• Proposes a frequency-learning adversarial network (FLAN) framework that leverages frequency-domain stability for robust cross-scenario signal modulation classification.
• Introduces frequency adaptation (FA) and fitting channel adaptation (FCA) modules to effectively reduce domain shift caused by channel variations.
• Achieves approximately 5.2 percentage points improvement in top-1 classification accuracy over state-of-the-art transfer learning methods in high SNR conditions.
• Validates the method on a real collected dataset (CSRC2023), demonstrating practical applicability in wireless communication systems.