A Dynamic Data Privacy Protection Method Based on the Lyapunov-SNC Collaborative Computing Model

Abstract

A dynamic data privacy protection method based on the Lyapunov-SNC collaborative computing model belongs to the field of specific computing model fusion technology. It includes real-time acquisition of system operating status data, time-series smoothing processing, and extraction of environmental parameters and data features; parallel execution of Lyapunov stability control and stochastic network calculus (SNC) reliability assessment; construction of a collaborative queue management mechanism; modeling a multi-objective privacy protection optimization problem, using the NSGA-II algorithm to solve the Pareto solution set, and dynamically selecting privacy protection strategies; converting these into execution instructions, real-time monitoring of system performance indicators, and dynamic parameter adjustment based on feedback to form a closed-loop adaptive optimization. This invention solves the system dynamic stability control problem through Lyapunov optimization, solves the service quality quantification assessment problem through SNC, and dynamically selects and combines the optimal data privacy protection mechanism based on machine learning algorithms, data characteristics, application scenarios, and threat models, providing a unified optimization framework and reducing the complexity of combining multiple privacy technologies.

Description

Technical Field

[0001] This invention belongs to the field of specific computing model fusion technology, and specifically relates to a dynamic data privacy protection method based on the Lyapunov-SNC collaborative computing model. Background Technology

[0002] With the rapid development of big data and artificial intelligence technologies, data privacy protection has become a significant challenge for all industries. In the energy sector, with the widespread adoption of smart grids and renewable energy, electricity data contains a large amount of sensitive information, such as user electricity consumption habits and energy production status, making the protection of this data particularly important.

[0003] Existing data privacy protection technologies, such as differential privacy, federated learning, homomorphic encryption, and data anonymization, each have their advantages but also significant limitations. Differential privacy is suitable for statistical queries but not for machine learning model training; federated learning is suitable for distributed training but incurs high communication overhead; homomorphic encryption supports ciphertext computation but has extremely low computational efficiency; data anonymization is simple to process but offers limited privacy protection.

[0004] In terms of dynamic optimization, existing technologies exhibit a clear tendency towards simplification. Some solutions employ the Lyapunov optimization framework for dynamic system stability control, achieving long-term stability by constructing Lyapunov functions and minimizing drift plus penalty functions. However, these solutions have serious drawbacks: while guaranteeing system stability, they lack the ability to quantitatively assess service quality and cannot provide theoretical guarantees for key service quality indicators such as queue overflow probability and latency boundaries. Another approach uses Stochastic Network Calculus (SNC) for service quality assessment, constructing arrival and service curves to rigorously mathematically model queue overflow probability and provide quantifiable service quality guarantees. However, this approach also has significant shortcomings: it focuses on static service quality assessment, lacks dynamic stability control mechanisms, and cannot achieve long-term stable operation of the system in time-varying environments.

[0005] Among the many existing data privacy protection schemes, for example, Huang et al., in their paper "MEC-Enabled TaskReplication With Resource Allocation for Reliability-Sensitive Services in 5GmMTC Networks" (IEEE Transactions on Services Computing, vol. 18, no. 1, pp.253-269, Jan.-Feb. 2025), focused on methods to ensure service reliability in mobile edge computing (MEC) environments through task replication and resource allocation. This scheme embodies a typical approach to optimization decisions in dynamic, resource-constrained network environments. However, such schemes and other existing technologies still have the following shortcomings: First, existing schemes generally adopt a single optimization paradigm, either Lyapunov optimization, SNC evaluation, or other optimization methods, which cannot simultaneously ensure the coordinated protection of system stability and service quality. Second, there is a lack of research on the collaborative working mechanism of Lyapunov optimization and SNC evaluation, and this technological gap seriously restricts the further improvement of the performance of dynamic privacy protection systems.

[0006] In summary, the main drawbacks of existing technologies include:

[0007] (1) Limited application scenarios: Single data privacy protection mechanisms are often only applicable to specific scenarios and are difficult to adapt to diverse energy data protection needs.

[0008] (2) Difficulty in balancing privacy and availability: Existing technologies are difficult to adaptively adjust the strength of privacy protection, often resulting in either insufficient privacy protection or excessive sacrifice of data availability.

[0009] (3) Lack of dynamic adjustment capability: Traditional solutions mostly use static parameter configuration, which cannot dynamically adjust the protection strategy according to data characteristics, attack threats and environmental changes.

