Federated Deep Unfolding for Compressive Sensing

The deep unfolding-based compressive sensing (CS) method has garnered significant success in recent years due to its interpretability and convergence guarantees. However, existing methods are centralized, which does not align with the decentralized and privacy-sensitive nature of real-world data. To address these issues, this paper proposes federated deep unfolding (FDU) for CS. We observed that, given the semi-data-driven and semi-model-driven nature of deep unfolding networks, the trained model parameters exhibit minimal differences even when trained in a federated learning setting. Experiments on datasets Set11, Set5, and CBSD68 confirms that the proposed FDU can achieve results comparable to, and occasionally slightly superior to, centralized approaches.