arxivcs.CV2026-07-23
Towards Privacy-Preserving Federated Prompt Tuning under Data Heterogeneity: A Subspace-Decomposed Expert Approach
Yuhua Wang, Xiaodong Li, Yihao Guo, Yuxiang Jia, Qinnan Zhang, Yifan Sun, et al.
Federated prompt tuning (FPT) enables collaborative adaptation of vision--language models (VLMs) using lightweight prompts. Existing methods often address heterogeneity and privacy through a split-prompt design under local differential privacy (DP), combining a shared prompt for…