Few-Shot Transfer Learning for Spectrum Awareness in Dynamic Propagation Environments

Deep learning models have emerged as the state of the art for spectrum sensing applications due to their expressivity, adaptivity, and robustness. In recent years, projects such as the IARPA SCISRS program [1] have continued to emphasize the necessity of highly adaptive algorithms for applications such as anomalous signal detection. To accurately model spectral behaviors in complex propagation environments, deep learning models typically require many labeled samples. However, large-scale over-the-air data collection and labeling is often prohibitively expensive, especially in adversarial or dynamic environments. As a result, developing methods to reduce the number of labeled samples needed to train robust deep learning models is a critical research area. Transfer learning, one such approach that reduces the amount of task-specific data necessary by pretraining models on massive external datasets to capture task-generic information, has been shown to be successful in environmental adaptation tasks [2] . Furthermore, prior work [3] has demonstrated the effectiveness of using simulated data to augment spectrum awareness. We hypothesize that pretraining deep neural networks using simulated data will enable the networks to learn high-level structural characteristics of communications waveforms, improving their performance on downstream few-shot over-the-air spectrum sensing tasks. As visualized in Figure 1 , we consider the few-shot spectrum sensing problem in dynamic environments where few labeled samples are available. By using simulated examples of known representative waveforms to pretrain deep neural networks, we hope to achieve more robust spectrum sensing performance on data collected from previously unseen transmitters. To the best of our knowledge, there is limited research making use of simulated data to improve model performance on spectrum sensing tasks in few-shot over-the-air dynamic propagation environments.