Video user data security analysis and prediction method based on privacy computing

Abstract

This invention discloses a method for secure analysis and prediction of video user data based on privacy computing. The method includes the following steps: S1: Collect video user data and construct a dataset; S2: Allocate privacy budgets to different processing stages of the video user data through multi-level privacy budget partitioning; S3: Adopt an adaptive noise injection mechanism, define a noise injection adjustment formula, and obtain preprocessed video user data; S4: Construct a privacy computing framework and set a loss function by combining a balanced objective function and a graph reconstruction loss function; S5: Train the privacy computing framework using the preprocessed video user data; S6: Execute video user data analysis and prediction using the privacy computing framework and output prediction results. This invention provides an efficient and scientific optimization solution for secure analysis and prediction of video user data, bringing significant technical value and economic benefits to practical applications.

Description

Technical Field

[0001] The present invention relates to the field of data security technology, and in particular to a video user data security analysis and prediction method based on privacy computing. Background Art

[0002] Existing video user data analysis and prediction technologies primarily rely on traditional machine learning methods and deep learning models. While these methods can predict user behavior and preferences to a certain extent, they generally face risks of privacy protection and data leakage. With the increasing stringency of data protection regulations and users' heightened concern for personal privacy, how to effectively analyze and predict data while protecting user privacy has become a pressing issue. Traditional user data analysis and prediction methods typically require large amounts of personal data, which greatly increases the risk of privacy leakage. This is especially true in the context of big data and cloud computing, where the storage and analysis of user behavior data can lead to data leakage or misuse.

[0003] Among these technologies, the research on privacy-preserving data analysis mainly focuses on ensuring data security through encryption technology. Existing encryption methods, such as homomorphic encryption and secure multi-party computing, can effectively protect data privacy, but their computational complexity is high and processing efficiency is low, making it difficult to achieve large-scale, efficient real-time data analysis in actual commercial applications. Especially in video user data analysis, the diversity and complexity of user data require efficient and accurate analysis capabilities, while existing encryption technologies often cannot meet the requirements of real-time and efficiency.

[0004] In addition, traditional video user behavior prediction methods mostly rely on classic deep learning models such as graph neural networks and convolutional neural networks. Although these models have made certain progress in processing structured data, for large-scale and complex user behavior data, especially in multi-dimensional and multi-modal scenarios, the generalization ability and prediction accuracy of the models are still insufficient. These models usually have certain limitations in processing heterogeneous data, long-term dependencies and data noise, resulting in poor stability and reliability of the models in practical applications.

[0005] Existing technologies also face the challenge of maintaining data analysis accuracy while protecting user privacy. Privacy protection techniques are typically implemented by injecting noise into the data. However, excessive noise injection can significantly reduce the effectiveness of the data, thereby affecting the accuracy of the analysis results. Differential privacy, as a common privacy protection method, can effectively control the risk of data leakage. However, ensuring data privacy while preventing noise from excessively affecting model performance remains a technical challenge. Currently, the application of differential privacy technology is mostly limited to simple mechanisms, such as Laplace noise injection or Gaussian noise injection. These mechanisms cannot effectively balance the contradiction between privacy protection and data quality in complex scenarios.

[0006] Against this backdrop, secure data analysis methods based on privacy-preserving computing have gradually attracted the attention of researchers. Privacy-preserving computing, by encrypting or anonymizing sensitive data and combining it with techniques like differential privacy and homomorphic encryption, can analyze and predict user behavior data without exposing their privacy. However, existing privacy-preserving computing frameworks often struggle to strike an effective balance between computational efficiency, noise injection control, and model accuracy. Especially when processing large amounts of user data, maintaining model efficiency and accuracy while ensuring privacy remains a pressing challenge.

[0007] Based on this, combining the privacy computing framework and optimizing the differential privacy mechanism, combined with deep learning methods such as graph neural networks, can provide more accurate and efficient solutions for video user data analysis and prediction. By proposing a new privacy-preserving data analysis method, this method can dynamically adjust the noise injection intensity while ensuring privacy protection, thereby improving analysis accuracy and efficiency, and further overcoming the defects in existing technologies. Summary of the Invention

[0008] One purpose of the present invention is to propose a video user data security analysis and prediction method based on privacy computing. The present invention can provide an efficient and scientific optimization solution in video user data security analysis and prediction, bringing significant technical value and economic benefits to practical applications.

