GRL learning and adaptive ADAGAN-based generative privacy data protection method and system

Abstract

The application discloses a kind of dual-module generative privacy computing framework (GRL-ADAGAN) based on GRL learning and adaptive ADAGAN, it is related to privacy computing and artificial intelligence technical field, solve the problems of large noise interference, low computing efficiency, insufficient feature retention in the prior art in data asset transaction.The framework includes: discrete data processing module, privacy protection and high-fidelity simulation are carried out on discrete features based on graph-based reinforcement learning (GRL);Continuous data generation module, based on adaptive ADAGAN, generate continuous data with consistent statistical characteristics;Cooperative optimization mechanism, dynamic cooperation is realized through bidirectional parameter feedback, and offline generation-online application strategy is used to reduce overhead;Evaluation system module, privacy protection effect is quantitatively evaluated by multidimensional scoring.The application is suitable for data security circulation and value mining in the fields of finance, medical treatment and the like.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of privacy computing and artificial intelligence, and particularly discloses a dual-module generative privacy computing framework (GRL-ADAGAN) based on GRL and adaptive ADAGAN. The framework combines reinforcement learning and generative adversarial networks to solve the privacy protection problem in data asset transactions and realize "data usability and invisibility". It is suitable for financial, medical and other scenarios, protects sensitive information while ensuring that the generated data is statistically consistent with the original data, and supports direct application of machine learning models. Through dynamic optimization of privacy parameters and hierarchical collaborative mechanisms, the framework overcomes the limitations of traditional privacy protection technologies in data utility, efficiency and feature relevance, and provides an innovative solution for data asset security circulation and value mining. BACKGROUND

[0002] In the digital economy era, data has become a key production factor, and its assetization process is accelerating. As the scale of data transactions continues to expand, how to realize data value mining while ensuring "data usability and invisibility" has become a major challenge faced by the industry and academia. The current mainstream privacy protection technologies all have significant defects in practical application: differential privacy technology can provide theoretical privacy protection, but noise injection will significantly reduce data quality; homomorphic encryption technology is highly secure, but its computational complexity makes it difficult to meet the efficiency requirements of actual business; traditional generative adversarial networks generated synthetic data often cannot fully preserve the statistical properties of the original data and have the risk of privacy leakage; federated learning realizes data non-domain, but faces technical bottlenecks such as high communication cost and slow convergence speed. More importantly, existing technologies lack a unified evaluation system, making it difficult to quantitatively compare the comprehensive performance of different methods in terms of privacy protection strength, data utility preservation, and computational efficiency.

[0003] This technical gap has seriously restricted the healthy development of the data factor market. On the one hand, data holders are reluctant to share data due to concerns about privacy leakage; on the other hand, data demanders find it difficult to obtain data resources that meet privacy requirements and maintain high usability. In particular, in the fields of financial risk control and medical research, which have strict requirements for data quality, traditional privacy protection technologies often fail to meet actual needs. In addition, with the growth of data asset valuation and transaction demand, the market urgently needs a new type of privacy computing solution that can adapt to different types of data characteristics, dynamically optimize privacy parameters, and seamlessly integrate with existing machine learning frameworks.

[0004] In this context, the present application innovatively proposes a technical route combining reinforcement learning with generative adversarial networks, effectively solving the core contradiction of balancing "privacy-utility-efficiency" faced by existing technologies by constructing a dual-module collaborative optimization privacy computing framework, providing an innovative solution for the safe circulation and value release of data assets. SUMMARY

[0005] The present application aims to overcome the technical bottlenecks of existing privacy protection technologies in data asset transactions, such as large noise interference, low computational efficiency, and insufficient feature preservation. It proposes a dual-module generative privacy computing framework based on GRL learning and adaptive ADAGAN (GRL-ADAGAN). This framework, through innovative hierarchical architecture design, achieves collaborative privacy protection for discrete and continuous data, effectively solving the key technical problem of "data usability invisibility". The core technical innovations are mainly reflected in the following aspects:

[0006] (1) Irreversible feature preservation generative privacy protection mechanism: a unique generation mechanism is designed to ensure that any multi-dimensional analysis of synthetic data cannot reconstruct the original data. The generated data is highly consistent with the original data in terms of statistical features and neural network representation, ensuring that the calculation results remain equivalent, and making the data "usable but invisible".

