Federated Learning

Federated Learning in the Data Marketplace

This section explores how the Data Marketplace infrastructure could support federated learning (FL), a privacy-enhancing approach to collaborative model training.

Federated Learning Framework

The Data Marketplace architecture supports federated learning (FL) implementations, allowing participants to train machine learning models collaboratively while keeping raw data local, aligning with privacy preservation principles. The process follows a defined workflow coordinated through off-chain workers:

Dataset Selection

Participants access datasets via the tiered system, selecting relevant data (e.g., medical records) at Tier 3, with zk-SNARKs verifying eligibility through the verification infrastructure without revealing token balances.

Local Training

Participants train local models using a stochastic gradient descent (SGD) algorithm with a learning rate of 0.01 and batch size of 32, processing data on their nodes to generate gradient updates.

Secure Update Submission

Updates are encrypted with AES-256 (GCM mode) and submitted with zk-SNARK proofs of correctness, costing approximately 210,000 weight equivalent in testnet simulations.

Aggregation

An off-chain aggregator coordinated through off-chain workers (using secure multi-party computation, SMPC) computes a weighted average of updates, weighted by dataset size, and submits a single proof for on-chain verification.size, and submits a single proof for on-chain verification.

Dataset Selection

Local Training

Participants train local models using a stochastic gradient descent (SGD) algorithm with a learning rate of 0.01 and batch size of 32, processing data on their nodes to generate gradient updates.

Secure Update Submission

Updates are encrypted with AES-256 (GCM mode) and submitted with zk-SNARK proofs of correctness, costing approximately 210,000 weight equivalent in testnet simulations.