Learning Physically Consistent Spectrogram-Based Representations for Underwater Acoustic Communications with Privileged Information

This work proposes a physics-guided representation learning framework for underwater acoustic communications based on privileged information available only during training. Physical channel descriptors extracted from paired TX-RX cycles are encoded by a descriptor-space teacher and distilled into a model that operates solely on received-signal spectrograms, enabling blind RX-only inference. Experiments on the SODAC dataset, across receivers and frequency bands, show that the distilled representation preserves physically relevant channel structure and improves robustness under receiver shift. For downstream evaluation, unseen non-paired waveforms are assessed in a shared coarse signaling-family label space, avoiding ill-posed evaluation on disjoint raw waveform identities.