[0007] (2) Adaptive dual-layer collaborative optimization: a dual-layer architecture is developed, including a GRL learning module for discrete data generation and conversion to continuous intervals, and an ADAGAN module for continuous data synthesis. A unified reward function and joint parameter tuning mechanism ensure the consistency and quality of the generated data, thereby improving privacy protection performance.

[0008] (3) Decoupling computation by "offline generation-online application": to solve the computational overhead bottleneck in privacy protection systems, a decoupling strategy is introduced to separate resource-intensive processes. Data desensitization and feature reconstruction are completed in the offline phase, while the online phase uses synthetic data that is fully compatible with the original data structure, thus eliminating the need for additional computational resources for downstream tasks.

[0009] (4) Quantitative privacy model evaluation through privacy, efficiency, and consistency: to compare the invisibility and usability of data protected by popular privacy protection algorithms, as well as their feasibility in practical applications, we designed a three-dimensional evaluation system for comprehensive scoring.

[0010] In summary, the new privacy protection computing framework GRL-ADAGAN proposed in the application combines the dynamic optimization capability of reinforcement learning with the feature learning capability of generative adversarial network, breaks through the performance bottleneck of traditional privacy protection technology, and provides a new technical path for the safe circulation and value mining of data assets. The modular design of the framework has good expansibility and can flexibly adapt to different types of data features and application requirements, which is of great significance to promote the healthy development of the data element market.

[0011] The application provides a dual-module generative privacy computing framework (GRL-ADAGAN) based on GRL learning and adaptive ADAGAN, which adopts a hierarchical collaborative modular architecture design and mainly includes the following core technical innovations:

[0012] (1) Upper reinforcement learning algorithm

[0013] The upper reinforcement learning module is responsible for coarse-grained prediction of discrete data. The module first performs standardization processing on the original data to prepare for the subsequent discretization step. The standardized data is divided into multiple discrete levels, so that the reinforcement learning algorithm can run in the defined state space and action space. By constructing and updating the reinforcement table, the reinforcement learning selects the optimal action according to the current state, and continuously optimizes the prediction strategy through the reward mechanism.

[0014] a. Data preprocessing and discretization

[0015] First, the original data is standardized to scale it to the range of [0, 1] for subsequent processing. The standardization formula is as follows:

[0016] wherein, is the original data, min( ) and max( ) are the minimum and maximum values in the data set, respectively. Subsequently, the standardized data is discretized into m levels, with a discretization range of [0, 999], and the discretization formula is:

[0017]

[0018] Here, ′ is the standardized data, and m is the preset number of discrete levels.

[0019] b. Reinforcement learning modeling

[0020] The core of reinforcement learning lies in the definition of state space, action space, and reward function. The state s represents the characteristics of the current path node, and the action a represents the selected next node value. The reward function P is calculated based on the difference between the predicted value and the true value. The larger the negative absolute error, the lower the reward, thus encouraging the algorithm to improve the prediction accuracy. The reward function formula is as follows:

[0021]

[0022] In reinforcement learning, the update of reinforcement value is a key step, and the formula is as follows:

[0023]

[0024] where α is the learning rate, controlling the learning speed; γ is the discount factor, measuring the importance of future rewards; is the maximum reinforcement value for the next state.

[0025] c. Discrete data prediction

[0026] The reinforcement table obtained by training is used for prediction. The reinforcement table contains the mapping from state s to action a and its corresponding reinforcement value. The prediction process selects the action with the highest reinforcement value as the predicted value:

[0027]

[0028] The reinforcement learning algorithm is based on the reward maximization principle and can provide accurate discrete data prediction results, providing an effective initial range for the continuous data prediction of the lower ADAGAN module. In this way, the upper reinforcement learning module not only improves the efficiency and interpretability of discrete data prediction, but also lays the foundation for achieving the optimal balance between privacy protection and data effectiveness for the entire model.

