The paper introduces FedOPAL, a framework designed to address the limitations of existing one-shot federated learning methods by combining analytical aggregation with visual prompt tuning. It adapts visual prompts as feature rectifiers to correct feature distribution misalignment in non-independent and identically distributed data environments.

  • Adapts visual prompts to actively correct heterogeneous data distributions into a linearly separable space using local proximal constraints.
  • Enables efficient gradient-free aggregation through least-squares closed-form solutions, eliminating the need for iterative fine-tuning or knowledge distillation.
  • Achieves zero server-side training costs while maintaining accuracy comparable to state-of-the-art iterative methods.
  • Significantly outperforms original analytical methods on several benchmarks.

FedOPAL provides a new engineering paradigm for efficient collaboration of large models on the edge by resolving the contradiction between communication bandwidth bottlenecks and static feature assumptions in analytical federated learning.