Generative privacy data protection method and system based on GRL learning and adaptive ADAGAN

Abstract

The invention discloses a dual-module generative privacy computing framework (GRL-ADAGAN) based on GRL learning and an adaptive ADAGAN, relates to the technical field of privacy computing and artificial intelligence, and solves the problems of large noise interference, low computing efficiency and insufficient feature retention in data asset transaction in the prior art. The framework comprises a discrete data processing module for performing privacy protection and high-fidelity simulation on discrete features through graph-based reinforcement learning (GRL); the continuous data generation module is used for generating continuous data with consistent statistical characteristics based on the self-adaptive ADAGAN; according to the collaborative optimization mechanism, dynamic collaboration is achieved through bidirectional parameter feedback, and the overhead is reduced by adopting an offline generation-online application strategy; and the evaluation system module quantitatively evaluates the privacy protection effect through multi-dimensional scores. The method is suitable for data security circulation and value mining in the fields of finance, medical treatment and the like.

Description

Technical Field

[0001] This invention relates to the fields of privacy computing and artificial intelligence technology. Specifically, it discloses a dual-module generative privacy computing framework (GRL-ADAGAN) based on GRL and adaptive ADAGAN. This framework integrates reinforcement learning and generative adversarial networks to address the privacy challenges in data asset transactions, achieving "data availability without visibility." Suitable for scenarios such as finance and healthcare, it ensures statistical consistency between generated data and original data while protecting sensitive information, supporting the direct application of machine learning models. By dynamically optimizing privacy parameters and employing a hierarchical collaborative mechanism, it overcomes the limitations of traditional privacy protection technologies in terms of data utility, efficiency, and feature relevance, providing an innovative solution for the secure circulation and value mining of data assets. Background Art

[0002] In the digital economy, data has become a key production factor, and its assetization is accelerating. As the scale of data transactions continues to expand, how to mine data's value while ensuring "data availability without visibility" has become a major challenge facing both industry and academia. Current mainstream privacy-preserving technologies suffer from significant shortcomings in practical applications: While differential privacy offers theoretical privacy guarantees, noise injection significantly degrades data quality; while homomorphic encryption offers high security, its computational complexity makes it difficult to meet the efficiency requirements of real-world applications; synthetic data generated by traditional generative adversarial networks often fails to fully preserve the statistical properties of the original data and poses privacy risks; and while federated learning ensures data remains within the domain, it faces technical bottlenecks such as high communication costs and slow convergence. More importantly, existing technologies lack a unified evaluation system, making it difficult to quantitatively compare the overall performance of different approaches in terms of privacy protection strength, data utility preservation, and computational efficiency.

[0003] This technological gap severely constrains the healthy development of the data factor market. On the one hand, data holders are hesitant to share data due to concerns about privacy breaches; on the other hand, data demanders struggle to access data resources that meet privacy requirements while maintaining high availability. Traditional privacy protection technologies often fall short of meeting practical needs, particularly in areas with stringent data quality requirements, such as financial risk management and medical research. Furthermore, with the growing demand for data asset valuation and trading, the market urgently needs a new privacy computing solution that can adapt to the characteristics of diverse data types, dynamically optimize privacy parameters, and seamlessly integrate with existing machine learning frameworks.

[0004] In this context, this invention innovatively proposes a technical route that combines reinforcement learning with generative adversarial networks. By constructing a privacy computing framework with dual-module collaborative optimization, it effectively solves the core contradiction of "privacy-utility-efficiency" that is difficult to balance faced by existing technologies, and provides an innovative solution for the safe circulation and value release of data assets. Summary of the Invention

[0005] This paper aims to overcome technical bottlenecks of existing privacy-preserving technologies in data asset transactions, such as high noise interference, low computational efficiency, and insufficient feature preservation. It proposes a dual-module generative privacy computing framework (GRL-ADAGAN) based on GRL learning and adaptive ADAGAN. Through an innovative layered architecture, this framework achieves collaborative privacy protection for both discrete and continuous data, effectively addressing the key technical challenge of "data availability without visibility." Core technological innovations are primarily reflected in the following aspects:

[0006] (1) Irreversible Feature Preserving Generative Privacy Protection Mechanism: A unique generation mechanism is designed to ensure that the original data cannot be reconstructed through any multidimensional analysis of the synthesized data. The generated data is highly consistent with the original data in terms of statistical features and neural network representation, thus ensuring that the computational results remain equivalent while making the data “usable but invisible”.