[0029] (2) Lower module of generative adversarial network (ADAGAN)

[0030] After the upper reinforcement learning module completes the coarse-grained prediction of discrete data, the lower ADAGAN module will take over the task and accurately predict continuous data. ADAGAN optimizes the data generation process through the adversarial training of the generator and discriminator, while protecting data privacy and learning the true distribution of continuous data.

[0031] Reinforcement learning algorithm processes discrete data and outputs discrete values These discrete values define the approximate range of continuous values. Then we use the lower ADAGAN algorithm to accurately predict continuous data. ADAGAN optimizes the data generation process through the adversarial training of the generator and discriminator, and its core goal is to learn the true distribution of continuous data, which is achieved through the adversarial optimization of the generator and discriminator.

[0032] a. ADAGAN model design

[0033] The generator B of ADAGAN receives random noise d and discrete embedding vectors As input, it generates predicted continuous data, mathematically expressed as:

[0034]

[0035] The task of the discriminator is to distinguish between real data and generated data, and to improve the generation quality of the generator through training.

[0036] b. Constraints

[0037] Distribution proportion constraint: control the proportion of generated data in multiple intervals, usually set the number of groups to 3.

[0038]

[0039] where, represents the mean of each segment, represents the distribution proportion of each segment.

[0040] Mean perturbation constraint (adaptive noise): requires the mean of the generated data in each interval to be close to the target value, allowing limited perturbation of the mean.

[0041]

[0042]

[0043] where, is a random perturbation coefficient, represents the target mean, and the mean size is adjusted by the random perturbation coefficient .

[0044] Three-segment distribution constraint function:

[0045]

[0046] If is true, the generated data is accepted; otherwise, it is regenerated.

[0047] c. Loss function and optimization

[0048] The training goal of ADAGAN is to achieve adversarial training by optimizing the loss functions of the generator and the discriminator. The discriminator loss function contains two parts:

[0049]

[0050] The first part measures the ability of the discriminator to judge real data, and the second part measures the ability of the discriminator to identify generated data. The loss function of the generator then tricks the discriminator by minimizing the loss:

[0051]

[0052] The optimization goal of the generator loss function is to generate more accurate predictions of real data.

[0053] d. Combination of discrete and continuous predictions

[0054] The generator and discriminator eventually reach a Nash equilibrium through mutual adversarial training, which can be formally represented as:

[0055]

[0056] In short, the discriminator aims to maximize its loss function to accurately distinguish between real data and generated data, while the generator tries to minimize its loss function to produce more realistic data and deceive the discriminator. Through adversarial training, both are constantly iteratively optimized until they reach an equilibrium state, namely Nash equilibrium.

[0057] e. Combination of discrete and continuous predictions

[0058] First, the prediction of discrete values quantizes the original continuous range to a discrete representation, for example, discrete level 5 corresponds to the original value interval [450, 550]. Then, ADAGAN is used to generate accurate continuous values:

[0059] Then, we use ADAGAN to generate accurate continuous values . ADAGAN receives random noise z and embedding vector E( ) of discrete values as input to the generator B, and outputs the predicted continuous data , which can be represented as:

[0060]

[0061] The simulation data generated by the lower ADAGAN module is consistent with the original data in the continuous feature dimension, providing the required data features for data analysis and machine learning models, and effectively protecting personal privacy.

[0062] (3) Double-layer collaborative mechanism

[0063] The core of the double-layer collaborative mechanism is how to combine the discrete prediction of the reinforcement learning module with the continuous prediction of the ADAGAN module to generate simulated data that not only conforms to the distribution characteristics of the original data but also protects privacy. The reinforcement learning module first performs coarse-grained prediction on discrete data to determine the approximate range of continuous data, and then the ADAGAN module generates accurate continuous data within the range.

[0064] This collaborative method uses reinforcement learning to predict discrete values at the upper layer (as shown in formula 5), fully leveraging the strong interpretability and high efficiency of reinforcement learning to provide a wide and accurate prediction range for the data. At the lower layer, ADAGAN is used to predict continuous values (as shown in formula 13), which learns the underlying distribution of continuous data and refines the wide range determined by the upper layer to generate accurate continuous value predictions.