[0009] A video user data security analysis and prediction method based on privacy computing according to an embodiment of the present invention includes the following steps:

[0010] S1. Collect video user data and build a dataset;

[0011] S2. In the data preprocessing stage, privacy budgets are allocated to different processing stages of video user data through multi-level privacy budget division based on data sensitivity, analysis requirements, and processing stages;

[0012] S3. Based on the privacy budget allocation, an adaptive noise injection mechanism is adopted to define a formula for adjusting the noise injection amount. The noise intensity is adjusted according to the noise injection amount to obtain pre-processed video user data.

[0013] S4. Build a privacy computing framework, set a balance objective function between privacy protection and analysis accuracy, and use a graph convolutional network to encode and decode video user data. Set the loss function by combining the balance objective function and the graph reconstruction loss function.

[0014] S5. Use the pre-processed video user data to train the privacy computing framework;

[0015] S6. Perform video user data analysis and prediction through the privacy computing framework and output the prediction results.

[0016] Optionally, S2 includes the following steps:

[0017] S21. According to the processing stage of video user data, a hierarchical structure of privacy budget allocation is set up, and multiple stages are defined;

[0018] S22. Define the privacy budget allocation factor:

[0019] ;

[0020] in, represents the privacy budget allocation factor of the i-th stage, is the data sensitivity weight of the i-th stage, is the computational complexity of the i-th stage, is the analysis requirement weight of the i-th stage, and m is the number of all stages;

[0021] S23. Calculate the privacy budget for each stage based on the privacy budget allocation factor for each stage and the preset overall privacy budget:

[0022] ;

[0023] in, is the preset overall privacy budget, is the privacy budget of the i-th stage;

[0024] S24. Adjust the privacy budget allocation for each stage according to the preset target of the privacy budget allocation.

[0025] Optionally, S3 includes the following steps:

[0026] S31. Setting a dynamic adjustment model for noise injection based on the sensitivity of the video user data, user needs, and computing context, and dynamically adjusting the amount of noise injection by real-time evaluation of privacy protection requirements and computing resources in the computing context;

[0027] S32. The adjustment formula for defining the noise injection amount is:

[0028] ;

[0029] in, represents the noise injection amount in the i-th stage, is the adjustment coefficient, is the data sensitivity of the i-th stage, is the dynamic adjustment factor, is the computational complexity of the i-th stage, is the analysis demand weight of the i-th stage, p and q are adjustment indexes, is the adaptive adjustment coefficient, To calculate the interference factor of the context, is the original data item size, Processing time for the stage;

[0030] S33. Based on the dynamic adjustment of the noise injection amount, a noise injection intensity adjustment formula is further defined:

[0031] ;

[0032] in, is the noise injection intensity of the i-th stage, is the adjustment coefficient of noise injection intensity, is the noise injection amount in the i-th stage, is the privacy budget of the i-th stage, is the exponential adjustment factor of the noise injection intensity, is an additional context adjustment factor, To calculate the load of resources, is the prediction accuracy requirement of stage i;

[0033] S34. Perform noise injection processing on the data according to the adjusted noise injection intensity:

[0034] ;

[0035] in, Represents the pre-processed video user data, is the original data item, is the noise injection intensity of the i-th stage, is the noise adjustment factor, is the noise intensity adjustment coefficient, is the weighting coefficient, is the sensitivity of the jth data item, and k is the total number of data items.

[0036] Optionally, the S4 includes the following steps:

[0037] S41. During the data analysis and prediction phase, set a balance between privacy protection and analysis accuracy, based on the preset privacy budget. and noise injection amount , define the equilibrium objective function as:

[0038] ;

[0039] in, is the balance objective function between privacy protection and analysis accuracy, is the balance coefficient, and its value range is [0,1]. is the data sensitivity of stage i, is the computational complexity of the i-th stage, is the analysis requirement weight of stage i, Preprocessed video user data With the original data item The error measure between them, n is the number of data items;

[0040] S42, by adjusting the balance coefficient , controls the weight distribution between privacy protection strength and analysis accuracy, and defines the optimization formula as:

[0041] ;

[0042] in, is the optimized equilibrium objective function, is the balance coefficient;

[0043] S43, using pre-processed video user data Extract node features of video user data , each node represents the behavioral feature vector of user i, each edge (i, j) represents the relationship between user i and user j, the connection relationship of the edge is represented by the adjacency matrix A, and the adjacency matrix of each node is expressed as , normalize the adjacency matrix A to obtain ;

