| With the rapid development of the Internet and big data technology,the need for data sharing among different institutions or organizations for collaboration or cooperative processing has emerged.However,the processing of distributed data involves coordination and collaboration among multiple data owners.Due to privacy and security concerns,the original data is usually not shared.Therefore,synthetic data sets have been proposed as an alternative solution,which generate new data sets through the analysis of the original data while preserving the statistical characteristics of the original data.Synthetic data has been widely researched and applied in various fields,but its privacy protection remains a significant challenge.Malicious attackers can use statistical features and background knowledge extracted from synthetic data sets to link them with other publicly available information and identify individuals’ sensitive information.Therefore,this thesis aims to explore the use of differential privacy techniques to protect the privacy of synthetic data,enabling secure sharing and utilization in a distributed environment.The specific aspects of the research are as follows:(1)Differential privacy-based data generation framework: This thesis proposes a data protection framework called DPWGAN-GP,which combines differential privacy techniques with Wasserstein Generative Adversarial Networks(WGAN-GP)to train synthetic data with differential privacy properties.A series of optimization strategies,including weight decay and dynamic noise,are employed to improve the framework.Furthermore,a Bidirectional Long Short-Term Memory(BI-LSTM)neural network model is introduced to enhance the performance of the framework by learning the mapping relationship of real data and generating more realistic time series data.Experimental results demonstrate that the proposed framework can generate high-quality data within a given privacy budget,achieving a high classification rate while preserving the privacy and utility of the data.Additionally,the stability of the model is verified through experiments,showing that the model does not suffer from mode collapse or gradient vanishing issues commonly found in GANs,resulting in better and more realistic data generation.(2)Differential privacy-based distributed algorithm: This thesis presents a differential privacy-based distributed algorithm that employs the concept of federated learning to distribute the data set among multiple participants,enabling local and aggregate computation results.By introducing differential privacy mechanisms,the privacy of each participant is ensured,making data sharing between participants more secure and reliable.Synthetic data sets are used to protect the privacy of the original data and improve the generalization performance of the model.Moreover,an online gradient descent algorithm is utilized to adapt to continuously incoming data streams,providing faster response and higher real-time performance.The performance of the algorithm is compared using synthetic and real data sets through experiments conducted on logistic regression.The results demonstrate that the proposed algorithm outperforms in regret performance,maintains a high classification rate under privacy protection,and exhibits faster convergence speed and more stable results.This further validates the effectiveness and feasibility of the differential privacy-based distributed algorithm and synthetic data generation in practical applications. |
