Research On Dynamic Differential Privacy Federated Learning Scheme Of Intermediate Parameters Based On Blockchain

With the continuous development of artificial intelligence technology,it has gradually penetrated into people’s production and life.The core of artificial intelligence is machine learning,and Deep learning is an important branch of machine learning that uses deep neural networks to implement various artificial intelligence functions.Deep learning builds deep neural networks to automatically learn abstract features from large amounts of data,and then shines in multiple fields such as image recognition,natural language processing and speech recognition.The problems brought by the rapid development of deep learning are also becoming more prominent.Among them,protecting personal data privacy has become an urgent and important task in the context of the continuous increase of massive data sets.Federated learning,as a framework for protecting data privacy and improving machine learning privacy,has received widespread attention.This paper analyzes the basic concepts and processes of federated learning,summarizes the sources of federated learning vulnerabilities as communication protocols,user data and central aggregation services,focuses on the privacy security threats existing in existing federated learning algorithms,including model poisoning attacks,parameter inference attacks and intermediate parameter aggregation server attacks,etc.,followed by the solution of this paper for solving privacy security problems.Differential privacy is a technique that can perform data analysis and sharing without exposing personal sensitive information.Its basic idea is to add a certain amount of noise to the data.The added random disturbance has a very small impact on the results,thus achieving the purpose of protecting privacy.Blockchain technology is a new computing model based on distributed networks and cryptographic technology and is widely used in decentralized,secure and highly transparent application scenarios.Its basic concept is to store data in multiple nodes in a decentralized manner to achieve data exchange and consensus through decentralization.Each blockchain node has the same data copy and uses encryption algorithms to ensure the integrity and immutability of the data while using consensus algorithms to ensure its consistency and credibility.Due to its advantages such as decentralization and high security,this paper proposes a blockchain-based federated learning model for the characteristics of federated learning and existing privacy security risks combined with differential privacy and blockchain technology.This paper presents the main research topics and novel contributions as follows:(1)An adaptive intermediate parameter gradient clipping mechanism is proposed to address the slow local training problem in federated learning.The gradient parameters of the previous round of training are used to perform gradient clipping on the parameters of this round to reduce model training time.(2)We propose a differential privacy method with weighted noise addition based on the Gaussian mechanism,which enhances the differential privacy technique against parameter inference attacks in federated learning.Using the Gaussian weighting method,we assign data privacy weights to local participants prior to training,which preserves the privacy of participants and also mitigates the noise effect on model accuracy.(3)A random K-N scheduling federated aggregation mechanism is proposed to address the slow convergence time of models caused by traditional federated learning aggregation processes requiring all participants to participate in aggregation.The federated learning aggregation method is improved and random K-N federal scheduling aggregation is adopted to reduce model training time.(4)A blockchain-based privacy-secure federated learning model system is designed and implemented.Federated learning still has problems such as model poisoning attacks and central aggregation server attacks.The blockchain system replaces the traditional central aggregation server in federated learning and designs a blockchain system suitable for this chapter including block design transaction data design smart contract design and network consensus mechanism selection etc.The characteristics of blockchain ensure that the entire federated learning model’s privacy security model security and fairness are further guaranteed.