[0044] S44. Use graph convolutional networks to encode video user behavior data and learn low-dimensional representations of node features through multiple layers of graph convolutional layers:

[0045] ;

[0046] in, represents the node feature representation of the k-th layer, is the normalized adjacency matrix, is the weight matrix of the kth layer, is the bias term, is the sigmoid activation function;

[0047] S45. Introduce the graph attention mechanism and give each edge a different attention weight:

[0048] ;

[0049] in, is the attention coefficient between node i and node j, a is the weight of the attention mechanism, is the feature representation of node i, is the feature representation of node j, Represents the vector connection operation, NE(i) represents the neighbor node set of node i, represents the natural exponential function, and LeakyReLU represents the linear rectification function;

[0050] S46. Update each node i:

[0051] ;

[0052] in, is the updated feature representation of node i, NE(i) is the set of neighbor nodes of node i, is the attention weight between node i and node j, W is the learned weight matrix, is the feature representation of node j, is the sigmoid activation function;

[0053] S47. Use the decoder to reconstruct the graph structure and restore the adjacency matrix:

[0054] ;

[0055] in, is the reconstructed adjacency matrix element, representing the weight of the edge between node i and node j, 、 is the low-dimensional representation of node i and node j, is the sigmoid activation function;

[0056] S48. Obtain prediction results:

[0057] ;

[0058] in, is the prediction result of the i-th data item, is the low-dimensional feature representation of the i-th data item, is the weight matrix used for prediction tasks;

[0059] S49, constructing a graph reconstruction loss function and a prediction accuracy loss function;

[0060] The graph reconstruction loss function is expressed as:

[0061] ;

[0062] in, is the reconstructed adjacency matrix element, is the real adjacency matrix element;

[0063] The prediction accuracy loss function is expressed as:

[0064] ;

[0065] in, is the true label of the i-th data item, is the prediction result of the i-th data item, and n is the number of data items;

[0066] The total loss function is defined as:

[0067] ;

[0068] in, is the optimized equilibrium objective function, Reconstruct the loss function for the graph, is the prediction accuracy loss function, and is the adjustment coefficient.

[0069] Optionally, the S5 includes the following steps:

[0070] S51. Perform privacy-preserving data analysis and prediction using a privacy-preserving computing framework, processing data and training models based on the allocated privacy budget.

[0071] S52. During the privacy protection data analysis process, encrypt and de-identify the data according to the preset privacy protection policy;

[0072] S53. Monitor the consumption of the privacy budget in real time. If it is detected that the privacy budget is about to be overconsumed, adjust the data processing or noise injection strategy in the calculation process.

[0073] The beneficial effects of the present invention are:

[0074] (1) The video user data security analysis and prediction method based on privacy computing proposed in this paper successfully solves the problem of improving data analysis accuracy and prediction effect while ensuring data privacy protection through innovative optimization of existing technologies. First, a multi-level privacy budget division mechanism is adopted to intelligently allocate privacy budgets at different processing stages based on data sensitivity, analysis requirements, and computing context. This method can dynamically adjust the intensity of privacy protection according to specific data processing requirements, ensure the rational use of privacy budgets, thereby avoiding excessive consumption of privacy budgets and improving privacy protection efficiency.

[0075] (2) The present invention combines an adaptive noise injection mechanism. By dynamically adjusting the amount of noise injection, the present invention can more accurately balance the contradiction between privacy protection and data analysis accuracy. The noise injection is not only based on the sensitivity of the data, user needs and computing context, but can also be dynamically adjusted according to the allocation of the privacy budget to ensure that the accuracy of data analysis is not excessively affected. This dynamic adjustment strategy effectively avoids the data distortion problem caused by excessive noise injection in traditional privacy protection methods.

[0076] (3) This invention successfully achieves a balance between privacy protection and accuracy in video user data analysis and prediction by introducing innovative privacy protection technology and efficient data analysis methods. It not only improves the accuracy of data analysis, but also optimizes computing efficiency while protecting user privacy, thus resolving the contradiction between privacy protection and data quality in existing technologies. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0078] Figure 1 This is a flowchart of a video user data security analysis and prediction method based on privacy computing proposed by the present invention;

[0079] Figure 2 This is a flowchart of privacy budget allocation in the video user data security analysis and prediction method based on privacy computing proposed by the present invention. DETAILED DESCRIPTION

[0080] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0081] refer to Figure 1-Figure 2 A video user data security analysis and prediction method based on privacy computing includes the following steps:

[0082] S1. Collect video user data and build a dataset;

[0083] S2. In the data preprocessing stage, privacy budgets are allocated to different processing stages of video user data through multi-level privacy budget division based on data sensitivity, analysis requirements, and processing stages;

[0084] S3. Based on the privacy budget allocation, an adaptive noise injection mechanism is adopted to define a formula for adjusting the noise injection amount. The noise intensity is adjusted according to the noise injection amount to obtain pre-processed video user data.

