Research On Privacy Preservation Of Realiable Deep Learning On Blockchain

In recent years,with the advent of the era of big data,deep learning is not only adopted by computer science,but also widely used in the fields of Internet of Things Engineering,medical health and education.In these areas,Enterprises have high demand for collaborative learning models,and many enterprise users collaborate to train common models by providing their own data.Therefore,business needs in actual scenarios are more inclined to deep learning models that can obtain valuable feature data from the large scale distributed data owned by each participant in each region.However,the distributed deep learning system in the existing research still has many problems to be solved urgently in the scene of large-scale user participation,such as user data privacy,node credibility problems,and model evaluation problems.Focusing on the reliable framework and security protocol of distributed deep learning based on blockchain,our work studies the data sharing and model training of deep learning,and designs different security solutions according to different application requirements.The main research content of this paper includes the following aspects:Aiming at the privacy and security issues of users participating in collaborative training in blockchain-based deep learning platform,this paper proposes a decentralized,secure and efficient deep learning algorithm in a mobile scenario,so that participants can use the blockchain and smart contract custody and continuously updated model.In this framework,we use encryption weights based on training errors to construct a dynamic weighting algorithm to realize the sharing and aggregation of model parameters,and protect the privacy of local training data without sacrificing the accuracy of the aggregation model.Through the theory analysis and computer simulations,it is proved that the framework guarantees the security of data and the effectiveness of the algorithm,and reduces the transaction cost and scale of model parameters.The fusion algorithm of artificial intelligence introduces and blockchain introduces the advantages of deep learning to application areas where data owners cannot share their data due to data machine privacy issues.Aiming at the problem of node reliability and model evaluation in the process of data sharing in the blockchain-based deep learning framework,we propose a Validated Deep Learning Blockchain(VDLChain)based on blockchain,and use the results of decentralized evaluation to design an incentive mechanism based on the quality of model verification.Through theoretical analysis and experimentation,VDLChain promotes the reliability of users,and reduces the influence of irregular nodes on the aggregation model under the condition of ensuring the privacy and precision of deep learning models.Aiming at the reliable model aggregation of distributed deep learning in the process on the blockchain,we propose a distributed deep learning algorithm based on an efficient consensus mechanism.By improving the byzantine fault-tolerant consensus algorithm based on the voting mechanism,the algorithm improves the efficiency of model aggregation,effectively reduces the degree of centralization of the system,and realizes reliable model aggregation.Secondly,the algorithm constructs the Krum algorithm based on differential privacy during the model aggregation calculation process,and calculates the aggregation model in a clustering manner for the uploaded encryption parameters to ensure privacy security and algorithm convergence.Through the theory analysis and computer simulations,the algorithm can still effectively resist attacks from malicious nodes,ensure the security of the algorithm.At the same time,the algorithm reduces the communication complexity of model aggregation and improves the efficiency of the algorithm.