Field Note #47. On "Federated Learning Précis"

Source: Dongkun Hou et. al., "A Systematic Literature Review of Blockchain-Based Federated Learning: Architectures, Applications and Issues" in 2nd Info. Comm. Tech. Conf. pp. 302-307 (May 2021)

The Thesis.

"Federated learning (FL) can realize a distributed training machine learning models in multiple deices while protecting their data privacy, but some defect still exists such as single point failure and lack of motivation.  Blockchain as a distributed ledger can be utilized to provide a novel FL framework to address those issues." (Id. at 302)

The Key Points.

4 Problems with Current Federated Learning Approaches (which Blockchain Architectures Can Address):

• Single point of failure
• Poison attack
• Lack of motivation (lack of voluntary data contribution)
• Privacy leakage

The Details.