[0065] Finally, the prediction results of the two levels are integrated: the upper-layer discrete prediction determines the range boundary, and the lower-layer continuous prediction refines the data within the boundary. This hierarchical processing mechanism not only ensures the accuracy of data generation but also effectively protects the privacy characteristics of the original data through the discrete-continuous conversion process.

[0066]

[0067] Total reward function:

[0068]

[0069] where, control the influence weight of reinforcement learning prediction on reward, regulate the contribution degree of generation quality to reward.

[0070] The entire double-layer model achieves collaborative optimization through its respective optimization objectives: the reinforcement learning module optimizes the discrete prediction strategy by maximizing cumulative rewards to ensure accuracy and efficiency of prediction; and the ADAGAN module accurately models the continuous data distribution through the adversarial training of the generator and discriminator to achieve high-fidelity continuous data generation. This double-layer structure not only improves the overall prediction performance but also effectively safeguards data privacy and security through the hierarchical processing mechanism.

[0071] The system workflow includes: 1) data standardization and discretization preprocessing; 2) upper-layer reinforcement learning module for discrete data privacy protection; 3) lower-layer ADAGAN module for continuous data generation; 4) two-way parameter feedback collaborative optimization; 5) PECS system comprehensive evaluation; 6) generation of high-fidelity privacy-protected data. This technical solution achieves the optimal balance between privacy protection and data utility through modular design, providing safe and efficient technical support for data asset transactions. BRIEF DESCRIPTION OF DRAWINGS

[0072] Figure 1 The GRL-ADAGAN algorithm framework design is illustrated in the flowchart of this invention. Detailed Implementation

[0073] This invention provides a graph-based reinforcement learning and adaptive ADAGAN dual-module generative privacy computing framework (GRL-ADAGAN), the specific implementation process of which includes the following key steps:

[0074] (1) System initialization and parameter configuration

[0075] The system first loads the original dataset and performs standardized preprocessing, using the min-max normalization method to normalize the data to a uniform interval. Then, it configures the core parameters of the reinforcement learning module, including the discretization level, learning rate, and discount factor, ensuring the model has good exploratory capabilities in the early stages of training. Simultaneously, it initializes key parameters of the ADAGAN module, such as distribution constraint weights and mean perturbation coefficients, laying the foundation for subsequent adversarial training.

[0076] (2) Discrete data processing stage

[0077] The normalized data is discretized using the K-means algorithm, mapping continuous features to multiple discrete levels. A path graph model is constructed to represent the relationships between data features, defining state and action spaces, and a dynamic strategy is employed to select the optimal action. Based on the designed reward function, the system continuously updates the reinforcement table to optimize the decision-making process, ultimately outputting high-quality discrete prediction results.

[0078] (3) Continuous data generation stage

[0079] After obtaining discrete prediction results, the system uses the ADAGAN module to generate continuous data. The generator receives discrete feature embedding vectors and random noise as input and learns the distribution characteristics of the original data through a multi-layer neural network. The discriminator combines distribution ratio constraints and mean perturbation constraints to ensure that the generated data is highly consistent with the original data in terms of statistical properties. Through an adversarial training mechanism, the system continuously optimizes the performance of the generator and discriminator until a Nash equilibrium is reached.

[0080] (4) Collaborative optimization and evaluation

[0081] The system achieves collaborative optimization between upper and lower layer modules through a two-way parameter feedback mechanism, dynamically adjusting the hyperparameters of each module to balance privacy protection and data utility. Simultaneously, the system employs the PECS scoring system to comprehensively evaluate the privacy protection strength, computational efficiency, and feature consistency of the generated data, ensuring its reliability in practical applications. The final synthesized data not only meets the requirement of "data usable but not visible" but also effectively supports various machine learning tasks.

