Towards Privacy-Preserving Light Field Super-Resolution: A Federated Learning Approach

Light field super-resolution is essential for enhancing the utility of commercial plenoptic cameras. However, centralized training requires large-scale data sharing, which is hindered by strict privacy regulations and the inherent heterogeneity of devices. Federated learning (FL) offers a privacy-preserving alternative, yet its effectiveness in LFSR remains underexplored.In this work, we propose a federated paradigm for light field super-resolution by extending representative deep learning models into their federated counterparts, namely Fed_EDSR, Fed_DistgSSR, and Fed_DPT. To better accommodate the 4D structural properties and device heterogeneity of light field data, we further incorporate geometric consistency constraints and hierarchical aggregation strategies.Extensive experiments across multiple benchmarks and upscaling factors demonstrate that federated models consistently improve reconstruction quality. Notably, FL reshapes the performance landscape by elevating weaker models and narrowing the performance gap, highlighting its capability to exploit distributed data diversity. To the best of our knowledge, this is the first systematic validation of FL for LFSR, offering a practical solution for privacy-aware collaborative processing of high-dimensional visual data.