FedSTIP-Tuning: Federated Black-Box Prompt Tuning with Aggregate-Only Feedback and Common-Candidate Optimization

Federated black-box prompt tuning enables organizations to adapt a proprietary frozen language model without centralizing their labeled data, but local search-state aggregation does not directly define a shared pooled-data objective, and client-specific feedback can expose private signals. We propose FedSTIP-Tuning, which composes a three-party permutation-based protected forward propagation with a common-candidate population and pairwise-masked loss aggregation in a cross-silo setting. All participating clients evaluate the same prompt candidates, and the coordinator ranks their sample-weighted aggregate losses, thereby optimizing the empirical objective over the fixed union of client shards under full participation. Under the stated semi-honest and non-collusion assumptions, the aggregation layer hides each released loss vector’s client-level decomposition, while the aggregate values and execution metadata remain observable to their authorized parties. On RoBERTa-large, FedSTIP-Tuning exceeds the strongest evaluated FedBPT configuration by 6.40, 2.66, and 7.60 percentage points on AG News, SST-2, and DBpedia, respectively, under matched forward budgets. Its protected forward preserves all candidate rankings across 600 paired generations and is 861× and 957× faster than a generic interactive two-party computation reference on the RoBERTa-base warm and cold paths, and 694× and 578× on RoBERTa-large.