Abstract
Machine learning's performance is fundamentally tied to the quality and diversity of its training data. However, high-quality data is often sequestered in private silos. While data marketplaces aim to facilitate the exchange of this information, they are plagued by the "buyer’s dilemma": buyers cannot verify data quality without access, and providers cannot grant access without guaranteed payment. This paper introduces martFL, a novel federated learning (FL) architecture designed to create a secure, fair, and robust data marketplace. By combining quality-aware aggregation with zero-knowledge proofs, martFL ensures that data providers are compensated based on the actual utility of their contributions while protecting privacy and mitigating the impact of biased buyer data.
Introduction
Traditional data marketplaces face significant hurdles regarding trust and verification. Federated Learning (FL) offers a potential solution by allowing models to be trained on local data without sharing raw datasets. However, existing FL frameworks are often ill-suited for a commercial marketplace for three primary reasons:
- Quality Assessment: Buyers lack a private method to assess model utility before purchase.
- The Bias Trap: Robust FL methods often rely on a buyer-provided "root dataset" to filter updates; if this root dataset is biased, the system may exclude high-quality, diverse data from providers.
- Verifiable Billing: There is rarely a trustless mechanism to prove that a buyer utilized a provider's update and compensated them fairly.
martFL addresses these challenges through a robust aggregation protocol and a verifiable transaction system.
Methodology
The martFL architecture is built upon two primary pillars designed to ensure both the technical integrity of the model and the economic fairness of the marketplace.
Quality-Aware Model Aggregation
To break the tradeoff between robustness and inclusiveness, martFL avoids relying solely on a static buyer-provided root dataset. Instead, it employs a two-phased approach:
- Hierarchical Clustering: The system uses Gap-Statistics and K-Means clustering to group provider updates by similarity. This allows the system to identify and filter out obvious outliers or malicious "poisoned" models without needing to compare them to a reference set immediately.
- Dynamic Baseline Adjustment: To prevent the buyer's initial data bias from skewing the results, martFL identifies "Preferred Data Providers" (DPs) in early rounds. The reference model is then dynamically updated using these high-quality contributions, allowing the system to "learn" what good data looks like beyond the buyer's initial scope.
Verifiable Transaction Protocol
To ensure the Data Acquirer (DA) acts honestly, martFL utilizes zk-SNARKs (specifically the Groth16 protocol). The DA generates a succinct, zero-knowledge proof demonstrating that they aggregated the global model according to the specified weights for each provider.
Because proving an entire model aggregation in a zk-circuit is computationally prohibitive, martFL introduces Verifiable Sampling. The DA only proves the aggregation for a randomly selected subset of model parameters. This provides high statistical confidence in the DA's honesty while keeping proof sizes small and blockchain gas costs minimal.
Privacy is further maintained through Homomorphic Encryption (CKKS), which allows the DA to score and aggregate model updates while they remain in an encrypted state.
Experimental Evaluation
The authors evaluated martFL across four diverse datasets: image classification (FMNIST, CIFAR-10) and text classification (TREC, AGNEWS). The experiments specifically simulated "biased" scenarios where the buyer's root dataset contained only a fraction of possible labels, testing the system's ability to incorporate diverse provider data.
Key Results
- Improved Accuracy: In scenarios with biased buyer data, martFL improved model accuracy by up to 25% compared to existing server-driven methods like FLTrust.
- Economic Efficiency: The system reduced data acquisition costs by up to 64% because the DA only pays for updates that provide measurable utility.
- System Robustness: martFL remained resilient against backdoor, Sybil, and free-rider attacks, maintaining high accuracy even when up to 80% of providers were malicious.
- Low Operational Overhead: On-chain verification costs were recorded at less than $0.0025 USD, demonstrating the feasibility of the sampling-based zk-SNARK approach for real-world blockchain deployment.
Conclusion
martFL represents a significant step forward for decentralized data marketplaces. By moving away from static, potentially biased reference sets and toward a dynamic, verifiable aggregation model, it solves the "buyer's dilemma" while maintaining high performance and low cost. Future work will explore trust-free zk-schemes to remove the need for a trusted setup and investigate more complex economic models, such as Shapley Values, for even more granular data pricing.
Summary updated