[0085] S4. Build a privacy computing framework, set a balance objective function between privacy protection and analysis accuracy, and use a graph convolutional network to encode and decode video user data. Set the loss function by combining the balance objective function and the graph reconstruction loss function.

[0086] S5. Use the pre-processed video user data to train the privacy computing framework;

[0087] S6. Perform video user data analysis and prediction through the privacy computing framework and output the prediction results.

[0088] In this embodiment, S2 includes the following steps:

[0089] S21. According to the processing stage of video user data, a hierarchical structure of privacy budget allocation is set up, and multiple stages are defined;

[0090] S22. Define the privacy budget allocation factor:

[0091] ;

[0092] in, represents the privacy budget allocation factor of the i-th stage, is the data sensitivity weight of the i-th stage, is the computational complexity of the i-th stage, is the analysis requirement weight of the i-th stage, and m is the number of all stages;

[0093] S23. Calculate the privacy budget for each stage based on the privacy budget allocation factor for each stage and the preset overall privacy budget:

[0094] ;

[0095] in, is the preset overall privacy budget, is the privacy budget of the i-th stage;

[0096] S24. Adjust the privacy budget allocation for each stage according to the preset target of the privacy budget allocation.

[0097] In this embodiment, S3 includes the following steps:

[0098] S31. Setting a dynamic adjustment model for noise injection based on the sensitivity of the video user data, user needs, and computing context, and dynamically adjusting the amount of noise injection by real-time evaluation of privacy protection requirements and computing resources in the computing context;

[0099] S32. The adjustment formula for defining the noise injection amount is:

[0100] ;

[0101] in, represents the noise injection amount in the i-th stage, is the adjustment coefficient, is the data sensitivity of the i-th stage, is the dynamic adjustment factor, is the computational complexity of the i-th stage, is the analysis demand weight of the i-th stage, p and q are adjustment indexes, is the adaptive adjustment coefficient, To calculate the interference factor of the context, is the original data item size, Processing time for the stage;

[0102] S33. Based on the dynamic adjustment of the noise injection amount, a noise injection intensity adjustment formula is further defined:

[0103] ;

[0104] in, is the noise injection intensity of the i-th stage, is the adjustment coefficient of noise injection intensity, is the noise injection amount in the i-th stage, is the privacy budget of the i-th stage, is the exponential adjustment factor of the noise injection intensity, is an additional context adjustment factor, To calculate the load of resources, is the prediction accuracy requirement of stage i;

[0105] S34. Perform noise injection processing on the data according to the adjusted noise injection intensity:

[0106] ;

[0107] in, Represents the pre-processed video user data, is the original data item, is the noise injection intensity of the i-th stage, is the noise adjustment factor, is the noise intensity adjustment coefficient, is the weighting coefficient, is the sensitivity of the jth data item, and k is the total number of data items.

[0108] In this embodiment, S4 includes the following steps:

[0109] S41. During the data analysis and prediction phase, set a balance between privacy protection and analysis accuracy, based on the preset privacy budget. and noise injection amount , define the equilibrium objective function as:

[0110] ;

[0111] in, is the balance objective function between privacy protection and analysis accuracy, is the balance coefficient, and its value range is [0,1]. is the data sensitivity of stage i, is the computational complexity of the i-th stage, is the analysis requirement weight of stage i, Preprocessed video user data With the original data item The error measure between them, n is the number of data items;

[0112] S42, by adjusting the balance coefficient , controls the weight distribution between privacy protection strength and analysis accuracy, and defines the optimization formula as:

[0113] ;

[0114] in, is the optimized equilibrium objective function, is the balance coefficient;

[0115] S43, using pre-processed video user data Extract node features of video user data , each node represents the behavioral feature vector of user i, each edge (i, j) represents the relationship between user i and user j, the connection relationship of the edge is represented by the adjacency matrix A, and the adjacency matrix of each node is expressed as , normalize the adjacency matrix A to obtain ;