[0082] This invention proposes a dual-module generative privacy computing framework (GRL-ADAGAN) based on GRL learning and adaptive ADAGAN. By integrating an upper-layer GRL learning module and a lower-layer adaptive ADAGAN module, and introducing a bidirectional parameter feedback mechanism, it successfully solves the challenge of balancing data availability and privacy security in existing privacy protection technologies. The innovation of this framework is mainly reflected in the following aspects:

[0083] A hierarchical collaborative optimization mechanism is employed: the upper-layer GRL learning module processes discrete data, dynamically optimizes privacy parameters, and generates high-fidelity simulated discrete features. Simultaneously, the lower-layer adaptive ADAGAN module combines normal distribution constraints to generate continuous data, ensuring consistency between statistical properties and neural network representations. The two modules achieve collaborative optimization through bidirectional feedback, significantly improving privacy protection and data availability.

[0084] Irreversible feature preservation generation mechanism: Through an offline generation-online application computational decoupling strategy, it ensures that the synthesized data cannot be restored to the original data through reverse analysis, while maintaining a high degree of consistency with the original data in statistical distribution and machine learning tasks, truly achieving the privacy protection goal of "data usable but not visible".

[0085] Three-dimensional quantitative evaluation system: The Privacy-Efficiency-Consistency Scoring System (PECS) is innovatively proposed to comprehensively evaluate the algorithm from three dimensions: privacy protection strength, computational efficiency, and feature relevance, providing a scientific basis for the practical application of privacy protection technology.

[0086] In terms of practicality, the GRL-ADAGAN framework significantly outperforms existing privacy protection technologies. Compared to differential privacy, it excels in data utility and model performance; compared to homomorphic encryption, it is more computationally efficient, making it suitable for large-scale data applications; and compared to traditional generative adversarial networks, the synthetic data it generates more closely resembles the original data in statistical properties and machine learning tasks. This framework provides an efficient and reliable solution for data assetization and privacy protection in fields such as finance, healthcare, and retail, promoting the coordinated development of privacy security and value mining in the marketization of data elements.

[0087] In conclusion, the GRL-ADAGAN framework not only fills the gaps in intelligence, efficiency, and accuracy of existing privacy computing technologies, but also provides an innovative solution for privacy protection practices in the field of data asset management, possessing significant theoretical value and broad application prospects.

Claims

1. A generative privacy data protection method based on GRL learning and adaptive ADAGAN, characterized in that, Includes the following steps: Data preprocessing steps: Standardize and discretize the original continuous data, and map the standardized data into m discrete levels; The upper-layer GRL learning module processes the following steps: It performs coarse-grained predictions on the discretized data based on reinforcement learning, dynamically optimizing privacy parameters to generate high-fidelity simulated discrete features. Coarse-grained prediction includes: constructing the state space and action space, and applying the reward function... Update enhancement values ,in For the true value, Here, s represents the predicted value, and a represents the action. Represent the reinforcement value of the state-action pair; and select the action with the highest reinforcement value as the discrete prediction value. ; Lower-level ADAGAN module processing steps: Based on discrete predicted values Generative adversarial networks (GANs) are used to make accurate predictions on continuous data; the GAN consists of a generator and a discriminator; the generator receives random noise. and discrete predicted values Embedded vector Generate continuous synthetic data The discriminator combines distribution ratio constraints. and mean perturbation constraint This ensures that the synthesized data is consistent with the original data in terms of statistical characteristics; Among them, distribution ratio constraint Defined as: ,in, This represents the distribution proportion of the generated data in the m-th interval. This represents the distribution proportion of the original data in the m-th interval. These are the weighting coefficients; Mean disturbance approximately Defined as: ,in, This represents the mean of the generated data in the m-th interval. Indicates the target mean. The weighting coefficients are used; the target mean is determined by introducing a random perturbation coefficient. To build, that is ,in This represents the mean of the original data in the m-th interval, thus achieving an adaptive normal distribution constraint; Two-way parameter feedback and collaborative optimization steps: through the total reward function It connects the upper-layer GRL learning module and the lower-layer ADAGAN module to achieve collaborative optimization and dynamically balance privacy protection and data utility; among them... The reward for the upper-level GRL learning module is given by the formula. calculate, This is the loss function of the generator in the lower-level ADAGAN module. and These are the weighting coefficients.

2. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, The processor executes a computer program to implement the method as described in any one of claims 1.

3. A computer-readable storage medium having a computer program stored thereon, characterized in that, When a computer program is executed by a processor, it implements the method as described in any one of claims 1.

Images