Synchronization-Aware Diffusion Models for Intra-Family RF Signal Classification

Abstract

Classification of radio frequency (RF) signals in the presence of channel-induced synchronization errors remains a critical challenge in spectrum awareness systems. Traditional classification pipelines generally rely on fixed synchronization algorithms or assume aligned signals, which limits robustness under real world timing, phase, and frequency distortions. We introduce SyncDiff, a novel encoder-only diffusion model architecture that predicts synchronization parameters through iterative denoising steps prior to classification. By replacing conventional synchronization algorithms with a learned datadriven correction mechanism, our approach enables adaptive signal alignment based on current channel distortions in unsynchronized input data. SyncDiff employs a UNet based encoder to refine synchronization parameters across multiple inference steps, dynamically reducing channel-induced alignment errors while preserving the inherit modulation specific characteristics that allow these signals to be discriminable. Evaluations of the RadioML2018 RF standard benchmark data set [1] demonstrates improved classification accuracy across varying SNRs, modulation schemes and synchronization impairments. Our findings highlight the potential of diffusion-based synchronization learning to improve downstream RF classification without reliance on expert-engineered synchronization routines.