# Towards On-Device Federated Learning: A Direct Acyclic Graph-based Blockchain Approach

**Authors:** Mingrui Cao, Long Zhang, Bin Cao  
**arXiv:** [2104.13092](https://arxiv.org/abs/2104.13092)  
**Date:** 2021-04-27  
**Summary page:** https://www.summarizepaper.com/en/paper/2104.13092v1/

## Key points

- The paper addresses challenges in distributed characteristics of Federated Learning (FL)
- Blockchain is seen as a promising solution for decentralization, scalability, and security in FL
- Traditional consensus mechanisms like Proof of Work (PoW) consume excessive resources, reducing FL efficiency
- The paper proposes DAG-FL, a framework using Direct Acyclic Graph-based Blockchain for FL
- DAG-FL overcomes device asynchrony and anomaly detection issues without additional resource consumption
- DAG-FL consists of a three-layer architecture with two algorithms operating on different nodes
- A Poisson process model is formulated to discuss deployment parameters for stable operation of DAG-FL
- Simulations and experiments show that DAG-FL outperforms existing on-device FL systems in training efficiency and model accuracy
- DAG-FL offers improved performance while maintaining decentralization, scalability, and security in FL systems.

## Summary

The paper titled "Towards On-Device Federated Learning: A Direct Acyclic Graph-based Blockchain Approach" addresses the challenges of distributed characteristics in Federated Learning (FL), such as vulnerability of the global model and coordination among devices. The use of blockchain in FL has gained attention as a promising solution for decentralization, scalability, and security. Traditional consensus mechanisms like Proof of Work (PoW) result in excessive resource consumption which significantly reduces FL efficiency, especially for wireless and resource-limited devices. To overcome device asynchrony and anomaly detection issues without adding extra resource consumption from blockchain, the paper proposes a framework called DAG-FL (Direct Acyclic Graph-based Blockchain for Federated Learning). DAG-FL is introduced through a detailed three-layer architecture with two algorithms designed to operate on different nodes that explain the functioning of the DAG-FL consensus mechanism. The paper also formulates a Poisson process model to discuss setting deployment parameters that ensure stable operation of DAG-FL across various federated learning tasks. Extensive simulations and experiments demonstrate that DAG-FL outperforms existing on-device federated learning systems in terms of training efficiency and model accuracy. Overall, this research presents an innovative approach to address the challenges faced by FL using blockchain technology. By leveraging a Direct Acyclic Graph-based blockchain framework, DAG-FL offers improved performance while maintaining decentralization, scalability, and security in federated learning systems.