Research On Privacy Protection Of Multi-Source Data Based On Differential Privacy

Privacy protection of multi-source data is an important research content in the field of information security.The traditional method is to integrate multi-party data to train the model,which ignores the security of data and consumes a lot of costs and resources.Therefore,the research on privacy protection of multi-source data has important practical significance.Federated learning is an important technology to use when addressing the decentralized data,which combines privacy protection models with different ensemble methods and applies it to the framework of federated learning,which puts forward a new multi-source data privacy protection algorithm on the premise of ensuring the accuracy of the model to improve the security of data,the timeliness and the security of the model as the main goal.The main contributions are as follows:1.In order to solve the problem that traditional integration of multi-party data retraining models may increase the risk of data leakage,a federated ensemble algorithm is proposed.The algorithm uses different ensemble algorithms to integrate privacy protection models on multi-source data,and improves the security of data and model within the acceptable range of accuracy reduction.2.In view of the traditional data processing technology is outdated,and does not take into account the data and the model of security issues,an online federated incremental learning algorithm based on blockchain is proposed,which extends the federated ensemble algorithm to each time period,constructs an online federated incremental learning model,and uploads the model parameters in each time period to the blockchain.It improves the timeliness of the model,and further improves the security of data and model.3.Aiming at the problem that multi-source dynamic model may causes data leakage when updating model parameters,a differential optimization federated incremental learning algorithm based on blockchain is proposed.It adds differential privacy technology and optimizes model parameters in the process of training local model for each data source,so that the security of data and model is improved.Besides,the accuracy of the model is also improved.The security of multi-source data and the applicability of the model cannot be satisfied at the same time,but the combination of the two is more practical.Combined with federated learning,differential privacy and blockchain to build a multi-source privacy protection model,and conducted experiments on open data sets.The experimental results prove the feasibility and effectiveness of the model,which provides a new research idea for the construction of multi-source data privacy protection model.Figure 120;Table 26;Reference 72...