[0116] S44. Use graph convolutional networks to encode video user behavior data and learn low-dimensional representations of node features through multiple layers of graph convolutional layers:

[0117] ;

[0118] in, represents the node feature representation of the k-th layer, is the normalized adjacency matrix, is the weight matrix of the kth layer, is the bias term, is the sigmoid activation function;

[0119] S45. Introduce the graph attention mechanism and give each edge a different attention weight:

[0120] ;

[0121] in, is the attention coefficient between node i and node j, a is the weight of the attention mechanism, is the feature representation of node i, is the feature representation of node j, Represents the vector connection operation, NE(i) represents the neighbor node set of node i, represents the natural exponential function, and LeakyReLU represents the linear rectification function;

[0122] S46. Update each node i:

[0123] ;

[0124] in, is the updated feature representation of node i, NE(i) is the set of neighbor nodes of node i, is the attention weight between node i and node j, W is the learned weight matrix, is the feature representation of node j, is the sigmoid activation function;

[0125] S47. Use the decoder to reconstruct the graph structure and restore the adjacency matrix:

[0126] ;

[0127] in, is the reconstructed adjacency matrix element, representing the weight of the edge between node i and node j, 、 is the low-dimensional representation of node i and node j, is the sigmoid activation function;

[0128] S48. Obtain prediction results:

[0129] ;

[0130] in, is the prediction result of the i-th data item, is the low-dimensional feature representation of the i-th data item, is the weight matrix used for prediction tasks;

[0131] S49, constructing a graph reconstruction loss function and a prediction accuracy loss function;

[0132] The graph reconstruction loss function is expressed as:

[0133] ;

[0134] in, is the reconstructed adjacency matrix element, is the real adjacency matrix element;

[0135] The prediction accuracy loss function is expressed as:

[0136] ;

[0137] in, is the true label of the i-th data item, is the prediction result of the i-th data item, and n is the number of data items;

[0138] The total loss function is defined as:

[0139] ;

[0140] in, is the optimized equilibrium objective function, Reconstruct the loss function for the graph, is the prediction accuracy loss function, and is the adjustment coefficient.

[0141] In this embodiment, S5 includes the following steps:

[0142] S51. Perform privacy-preserving data analysis and prediction using a privacy-preserving computing framework, processing data and training models based on the allocated privacy budget.

[0143] S52. During the privacy protection data analysis process, encrypt and de-identify the data according to the preset privacy protection policy;

[0144] S53. Monitor the consumption of the privacy budget in real time. If it is detected that the privacy budget is about to be overconsumed, adjust the data processing or noise injection strategy in the calculation process.

[0145] Example:

[0146] In this example, a video streaming platform has over 50 million registered users and an average daily active user volume of 15 million. When users watch videos on this platform, the platform collects user behavior data, including viewing time, viewing types, click-through rates, and user interactions. This data is crucial for the platform, as it helps it make content recommendations, deliver targeted ads, and optimize video content. However, because this data involves user privacy, effectively analyzing the data and providing accurate predictions while ensuring privacy protection becomes a challenge.

[0147] Traditional video data analysis methods usually require collecting large amounts of user behavior data for analysis. However, when processing sensitive data, the risk of data privacy leakage is very high. Therefore, in this embodiment, the implementers will apply the above-mentioned "video user data security analysis and prediction method based on privacy computing" to effectively solve this problem through technical means such as multi-level privacy budget division, dynamic noise injection mechanism, and differential privacy protection.

[0148] On video streaming platforms, data collection is first performed to record various types of user behavior data, including the length of video viewing, the frequency of user ad clicks, comment interactions, etc. Through multi-level privacy budget division, each user behavior data is divided into different privacy processing levels, such as user personal information, viewing history, ad clicks, etc. For each type of data, the privacy budget is intelligently allocated according to its sensitivity and analysis needs.

[0149] Next, through an adaptive noise injection mechanism, appropriate noise is added to the user data processing at each stage to ensure that privacy protection is not over-consumed while maximizing the accuracy of data analysis. In the data preprocessing stage, a differential privacy mechanism is applied to inject noise into user behavioral data such as viewing time and interaction frequency, allowing the platform to maintain user privacy while still conducting reasonable statistical analysis of user behavior.

