Design And Implementation Of An Edge-Oriented Decentralized And Trusted Asynchronous Federated Learning Tool

With the increasing deployment of IoT,a large number of sensors and edge devices are collecting,storing and accumulating massive amounts of data at the edge side.Edge computing and federated learning provide an effective solution for edge devices to collaboratively train artificial intelligence models while protecting data privacy.However,traditional federated learning methods face challenges when applied to edge networks,such as the risk of a single point of failure caused by central servers and difficulty in ensuring data consistency and reliability in low-trust edge environments.Additionally,attackers can hijack devices to corrupt model aggregation due to weak security defenses.To address these challenges and ensure the reliability of edge devices for privacy computing at the edge of the network,we propose a blockchain-based federated learning approach.Specifically,we introduce an edge-oriented decentralized trusted asynchronous federated learning algorithm based on leaderless voting consensus and decentralized asynchronous federated learning.The algorithm ensures robustness and efficiency by identifying Byzantine nodes through node weight adjustment in the P2P network based on confidence.Moreover,we design a tool using distributed microservices and techniques such as event-centric,hot-synchronization and resource pooling to build a resilient and asynchronous decentralized federated learning network based on the proposed algorithm.The full paper is specified as follows.(1)To address the issues of low efficiency,high resource consumption,high synchronization delay,and unguaranteed model correctness in current blockchain-based federated learning methods,a novel decentralized asynchronous federated learning algorithm is proposed.This algorithm is based on leaderless state voting consensus and comprises four essential components.Firstly,it employs a DAG(Directed Acyclic Graph)based ledger and a confidence-based leaderless voting consensus mechanism to confirm transaction status via state voting.Additionally,the transactions are transmitted to the entire blockchain network through continuous random sampling,which effectively reduces communication overhead while ensuring data consistency.These operations facilitate highperformance and reliable transaction verification.Secondly,the algorithm quantifies node confidence and regulates the weight of blockchain nodes and generative models in consensus and aggregation,respectively.Furthermore,it minimizes the impact of Byzantine nodes.Thirdly,the algorithm implements a leaderless,decentralized asynchronous federated learning approach based on a local DAG ledger view built on top of the lightweight blockchain architecture.Lastly,the algorithm constructs a Byzantine node identification policy based on confidence and voting consensus.This policy consists of discovery,confirmation,voting,expulsion,or release,which effectively prevents Byzantine nodes from participating in blockchain activities.(2)Based on the proposed algorithm,we have designed and implemented an edge-oriented decentralized trusted asynchronous federation learning tool,which follows a loosely coupled modular concept.The service interface is described using IDL(Interface Description Language)and provides remote procedure invocation capability to the client through the microservice server.In addition,the tool provides blockchain consensus service based on an event center and has designed functionally isolated event processors for different message responses to achieve real-time response for multiple consensus requests.To enhance the availability of the ledger service,the tool stores semi-structured ledger data based on a document-based database and utilizes a hot-synchronization mechanism to complete data synchronization without suspending node services.The federated learning process is divided into three basic atomic modules:aggregation,training,and testing,which are completed through multi-module linkage.Furthermore,the tool utilizes caching and pooling techniques to improve resource reuse.In terms of functionality,the tool is equipped with trusted asynchronous federated learning and Byzantine fault tolerance.It provides fast networking,ledger synchronization,and multitasking.Additionally,the tool supports rapid deployment solutions and WEB clients.