FedOPAL Boosts One-Shot Federated Learning Efficiency

The widespread deployment of foundational models in edge intelligence faces a significant challenge: communication bandwidth bottlenecks in federated learning. While one-shot federated learning aims to reduce communication rounds, existing methods still incur high server-side computational costs and are sensitive to hyperparameters. Traditional analytical federated learning offers gradient-free aggregation but struggles with non-independent and identically distributed (non-IID) data due to static feature assumptions, leading to feature manifold misalignment and poor performance. FedOPAL addresses this by adapting visual prompts as "feature rectifiers." This framework actively corrects the feature distribution of heterogeneous data into a linearly separable space using local proximal constraints, thereby satisfying the theoretical assumptions of analytical federated learning. Experimental results show FedOPAL significantly outperforms original analytical methods and achieves accuracy comparable to state-of-the-art iterative methods, all while maintaining zero server-side training costs. This presents a new paradigm for efficient large model collaboration at the edge.