Fed-SRC: Private Anytime Selective-Risk Certification for Federated RAG
A team of researchers has come up with an innovative method called Fed-SRC, designed to offer selective-risk certificates for federated retrieval-augmented generation (RAG) while ensuring differential privacy. This certification framework, which is score-agnostic, is explained in an arXiv paper (2608.07913) and guarantees that the outputs meet a certain error threshold. To keep data private, clients only need to share Gaussian-perturbed score and loss histograms. The method uses specialized martingales to control target-risk contrast and accepted mass over different thresholds and rounds. Impressively, during testing, no violations of simultaneous bounds were found across any cells, privacy levels, or policies, with operational effectiveness depending on the score and populace, aiming for r*=0.
Key facts
- Fed-SRC is a score-agnostic certificate for federated, differentially private, adaptively monitored retrieval-augmented generation.
- Clients release only Gaussian-perturbed score and loss histograms.
- Record-indexed and noise-variance-indexed martingales jointly bound target-risk contrast and accepted mass.
- The method permits predictable recruitment, dropout, threshold selection, and optional stopping.
- A range-one total-variation term transfers the calibration mixture to a declared deployment mixture.
- Empirically, no simultaneous-bound violation occurs in any evaluated cell, privacy level, or policy.
- The primary target r*=0.
- The paper is available on arXiv with identifier 2608.07913.
Entities
Institutions
- arXiv