[0007] (2) Adaptive Two-Layer Collaborative Optimization: A two-layer architecture is developed, consisting of a GRL-based 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 generated data, thereby improving privacy protection performance.

[0008] (3) Computational decoupling through “offline generation – online application”: To address the computational overhead bottleneck in privacy-preserving 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 utilizes synthetic data that is fully compatible with the original data structure, eliminating the need to consume additional computational resources for downstream tasks.

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

[0010] In summary, the novel privacy-preserving computing framework GRL-ADAGAN, proposed in this paper, organically combines the dynamic optimization capabilities of reinforcement learning with the feature learning capabilities of generative adversarial networks. This framework not only overcomes the performance bottlenecks of traditional privacy-preserving technologies but also provides a new technical path for the secure circulation and value mining of data assets. The framework's modular design enables excellent scalability and flexible adaptation to different types of data features and application requirements, which is of great significance for promoting the healthy development of the data factor market.

[0011] This paper provides a dual-module generative privacy computing framework (GRL-ADAGAN) based on GRL learning and adaptive ADAGAN. It adopts a layered and collaborative modular architecture design and mainly includes the following core technical innovations:

[0012] (1) Upper-level reinforcement learning algorithm

[0013] The upper-level reinforcement learning module is responsible for making coarse-grained predictions on discrete data. This module first normalizes the raw data to prepare for the subsequent discretization step. The normalized data is then divided into multiple discrete levels, enabling the reinforcement learning algorithm to operate within a defined state and action space. By constructing and updating the reinforcement table, reinforcement learning selects the optimal action based on the current state and continuously optimizes the prediction strategy through a reward mechanism.

[0014] a. Data preprocessing and discretization

[0015] First, the original data is normalized and scaled to the range of [0, 1] for ease of subsequent processing. The normalization formula is as follows:

[0016] in, 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]. 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 the state space, action space, and reward function. State s represents the characteristics of the current path node, and action a represents the value of the next node selected. 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 incentivizing the algorithm to improve prediction accuracy. The reward function formula is as follows:

[0021]

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

[0023]

[0024] Among them, α is the learning rate, which controls the learning speed; γ is the discount factor, which measures the importance of future rewards; is the maximum reinforcement value for the next state.

[0025] c. Discrete data prediction

[0026] The reinforcement table obtained through 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] Based on the principle of reward maximization, the reinforcement learning algorithm can provide accurate discrete data prediction results, providing an effective initial range for the continuous data prediction of the lower-level ADAGAN module. In this way, the upper-level reinforcement learning module not only improves the efficiency and interpretability of discrete data prediction, but also lays the foundation for the entire model to achieve an optimal balance between privacy protection and data validity.

[0029] (2) Lower layer module Generative Adversarial Network (ADAGAN)

[0030] After the upper-level reinforcement learning module completes the coarse-grained prediction of discrete data, the lower-level ADAGAN module takes over and makes precise predictions for continuous data. ADAGAN optimizes the data generation process through adversarial training between the generator and the discriminator, learning the true distribution of continuous data while protecting data privacy.

[0031] Reinforcement learning algorithms process discrete data and output discrete values , these discrete values ​​define the approximate range of continuous values. Subsequently, we use the underlying ADAGAN algorithm to accurately predict continuous data. ADAGAN optimizes the data generation process through adversarial training between the generator and the discriminator. Its core goal is to learn the true distribution of continuous data, which is achieved through adversarial optimization of the generator and the discriminator.

