1: \begin{abstract}

2: When personalized federated learning (FL) meets large foundation models, new challenges arise from various limitations in resources.

3: In addition to typical limitations such as data, computation, and communication costs, access to the models is also often limited.

4: This paper endeavors to solve both the challenges of \textit{limited resources} and \textit{personalization}. i.e., distribution shifts between clients.

5: To do so, we propose a method named \textbf{\method} that uses

6: \textbf{Z}eroth-\textbf{O}rder \textbf{O}ptimization

7: for \textbf{P}ersonalized \textbf{F}ederated \textbf{L}earning.

8: \method avoids direct interference with the foundation models and instead learns to adapt its inputs through zeroth-order optimization.

9: In addition, we employ simple yet effective linear projections to remap its predictions for personalization.

10: To reduce the computation costs and enhance personalization, we propose input surgery to incorporate an auto-encoder with low-dimensional and client-specific embeddings.

11: We provide theoretical support for \method to analyze its convergence.

12: Extensive empirical experiments on computer vision and natural language processing tasks using popular foundation models demonstrate its effectiveness for FL on black-box foundation models.

13: \end{abstract}

14: