- https://arxiv.org/pdf/2009.09338.pdf
- Source: arxiv
TL;DR
This paper proposes replacing the trusted central server in federated learning with a blockchain aggregation layer using smart contracts, reputation-based client selection, and token incentives to achieve decentralized, Byzantine-tolerant model training.
Summary
Standard federated learning depends on a trusted central server — a single point of failure for trust and auditability. This paper replaces it with a blockchain layer where smart contracts automate model aggregation and record updates as immutable transactions. Reputation-based client selection filters Byzantine participants before aggregation, and a token incentive mechanism encourages honest gradient contributions. Evaluated on image classification, the system matches centralized FL accuracy while adding auditability and removing the central trust assumption.
Key contributions
- Proposes replacing the FL central server with a blockchain aggregation layer using smart contracts for coordination.
- Introduces reputation-based client selection as a Byzantine filtering mechanism before gradient aggregation.
- Demonstrates a token incentive system that sustains honest participation in open federated networks.
- Shows that decentralized aggregation achieves accuracy comparable to centralized FL on image classification benchmarks.
When to cite
- When proposing trustless or decentralized federated learning architectures for regulated or adversarial settings.
- When discussing smart contracts as transparent, auditable coordination mechanisms for ML aggregation.
- When evaluating incentive mechanisms for sustaining honest participation in open FL networks.
- When comparing reputation-based Byzantine filtering against cryptographic Byzantine defenses in FL.