Federated Deblurring as a Jigsaw Puzzle: A Privacy-Preserving Consensus Framework

Xuesong Chen, Liyan Ma, Fang Li

Keywords: Federated Optimization, Image Inverse Problems, Plug-and-Play (PnP) Priors, Privacy-Preserving

Abstract: Multi-observation deblurring seeks to recover a clear image from multiple blurred observations, yet most methods assume centralized access to full-scene data, which is unrealistic in practical scenarios where images are captured by distributed, independent devices. We introduce a more realistic federated setting where each client holds a private, partially overlapping view of a larger scene. This creates an “information jigsaw puzzle” that must be solved under strict privacy and regulatory constraints, without sharing raw images, kernels, or intermediate image estimates. We propose FedDeblur, a principled federated optimization framework based on consensus that decouples client-side local data fidelity updates and a server-side global image prior update. Clients transmit only carefully designed, desensitized variables, while the server coordinates global consensus. The modularity of our framework enables the server to flexibly incorporate diverse image priors, bridging classic regularizers like total variation with modern deep plug-and-play (PnP) denoisers, all transparently to clients. Crucially, all client-side updates admit efficient closed-form solutions, eliminating the need for inner iterations and making our framework practical for resource-constrained edge devices. Experiments demonstrate that FedDeblur seamlessly integrates fragmented information from partial views, effectively solving the jigsaw puzzle and achieving performance close to an idealized, non-private centralized oracle.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 7450