Cross-Chain Consensus for Clustered Federated Learning Updates
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing blockchain-based federated learning (BFL) systems face challenges with low consensus and learning efficiency when the number of participants is large, leading to high communication overheads and privacy risks due to the nature of the BFL structure.
Innovation Solution
Implement a cross-chain consensus mechanism that splits computing nodes into multiple small clusters within the same organization, conducting intra-cluster single-chain BFL and collecting local update information, followed by cross-cluster gradient exchange and verification to ensure data privacy and model integrity.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If blockchain-based federated learning is used to ensure decentralized control and privacy protection, then system reliability and security are improved, but communication overhead increases and learning efficiency decreases
Solution Approach 1:
The patent segments the federated learning system into multiple clusters, each with its own blockchain network. This segmentation reduces the communication scope from all nodes to only within-cluster nodes, thereby maintaining the reliability benefits of blockchain while reducing communication overhead and improving learning efficiency.
Solution Approach 2:
The patent introduces cluster heads as intermediary nodes that aggregate gradients from member nodes before cross-cluster communication. This intermediary mechanism reduces the number of direct blockchain interactions required, lowering communication overhead while maintaining system security and reliability.
2Reliability
If all nodes participate in consensus for every model update, then decentralization and security are maintained, but communication overhead increases significantly
Solution Approach 1:
The patent divides the network into clusters where only cluster heads participate in cross-cluster consensus. This segmentation reduces the number of nodes involved in consensus from all nodes to just the cluster heads, significantly reducing communication overhead while maintaining decentralization through the distributed cluster structure.
Solution Approach 2:
The patent implements partial consensus by having only cluster heads participate in cross-cluster model update consensus, rather than requiring all nodes to participate. This partial action approach maintains the essential security and decentralization benefits while dramatically reducing communication overhead.
3Productivity
If cross-cluster gradient exchange is implemented to improve learning efficiency, then model convergence speed increases, but privacy protection becomes more challenging
Solution Approach 1:
The patent uses cluster heads as intermediaries that aggregate and anonymize gradient information before cross-cluster exchange. This intermediary mechanism allows efficient cross-cluster learning while protecting individual node privacy by preventing direct observation of raw gradient data from specific nodes.
Solution Approach 2:
The patent implements different privacy protection mechanisms for different levels of the system: strong privacy protection within clusters using blockchain immutability, and aggregated/anonymous sharing between clusters. This local quality approach optimizes privacy protection for each level while enabling efficient cross-cluster learning.
Data Source
AI summary
The present invention relates a method and a system for cross-chain consensus oriented to federated learning, comprising: conducting intra-cluster single-chain federated learning and collecting local update information; sending updates after consensus to a second federation so as to execute cross-cluster gradient exchange; receiving a verification result of cross-cluster gradient update consensus fed back from the second federation; and conducting local model update based on the verification result. After implementation of the update consensus, the present invention provides rewards and punishments based on the contributions of the cluster representatives, thereby encouraging the cluster representatives in the computing nodes to act honestly, so that the participants can actively help the model update.