Radio-frequency (RF) signal processing has seen significant advancements with the advent of deep learning, providing more accurate and efficient solutions for tasks such as signal classification and generation. However, most existing methods are heavily dependent on large labeled datasets, which are often scarce and costly to obtain in real-world RF environments. Furthermore, these approaches tend to be task-specific, limiting their ability to generalize across various RF applications. To address these challenges, this paper proposes RF-MAE, a self-supervised adaptive frequency masked autoencoder. RF-MAE leverages self-supervised learning (SSL) to capture intrinsic patterns from large-scale unlabeled RF data. Central to RF-MAE is a novel Adaptive Frequency Masked (AFM) strategy, which dynamically masks frequency components based on their energy distribution. Supported by a robust theoretical foundation, AFM ensures the model focuses on the most informative signal components, thereby enhancing generalization across RF tasks. By pretraining on unlabeled data and fine-tuning on specific tasks, RF-MAE significantly reduces the reliance on labeled datasets while improving adaptability across diverse RF signal processing tasks. Experimental results demonstrate that RF-MAE consistently outperforms traditional models, underscoring its potential to generalize across tasks and deliver superior performance in a wide range of RF signal applications.

Radio Frequency Fingerprinting Identification (RFFI) has become a critical technology in the physical-layer security (PLS) field, with deep learning emerging as the dominant approach over the past decade. However, most deep learning-based models rely on the assumption that training and testing data follow an independent and identical distribution (i.i.d.), which often does not hold in real-world scenarios. This mismatch significantly degrades model performance in cross-domain settings, making cross-domain RFFI a challenging task. Traditional unsupervised domain adaptation (UDA) methods attempt to address this issue by jointly optimizing task loss and domain loss which is able to reduce the distribution gap between training and testing data. However, we observe that during training, the gradients of these two losses often conflict, hindering effective optimization and limiting cross-domain performance improvements. To address these challenges, we propose a novel framework, Gradient Collaborative Optimization with Dynamic Weighted Feature Alignment (GCODWFA). Specifically, GCODWFA introduces a novel Gradient Collaborative Optimization (GCO) loss, which explicitly adjusts the gradient interaction between task and domain losses by optimizing their angular relationship. Additionally, it incorporates a Dynamic Weighted Feature Alignment (DWFA) strategy, which dynamically adjusts the layer-specific weights for feature alignment based on the angular similarity of task and domain gradients. Extensive experiments conducted on multiple datasets demonstrate the superiority of GCODWFA over existing methods.