[0032] a. ADAGAN model design

[0033] The generator B of ADAGAN receives random noise d and a discrete-valued embedding vector As input, generate the predicted continuous data, the mathematical expression is:

[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: controls the proportion of generated data in multiple intervals. The number of groups is usually set to 3.

[0038]

[0039] in, represents the mean of each segment, Indicates the distribution ratio of each segment.

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

[0041]

[0042]

[0043] in, is the random disturbance coefficient, Represents the target mean and uses the random perturbation coefficient To adjust the mean value.

[0044] Three-stage distribution constraint function:

[0045]

[0046] like If true, accept the generated data; otherwise, regenerate.

[0047] c. Loss Function and Optimization

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

[0049]

[0050] The first part measures the discriminator's ability to identify real data, and the second part measures the discriminator's ability to identify generated data. The generator's loss function tries to deceive the discriminator by minimizing the loss:

[0051]

[0052] The optimization goal of the generator loss function is to generate prediction results that are closer to the real data.

[0053] d. Combining discrete and continuous predictions

[0054] The generator and the discriminator eventually reach Nash equilibrium through mutual adversarial training. The process can be formally expressed as:

[0055]

[0056] In short, the discriminator aims to maximize its loss function to accurately distinguish real data from generated data, while the generator attempts to minimize its loss function to produce more realistic data and deceive the discriminator. Through adversarial training, the two are continuously optimized iteratively until they reach a balance state, known as the Nash equilibrium.

[0057] e. Combination of discrete and continuous predictions

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

[0059] We then use ADAGAN to generate accurate continuous values ADAGAN receives random noise z and discrete values The embedding vector E( ) as the input of generator B and output the predicted continuous data , the process can be expressed as:

[0060]

[0061] The simulated data generated by the lower-level ADAGAN module is consistent with the original data in the continuous feature dimension, which not only provides the required data features for data analysis and machine learning models, but also effectively protects personal privacy.

[0062] (3) Two-tier coordination mechanism

[0063] The core of this two-layer collaborative mechanism lies in combining the discrete predictions of the reinforcement learning module with the continuous predictions of the ADAGAN module to generate simulated data that both conforms to the distribution characteristics of the original data and protects privacy. The reinforcement learning module first makes coarse-grained predictions on the discrete data to determine the approximate range of the continuous data. The ADAGAN module then generates precise continuous data within this range.

[0064] This collaborative approach leverages reinforcement learning at the upper layer to predict discrete values ​​(as shown in Equation 5), leveraging the interpretability and efficiency of reinforcement learning to provide a broad and accurate prediction range for the data. At the lower layer, ADAGAN is used to predict continuous values ​​(as shown in Equation 13). By learning the underlying distribution of continuous data, it refines the broad range determined by the upper layer to generate accurate continuous value predictions.

[0065] Ultimately, the prediction results from the two layers are integrated: the upper layer, discrete prediction, defines the range boundaries, while the lower layer, continuous prediction, refines the data within these boundaries. This layered processing mechanism ensures the accuracy of data generation while effectively protecting the privacy of the original data through the discrete-to-continuous conversion process.

[0066]

[0067] Total reward function:

[0068]

[0069] in, Controls the weight of the influence of reinforcement learning predictions on rewards, Regulates the contribution of build quality to rewards.

[0070] The entire two-layer model achieves collaborative optimization through their respective optimization objectives: the reinforcement learning module optimizes the discrete prediction strategy by maximizing cumulative rewards, ensuring prediction accuracy and efficiency; while the ADAGAN module accurately models the continuous data distribution through adversarial training between the generator and the discriminator, achieving high-fidelity continuous data generation. This two-layer structure not only improves overall prediction performance, but also effectively protects data privacy and security through a layered processing mechanism.

