Research On Application Of Differential Privacy Protection For Film Recommendation System

Although recommendation algorithms based on deep learning training provide users with more accurate recommendations,more and more real datasets are used for training.These datasets contain a large amount of personal information,and these unprocessed personal information will cause privacy leakage during the recommendation training process.An attacker can infer user privacy by repeatedly querying the system to master model parameters and combining the acquired background knowledge.However,the traditional differential privacy method uses global sensitivity to calculate the noise scale when solving this problem,resulting in insufficient model training accuracy.This paper focuses on improving the noise injection method of gradient matrix on the basis of satisfying the differential privacy theory in order to improve the accuracy of the algorithm under the same privacy budget.The main research contents are as follows:(1)There is a large gradient redundancy in the gradient descent process of deep neural network,and excessive noise will be introduced when applying differential privacy mechanism to resist member inference attacks.In view of the above problems,the gradient matrix is decomposed by the Funk-SVD matrix decomposition algorithm,noise is added to the lowdimensional feature subspace matrix and the residual matrix,and the redundant gradient noise is eliminated by the gradient reconstruction process.We recalculate the factorization matrix norm and reduce the noise scale in combination with smoothing sensitivity.At the same time,according to the correlation between input features and output,more privacy budget is allocated to features with large correlation coefficients to improve training accuracy.Finally,an adaptive gradient clipping algorithm is proposed to solve the problem of slow convergence based on the mean value of the norm of the decomposition matrix.The algorithm calculates the cumulative privacy loss under various optimization strategies by using moment statistics.Practical results show that the algorithm bridges the gap with non-privacy models more effectively.(2)A new differential privacy algorithm is proposed by introducing the method of local low-rank matrix approximation to solve the problem of adding excessive noise due to ignoring the influence of gradient local features when differential privacy is applied to deep neural networks.The gradient of the generative adversarial network model is decomposed by the method of local low-rank matrix approximation,and noise is added to the decomposed matrix.The gradient matrix is approximated as the weighted sum of low-rank matrices,the local characteristics of noise are fully considered,and noise is added to the decomposed local matrix,which can effectively eliminate excess noise caused by gradient redundancy.The experimental results show that this method can effectively improve the quality of privacy model generation.(3)The optimization ability of the gradient descent of the differential privacy recommendation algorithm is affected by the excessive environmental parameters,and the highdimensional environmental parameters will lead to a decrease in the recommendation accuracy.To solve this problem,a differential privacy recommendation algorithm based on deep autoencoder is proposed.The algorithm uses the self-attention model of perceived time interval to obtain the hidden relationship between the item and the interaction time series,uses the selfattention mechanism model to capture the attribute information of the item,and finally combines the captured information of the two modules to represent.The fully connected layer gradient trained by the recommendation model is input into the auto-encoder,and Gaussian noise is injected into the encoded compressed gradient,and the symmetric mapping relationship between encoding and decoding is used to reduce the influence of environmental parameters on the loss function,thereby improving the training accuracy.The experimental results show that the proposed method can effectively improve the recommendation effect.Finally,based on the above theoretical research,a video recommendation system with differential privacy as the core is built.