[0150] During the data analysis phase, the amount of noise injected is dynamically adjusted based on the set privacy budget and noise injection intensity to ensure that the platform can obtain effective analysis results. At the same time, by performing data analysis and prediction through the privacy computing framework, the platform can quickly calculate the interest preferences of each user and combine it with the privacy protection mechanism to perform personalized content recommendations and precise advertising delivery.

[0151] To verify the effectiveness of the method, experimental data was collected from a video streaming platform that collected behavioral data from approximately 30 million active users over the past month. To ensure user privacy, the platform applied the privacy budget partitioning and noise injection techniques of the present invention during the analysis process. When predicting user interests, the platform used different privacy protection strategies (such as traditional noise injection and the method of the present invention) for comparative testing.

[0152] The following is a comparison of the analysis results of the method of the present invention and the traditional method under different privacy protection strategies:

[0153] ;

[0154] Throughout the embodiment, the implementers, through the method of the present invention, not only resolved the contradiction between privacy protection and data analysis accuracy in the process of video user data analysis, but also greatly improved the data analysis efficiency and prediction accuracy. By introducing a dynamic noise injection mechanism and a privacy budget division strategy, the present invention can optimize the data analysis process while ensuring user privacy, thereby improving the accuracy of video user behavior prediction.

[0155] This method intelligently allocates privacy budgets and incorporates differential privacy mechanisms to dynamically adjust noise injection intensity at each data processing stage, effectively avoiding the data distortion caused by excessive noise in traditional methods. Compared to traditional static noise injection methods, this method more precisely balances privacy protection with data validity, eliminating the need to sacrifice analytical accuracy for privacy protection, thereby improving the model's predictive performance.

[0156] With the support of the privacy computing framework, this invention successfully realizes the efficient processing and real-time analysis of large-scale user data. Through the adaptive noise injection mechanism, the stability and response speed of the prediction algorithm are optimized. By dynamically adjusting the amount of noise injection, the platform can effectively ensure the privacy of data and the accuracy of analysis results under different privacy budgets, so that in practical applications, it can flexibly respond to data privacy protection needs in different scenarios.

[0157] This invention not only improves the prediction accuracy of video user behavior analysis, but also achieves high efficiency and real-time data analysis while protecting privacy. By comprehensively considering the sensitivity of the data, user needs and computing context, the present invention provides an efficient, flexible and scalable solution that can be optimized and adapted in multiple practical application scenarios, always maintaining the dual optimization of privacy protection and data analysis.

[0158] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A video user data security analysis and prediction method based on privacy computing, characterized by: The steps include: S1. Collect video user data and build a dataset; S2. In the data preprocessing stage, privacy budgets are allocated to different processing stages of video user data through multi-level privacy budget division based on data sensitivity, analysis requirements, and processing stages; S3. Based on the privacy budget allocation, an adaptive noise injection mechanism is adopted to define a formula for adjusting the noise injection amount. The noise intensity is adjusted according to the noise injection amount to obtain pre-processed video user data. S4. Build a privacy computing framework, set a balance objective function between privacy protection and analysis accuracy, and use a graph convolutional network to encode and decode video user data. Set the loss function by combining the balance objective function and the graph reconstruction loss function. S5. Use the pre-processed video user data to train the privacy computing framework; S6. Perform video user data analysis and prediction through the privacy computing framework and output the prediction results; The S4 comprises the following steps: S41. During the data analysis and prediction phase, set a balance between privacy protection and analysis accuracy, based on the preset privacy budget. and noise injection amount , define the equilibrium objective function as: ； in, is the balance objective function between privacy protection and analysis accuracy, is the balance coefficient, and its value range is [0,1]. is the data sensitivity of stage i, is the computational complexity of the i-th stage, is the analysis requirement weight of stage i, Preprocessed video user data With the original data item The error measure between them, n is the number of data items; S42, by adjusting the balance coefficient , controls the weight distribution between privacy protection strength and analysis accuracy, and defines the optimization formula as: ； in, is the optimized equilibrium objective function, is the balance coefficient; S43, using pre-processed video user data Extract node features of video user data , each node Represents the behavioral feature vector of user i, each edge Represents the relationship between user i and user j, and the edge connection relationship is represented by the adjacency matrix A, and the adjacency matrix of each node is expressed as , for the adjacency matrix Normalize to obtain ; S44. Use graph convolutional networks to encode video user behavior data and learn low-dimensional representations of node features through multiple layers of graph convolutional layers: ； in, represents the node feature representation of the k-th layer, is the normalized adjacency matrix, is the weight matrix of the kth layer, is the bias term, is the sigmoid activation function; S45. Introduce the graph attention mechanism and give each edge a different attention weight: ； in, is the attention coefficient between node i and node j, is the weight of the attention mechanism, is the feature representation of node i, is the feature representation of node j, Represents a vector concatenation operation, represents the set of neighbor nodes of node i, represents the natural exponential function, represents the linear rectification function; S46. Update each node i: ； in, is the updated feature representation of node i, is the set of neighbor nodes of node i, is the attention weight between node i and node j, is the learned weight matrix, is the feature representation of node j, is the sigmoid activation function; S47. Use the decoder to reconstruct the graph structure and restore the adjacency matrix: ； in, is the reconstructed adjacency matrix element, representing the weight of the edge between node i and node j, 、 is the low-dimensional representation of node i and node j, is the sigmoid activation function; S48. Obtain prediction results: ； in, is the prediction result of the i-th data item, is the low-dimensional feature representation of the i-th data item, is the weight matrix used for prediction tasks; S49, constructing a graph reconstruction loss function and a prediction accuracy loss function; The graph reconstruction loss function is expressed as: ； in, is the reconstructed adjacency matrix element, is the real adjacency matrix element; The prediction accuracy loss function is expressed as: ； in, is the true label of the i-th data item, is the prediction result of the i-th data item, and n is the number of data items; The total loss function is defined as: ； in, is the optimized equilibrium objective function, Reconstruct the loss function for the graph, is the prediction accuracy loss function, and is the adjustment coefficient.

