Research And Implementation Of Privacy Preserving Methods For Government Image Data

The open sharing of government big data is the key to promoting the application and development of big data in government affairs.It breaks business barriers and the phenomenon of”Isolated Data Island”,and realizes cross-departmental and cross-level sharing of information.However,it causes data information to face the risk of privacy leakage during the sharing process.Traditional technologies of privacy preserving cannot meet the multi-purpose sharing needs of open sharing platforms;it is also difficult to ensure the performance of collaborative training while preserving data privacy.As we all know,there are various storage forms of government big data,such as text,images,and videos.Different types of data have different data characteristics,and the methods for implementing privacy preserving are also different.This thesis divides government image data into unconditionally shared and open public data according to the security preserving level of the data itself,shared and open to higher-level departments,conditionally shared and open secret data to the same level and lower-level related departments,and top-secret shared and open only to national ministries and commissions.According to the actual needs of the open sharing platform,the following studies have been conducted on the privacy preserving methods of secret and top-secret data:For the secret image data that can be shared with related departments,a privacy preserving method for shared secret data based on target detection algorithm and encryption technology is designed.Target detection technology is used to reduce the data size during sharing;combined with RSA algorithm and AES algorithm,the shared data is mixed encryption,compared with a single RSA encryption algorithm,it reduces data while preserving data privacy and security.The time consumption of the encryption and decryption process.The usability of the shared secret data privacy preserving method is verified by the hybrid encryption algorithm;the efficiency of the method is verified by the comparison of the encryption and decryption time of the AES algorithm and the RSA algorithm.For the secret image data that cannot be directly shared,a Deep Convolution Generative Adversarial Network with differential privacy model is designed.The model uses differential privacy technology in the Deep Convolutional Generative Adversarial Networks and adds Gaussian noise to the gradient passed from the discriminator back to the generator in Deep Convolutional Generative Adversarial Networks to achieve the effect of protecting data privacy while generating data similar to the original data.Experimental results show that the data generated by this model can be applied to subsequent tasks,and the quality of the model is 0.1% higher than that of Deep Convolutional Generative Adversarial Networks without privacy preserving.For the top-secret image data that can be used for departmental collaborative training,a collaborative training privacy preserving model based on federated learning is designed.According to the phenomenon that most of the tasks that require inter-departmental cooperation in the government affairs big data environment are complex and relevant tasks,multi-task learning and federated learning are combined.During training,each participant transfers the encrypted weight vector to the credible three parties.And the third party aggregates and averages the upload weights,so as to avoid the direct aggregation of the data of the parties and provide the data privacy of the parties.The model allows participants to retain their own data privacy while still benefiting from the models of other participants,avoiding local optima.Taking the horizontal federated learning and multi-task learning model HyperFace as an example,the model is applied to government image data.The final result proves that the collaborative training privacy preserving model will not reduce the usability of the model while preserving data privacy.The above three models and methods correspond to the three types of government image data with different privacy levels in the open sharing platform.According to the privacy preserving requirements of the platform of different data,corresponding privacy preserving strategies are designed to provide privacy for the sharing of government image data.