End-to-End Learning for Length-Invariant Radio Identification

Radio Frequency Fingerprint Identification (RFFID) has rapidly become an advanced technique for identifying wireless devices, exploiting devices’ unique signal characteristics. This key technology promises a wide range of applications, from wireless security to sensing by differentiating signals of various devices. Despite their potential, previous Machine Learning methods, including Convolutional Neural Networks (CNNs), struggle with the varying length nature of RF signals due to their fixed-size input requirements. These models also fail to capture sequential relationships in signals, which impacts model performance. Moreover, some approaches require signal preprocessing, increasing computational overhead. This paper proposes an end-to-end hybrid model combining CNN and Long Short-Term Memory (LSTM) networks. Our approach extracts spatial features from RF signals while capturing sequential dependencies and accommodating variable signal lengths. We validate our method using RF fingerprints collected from 20 devices with a signal-to-noise ratio (SNR) of approximately 20 dB. The proposed model achieves 91.20% accuracy, surpassing current state-of-the-art performance by an average point of 20.57% across various SNR conditions while maintaining model size with fewer parameters.