2. The video user data security analysis and prediction method based on privacy computing according to claim 1 is characterized in that: The S2 comprises the following steps: S21. According to the processing stage of video user data, a hierarchical structure of privacy budget allocation is set up, and multiple stages are defined; S22. Define the privacy budget allocation factor: ； in, represents the privacy budget allocation factor of the i-th stage, is the data sensitivity weight of the i-th stage, is the computational complexity of the i-th stage, is the analysis requirement weight of the i-th stage, and m is the number of all stages; S23. Calculate the privacy budget for each stage based on the privacy budget allocation factor for each stage and the preset overall privacy budget: ； in, is the preset overall privacy budget, is the privacy budget of the i-th stage; S24. Adjust the privacy budget allocation for each stage according to the preset target of the privacy budget allocation.

3. The video user data security analysis and prediction method based on privacy computing according to claim 1 is characterized in that: The S3 includes the following steps: S31. Setting a dynamic adjustment model for noise injection based on the sensitivity of the video user data, user needs, and computing context, and dynamically adjusting the amount of noise injection by real-time evaluation of privacy protection requirements and computing resources in the computing context; S32. The adjustment formula for defining the noise injection amount is: ； in, represents the noise injection amount in the i-th stage, is the adjustment coefficient, is the data sensitivity of the i-th stage, is the dynamic adjustment factor, is the computational complexity of the i-th stage, is the analysis demand weight of the i-th stage, p and q are adjustment indexes, is the adaptive adjustment coefficient, To calculate the interference factor of the context, is the original data item size, Processing time for the stage; S33. Based on the dynamic adjustment of the noise injection amount, a noise injection intensity adjustment formula is further defined: ； in, is the noise injection intensity of the i-th stage, is the adjustment coefficient of noise injection intensity, is the noise injection amount in the i-th stage, is the privacy budget of the i-th stage, is the exponential adjustment factor of the noise injection intensity, is an additional context adjustment factor, To calculate the load of resources, is the prediction accuracy requirement of stage i; S34. Perform noise injection processing on the data according to the adjusted noise injection intensity: ； in, Represents the pre-processed video user data, is the original data item, is the noise injection intensity of the i-th stage, is the noise adjustment factor, is the noise intensity adjustment coefficient, is the weighting coefficient, is the sensitivity of the jth data item, and k is the total number of data items.

4. The video user data security analysis and prediction method based on privacy computing according to claim 1 is characterized in that: The S5 comprises the following steps: S51. Perform privacy-preserving data analysis and prediction using a privacy-preserving computing framework, processing data and training models based on the allocated privacy budget. S52. During the privacy protection data analysis process, encrypt and de-identify the data according to the preset privacy protection policy; S53. Monitor the consumption of the privacy budget in real time. If it is detected that the privacy budget is about to be overconsumed, adjust the data processing or noise injection strategy in the calculation process.

Images