Architecture-Agnostic Wavelet Compression in Clustered Federated Learning for SNR-Heterogeneous Automatic Modulation Recognition

Federated learning (FL) for SNR-heterogeneous automatic modulation recognition (AMR) is challenging because a single global model can be poorly matched to diverse client data distributions, particularly for lightweight networks. FedCAMR is presented as a one-shot clustered FL framework that groups clients using training dynamics and then trains cluster-specific models. An architecture-agnostic wavelet transform (WT) module, implemented via learnable 1-D convolutions, is further introduced to compress the downstream layer sizes. Experiments on a modified SNR-heterogeneous subset of RadioML 2016.10b show consistent gains over the conventional FL baseline, with the most significant improvements observed in low-SNR regimes and ultimately attain a performance close to the ideal centralized training with less than 1% accuracy gap. For WT-aided deep neural networks, such as convolutional networks, communication overhead and inference time are reduced by up to 99%, while achieving competitive AMR performance.