Transfer learning (TL) techniques have proven useful in a wide variety of applications traditionally dominated by machine learning (ML), such as natural language processing, computer vision, and computer-aided design. Recent extrapolations of TL to the radio frequency (RF) domain are being used to increase the potential applicability of RFML algorithms, seeking to improve the portability of models for spectrum situational awareness and transmission source identification. Unlike most of the computer vision and natural language processing applications of TL, applications within the RF modality must contend with inherent hardware distortions and channel condition variations. This paper seeks to evaluate the feasibility and performance trade-offs when transferring learned behaviors from functional RFML classification algorithms, specifically those designed for automatic modulation classification (AMC) and specific emitter identification (SEI), between homogeneous radios of similar construction and quality and heterogeneous radios of different construction and quality. Results derived from both synthetic data and over-the-air experimental collection show promising performance benefits from the application of TL to the RFML algorithms of SEI and AMC.
Radio Frequency Machine Learning (RFML) has in recent years become a popular method for performing a variety of classification tasks on received signals. Among these tasks is Specific Emitter Identification (SEI), which seeks to associate a received signal with the physical emitter that transmitted it. Many different model architectures, including individual classifiers and ensemble methods, have proven their capabilities for producing high accuracy classification results when performing SEI. Though the works studying different model architectures report on successes, there is a notable absence regarding the examination of systemic failures and negative traits associated with learned behaviors. This work studies those failure patterns for a 64-radio SEI classification problem by isolating common patterns in incorrect classification results across multiple model architectures and two distinct control variables: Signal-to-Noise Ratio (SNR) and the quantity of training data utilized. This work finds that many of the RFML-based models devolve to selecting from amongst a small subset of classes (≈10% of classes) as SNRs decrease and that observed errors are reasonably consistent across different SEI models and architectures. Moreover, our results validate the expectation that ensemble models are generally less brittle, particularly at a low SNR, yet they appear not to be the highest-performing option at a high SNR.