FedMark-FM：可审计、风险调整的数据市场用于联邦基础模型适配

Jul 16, 2026

Upload by Miles

Federated foundation-model adaptation depends on heterogeneous private artifacts such as retrieval corpora, prompts and demonstrations, LoRA adapters, and preference/safety data, yet common federated-learning incentives treat contributors as homogeneous. FedMark-FM proposes an auditable, risk-adjusted data-market framework that models clients as sellers of typed artifacts, estimates marginal value with S3Val secure surrogate Shapley valuation, and converts lower-confidence bounds into budget-feasible payments. The mechanism penalizes duplication, sybil splitting, poisoned adapters, privacy-budget gaming, and cost inflation, and introduces pipeline-ordered valuation.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages24

Reading Time1 h 0 min

Introduction

FedFM as an Economic Problem

Related Work and Gaps

Four Required Market Properties

Proposed FedMark-FM Framework

FedMark-FM: Auditable, Risk-Adjusted Data Markets for Federated Foundation-Model Adaptation Phat T. Tran-Truong 1,2, * , Xuan-Bach Le 1,2,†, and Minh Nhat Nguyen3 ~~ ~~ ✦ ~~ ~~ arXiv :2607 .07529v 1 [ cs .GT] 8 Jul 2026 Abstract—Federated foundation-model adaptation increasingly relies on heterogeneous private artifacts—retrieval corpora, prompts and demonstrations, LoRA adapters, preference and safety data, and update sketches—yet existing federated-learning incentive mechanisms price clients as homogeneous data or update providers. This assumption is poorly matched to foundation-model pipelines, where contribution value is heterogeneous, non-IID, pipeline-dependent, privacy-constrained, and vulnerable to strategic behavior. We propose FedMark-FM (a Federated Market for Foundation Models), an auditable, risk-adjusted data-market framework that models clients as sellers of typed artifacts, estimates marginal contribution with S3Val (Secure Surrogate Shapley Valuation)—a stratified, uncertainty-aware Shapley estimator that also supports pipeline-ordered valuation—and converts lowerconfidence-bound values into budget-feasible payments that penalize duplication, sybil splitting, poisoned adapters, privacy-budget gaming, and cost inflation. We evaluate FedMark-FM-Bench (the FedMark-FM Benchmark) across FEVER retrieval, held-out generator-backed RAG, and trained PEFT/LoRA tracks. Under a held-out prompt-injection poisoner, FedMark-FM improves downstream accuracy by 7.5–8.1 points over volume, leave-one-out, and FL-Shapley while selecting zero strategic clients. Split-conformal calibration reaches full lower-bound coverage at mean width 0.0141, versus 0.33 for naive intervals. We prove that pipeline-ordered valuation is the unique credit rule respecting serving causality, and show it materially changes credit assignment (Spearman 0.76, selected-set overlap 0.67) while leaving held-out task quality unchanged; the market also preserves rare specialists with audit-ready ledgers at 200–1000-client scale. FedMark-FM shows that incentives for federated foundation models can be engineered as auditable data infrastructure that couples valuation, mechanism design, privacy interfaces, and pipeline-order semantics. Index Terms—Federated foundation models, data markets, data valuation, incentive mechanisms, retrieval-augmented generation, LoRA adapters, data-centric AI. 1 INTRODUCTION Foundation models are increasingly deployed as data pipelines rather than monolithic networks. A production 1Faculty of Computer Science and Engineering, Ho Chi Minh City University of Technology (HCMUT), 268 Ly Thuong Kiet Street, Dien Hong Ward, Ho Chi Minh City, Vietnam. 2 Vietnam National University Ho Chi Minh City, Linh Xuan Ward, Ho Chi Minh City, Vietnam. 3RMIT University, Ho Chi Minh City, Vietnam. E-mail: [phatttt@hcmut.edu.vn](phatttt@hcmut.edu.vn); [lexuanbach@hcmut.edu.vn](lexuanbach@hcmut.edu.vn); [minh.nguyen244@rmit.edu.vn](minh.nguyen244@rmit.edu.vn). *First author. †Corresponding author. assistant, search stack, or enterprise copilot answers a query by composing retrieved passages, prompt templates, incontext demonstrations, parameter-efficient adapters, preference data, and safety probes, and these artifacts are refreshed continuously after deployment. In cross-silo settings they are owned by different organizations—hospitals, data vendors, safety labs—that cannot or will not centralize raw data. Realizing federated foundation models (FedFMs) at this scale is therefore as much an economic problem as an algorithmic one: a shared pipeline improves only if the owners of these private artifacts are paid for their contribution, and they participate only if that contribution is measured and rewarded fairly, at scale, and with evidence they can dispute. Two research lines bear on this problem. Federatedlearning (FL) incentive mechanisms allocate rewards for participation and risk sharing, typically through Shapleybased o

FAQs

为什么现有联邦学习激励机制难以适配联邦基础模型适配？

现有方法通常把客户端视为提供同质数据或更新者，但基础模型管线中贡献价值是异构、非IID、受隐私约束且容易被策略性行为操纵，因此简单的同质定价假设不匹配。

FedMark-FM 的核心机制是什么？

FedMark-FM 把客户端建模为“带类型的制品”卖家，并用 S3Val（安全替身的分层不确定性 Shapley 估计）估算边际贡献；再将下置信界转换为满足预算约束的支付，从而惩罚重复、Sybil 分裂、投毒适配器、隐私预算博弈和成本膨胀。

FedMark-FM 如何确保信用分配与服务顺序一致？

文中指出信用应沿“检索→提示→适配→安全”等可见服务因果顺序分配，并证明这种 pipeline-ordered valuation 是与服务因果一致的唯一 credit rule。

Jul 16, 2026

Upload by Miles