martFL: Enabling Utility-Driven Data Marketplace with a Robust and Verifiable Federated Learning Architecture

Abstract

The development of machine learning models requires a large amount of training data. Data marketplace is a critical platform to trade high-quality and private-domain data that is not publicly available on the Internet. However, as data privacy becomes increasingly important, directly exchanging raw data becomes inappropriate. Federated Learning (FL) is a distributed machine learning paradigm that exchanges data utilities (in form of local models or gradients) among multiple parties without directly sharing the original data. However, we recognize several key challenges in applying existing FL architectures to construct a data marketplace. (i) In existing FL architectures, the Data Acquirer (DA) cannot privately assess the quality of local models submitted by different Data Providers (DPs) prior to trading; (ii) The model aggregation protocols in existing FL designs cannot effectively exclude malicious DPs without “overfitting” to the DA's (possibly biased) root dataset; (iii) Prior FL designs lack a proper billing mechanism to enforce the DA to fairly allocate the reward according to contributions made by different DPs. To address above challenges, we propose martFL, the first federated learning architecture that is specifically designed to enable a secure utility-driven data marketplace. At a high level, martFL is empowered by two innovative designs: (i) a quality-aware model aggregation protocol that allows the DA to properly exclude local-quality or even poisonous local models from the aggregation, even if the DA's root dataset is biased; (ii) a verifiable data transaction protocol that enables the DA to prove, both succinctly and in zero-knowledge, that it has faithfully aggregated these local models according to the weights that the DA has committed to. This enables the DPs to unambiguously claim the rewards proportional to their weights/contributions. We implement a prototype of martFL and evaluate it extensively over various tasks. The results show that martFL can improve the model accuracy by up to 25% while saving up to 64% data acquisition cost.