Disentangled Representation Learning for RF Fingerprint Extraction under Unknown Channel Statistics

Deep learning (DL) applied to a device’s radio-frequency fingerprint (RFF) has attracted significant attention in physical-layer authentication due its extraordinary classification performance. Conventional DL-RFF techniques are trained by adopting maximum likelihood estimation (MLE). Although their discriminability recently been extended unknown devices open-set scenarios, they still tend overfit the channel statistics embedded training dataset. This restricts practical applications as it is challenging collect sufficient data capturing characteristics of all possible wireless environments. To address this challenge, we propose DL framework disentangled representation (DR) that first learns factor signals into device-relevant component and device-irrelevant via adversarial learning. Then, shuffles these two parts within dataset for implicit augmentation, which imposes strong regularization on RFF extractor avoid overfitting statistics, without collecting additional from channels. Experiments validate proposed approach, referred DR-based RFF, outperforms conventional methods terms generalizability under complicated propagation environments, e.g., dispersive multipath fading channels, even though collected simple environment with dominated direct line-of-sight (LoS) paths.