Transfer Learning with Radio Frequency Signals
Scott Kuzdeba, Josh Robinson, Joseph Carmack
Abstract
Transfer learning has allowed for more widespread adaptation and expanded use of deep learning models in fields such as computer vision and speech recognition. The radio frequency (RF) domain has lagged behind and has yet to experience similar gains. Yet, at the same time, transfer learning is more relevant to tasks the RF domain. Few large labeled RF datasets exist and when they are available there is typically a difference between the dataset, i.e., the source data, and any target data application due to the myriad of variances that are introduced to RF signals, systems, and environments. Transfer learning focuses directly on this disparity between source and target data and changes in task, making it an ideal candidate to enable more widespread adaptation of deep learning within the RF domain. In this paper we explore several applications of transfer learning to RF signals using both synthetic and real-world communications signals. We focus on an RF fingerprinting task, using our RiftNet model as the base of our deep learning models. This paper presents approaches and results for RF transfer learning in the context of pre-trained models, re-tuning, unsupervised feature learning, and novel device detection. Moving beyond the methods and results presented in this paper, we hope this work motivates others to explore the potential and benefits of RF applications of transfer learning.