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

[0072] Figure 1 . GRL-ADAGAN algorithm framework design, which is a schematic diagram of the process flow in this invention. DETAILED DESCRIPTION

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

[0074] (1) System initialization and parameter configuration

[0075] The system first loads the original dataset and performs standardization preprocessing, using the min-max method to normalize the data to a uniform interval. It then configures the core parameters of the reinforcement learning module, including the discretization level, learning rate, and discount factor, to ensure the model has good exploration capabilities during the initial training phase. It also initializes key parameters of the ADAGAN module, such as the distribution constraint weights and mean perturbation coefficient, to lay 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 into multiple discrete levels. The system constructs a path graph model to characterize the relationships between data features, defines the state space and action space, and employs a dynamic strategy to select the optimal action. Based on a 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 predictions, the system uses the ADAGAN module to generate continuous data. The generator receives discrete feature embeddings 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 statistically highly consistent with the original data. 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 of upper and lower-level modules through a bidirectional parameter feedback mechanism, dynamically adjusting each module's hyperparameters to balance privacy protection and data utility. Furthermore, 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 resulting synthetic data not only meets the requirement of "data available but not visible" but also effectively supports various machine learning tasks.

[0082] This paper proposes a dual-module generative privacy computing framework based on GRL learning and adaptive ADAGAN (GRL-ADAGAN). By integrating the upper-layer GRL learning module and the lower-layer adaptive ADAGAN module and introducing a bidirectional parameter feedback mechanism, it successfully solves the difficult balance between data availability and privacy security faced by existing privacy protection technologies. The innovations of this framework are mainly reflected in the following aspects:

[0083] Hierarchical collaborative optimization mechanism: The upper-level GRL learning module processes discrete data, dynamically optimizes privacy parameters, and generates high-fidelity simulated discrete features. Simultaneously, the lower-level adaptive ADAGAN module generates continuous data using normal distribution constraints, ensuring consistency between statistical properties and neural network representation. These two modules achieve collaborative optimization through bidirectional feedback, significantly improving privacy protection and data availability.

[0084] Irreversible feature-preserving generation mechanism: Through the computational decoupling strategy of offline generation and online application, the synthetic data is ensured to be unable to 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 available but invisible".

[0085] Three-dimensional quantitative evaluation system: The innovative Privacy-Efficiency-Consistency Scoring System (PECS) is proposed to comprehensively evaluate algorithms 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-preserving technologies. Compared to differential privacy, it offers superior data utility and model performance. Compared to homomorphic encryption, it offers higher computational efficiency and is suitable for large-scale data applications. Compared to traditional generative adversarial networks, the synthetic data it generates performs closer to 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 sectors such as finance, healthcare, and retail, promoting the coordinated development of privacy security and value mining in the marketization of data elements.

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

Claims

1. A dual-module generative privacy computing framework based on GRL learning and adaptive ADAGAN (GRL-ADAGAN), characterized by: The following steps are involved: The upper-layer GRL learning module is used to make coarse-grained predictions on discrete data and generate high-fidelity simulated discrete features by dynamically optimizing privacy parameters; The lower-layer ADAGAN module is used to accurately predict continuous data and generate synthetic data that is highly consistent with the original data in statistical properties and neural network representation through adaptive normal distribution constraints; A bidirectional parameter feedback mechanism connects the upper-layer GRL learning module and the lower-layer ADAGAN module to achieve collaborative optimization and dynamically balance privacy strength and data utility.

2. The method according to claim 1, characterized in that The lower layer ADAGAN module includes: Generator design, which receives random noise and discrete value embedding vectors and generates continuous data; The discriminator design combines distribution ratio constraints and mean perturbation constraints to ensure that the generated data is consistent with the original data in terms of statistical characteristics; A three-stage distribution constraint function is used to verify the validity of the generated data.

3. The method according to claim 1, characterized in that The bidirectional parameter feedback mechanism achieves collaborative optimization through the total reward function, which is: , where P Q is the reward of the Q learning module, S G is the loss function of the generator, and is the weight coefficient.

Images