offer a compelling alternative because their pretrained priors can support richer sequence understanding beyond task-specific training. We present SkyLLM, a foundation-model-driven framework that aligns RF time-series with the token space of a frozen LLM through patch-based embedding, cross-channel fusion, and prompt conditioning, so that waveform fragments can be interpreted in a representation space better suited to novelty-aware decision making. Experiments on real UAV RF data show that SkyLLM preserves strong discrimination while delivering markedly more reliable unknown rejection than lightweight baselines across increasingly open settings. More broadly, the study suggests that future trusted spectrum intelligence will depend not only on better classifiers, but on representation layers that remain stable as the airspace evolves. SkyLLM therefore offers a practical path toward 6G-era UAV monitoring systems that must combine recognition, uncertainty awareness, and operational trust.