[0010] (4) High complexity of combined applications: When multiple data privacy protection technologies are used in combination, there is a lack of a unified optimization framework, resulting in a complex and inefficient system. Summary of the Invention

[0011] To overcome the shortcomings of the existing technologies, the present invention aims to provide a dynamic data privacy protection method based on the Lyapunov-SNC collaborative computing model. This method expands the application scenarios by constructing a data privacy protection mechanism library, integrating multiple privacy protection technologies, and standardizing interfaces. It also establishes a multi-dimensional evaluation system to achieve the optimal balance between data privacy protection strength and data availability. Furthermore, it innovatively proposes a collaborative queue management mechanism combining the Lyapunov optimization framework and Stochastic Network Calculus (SNC). Lyapunov optimization addresses the dynamic stability control problem, while SNC addresses the quantitative evaluation of service quality. Based on machine learning algorithms, it dynamically selects and combines the optimal data privacy protection mechanisms according to data characteristics, application scenarios, and threat models, providing a unified optimization framework and reducing the complexity of combining multiple privacy technologies.

[0012] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0013] A dynamic data privacy protection method based on the Lyapunov-SNC collaborative computing model includes the following steps:

[0014] Step 1: Collect system operation status data in real time, perform time-series smoothing on the collected system operation status data, and extract environmental parameters and data features;

[0015] Step 2: Based on the system operating status data processed in Step 1, perform Lyapunov stability control and stochastic network calculus (SNC) reliability assessment in parallel; and construct a collaborative queue management mechanism.

[0016] Step 3: Based on the collaborative queue management mechanism built in Step 2, model a multi-objective privacy protection optimization problem, use the NSGA-II algorithm to solve the Pareto solution set of the multi-objective privacy protection optimization problem, and dynamically select a privacy protection strategy.

[0017] Step 4: Convert the privacy protection strategy selected in Step 3 into execution instructions, monitor system performance indicators in real time, and dynamically adjust parameters based on feedback to form a closed-loop adaptive optimization.

[0018] The specific method of step 1 includes:

[0019] Step 1.1: Collect and monitor system operation status data in real time;

[0020] Status data for the following queues is continuously collected through a distributed sensor network deployed at edge nodes:

[0021] Data queue length , representing the total amount of user data waiting to be processed in time slot t, in bits or number of tasks;

[0022] Privacy demand queue The quantitative accumulation of privacy protection requirements is calculated as follows:

[0023]

[0024] in, To meet the privacy protection requirements of user i in time slot t, This represents the amount of privacy protection already satisfied;

[0025] Utility demand queue This reflects the data availability maintenance requirements and is calculated as follows:

[0026]

[0027] in, For user i's utility needs in time slot t, To retain the amount of utility already realized;

[0028] Step 1.2: Use a time-series method to smooth queue fluctuations;

[0029] The queue status data collected in step 1.1 is preprocessed using time series analysis to eliminate instantaneous noise interference, resulting in processed queue status data; the window size for smoothing is adaptively adjusted according to the dynamic characteristics of the system.

[0030] Step 1.3: Extract multi-dimensional environmental parameters and data features;

[0031] By using multi-source information fusion technology, system operating environment parameters, including computing resource availability, are collected. Network bandwidth status Security Threat Level and real-time performance requirements Simultaneously, a lightweight machine learning model is used to analyze the characteristics of the data to be processed, including data sensitivity. Data scale Structural complexity and real-time requirements .

[0032] The specific method for step 2 includes:

[0033] Step 2.1: Construct the Lyapunov optimization framework and solve for the stability control strategy;

[0034] Based on the queue state data processed in step 1.2, the queue state data is mapped to the unified state vector required for Lyapunov optimization. The calculation method is as follows:

[0035]

[0036] in: For user queues;

[0037] Edge server queue Determined by the length of the relevant data queue The aggregation yields the result, calculated as follows:

[0038]

[0039] in, This refers to the weighting factor for user u, which is based on the proportion of computing resources reserved as specified in user u's Service Level Agreement (SLA). or minimum guaranteed bandwidth Dynamically determined; the calculation formula is:

[0040] Based on the proportion of reserved computing resources ,but:

[0041]

[0042] Based on the minimum guaranteed bandwidth ,but:

[0043]

[0044] in, For the set of all users assigned to edge server b, This represents the percentage of computing resources reserved for user u as agreed in the Service Level Agreement (SLA), with a value range of [value range missing]. , This represents the minimum guaranteed bandwidth agreed upon by user u in the Service Level Agreement (SLA).

[0045] Equivalent persistent queues to guarantee worst-case latency The ε-persistent queue length of its next time slot The calculation formula is:

[0046]

[0047] in, As a preset constant, For the service rate of server b, To discard task quantity, For the system's maximum service rate, This is an indicator function; it returns 1 when the condition is true and 0 otherwise.

[0048] Virtual reliability queue for handling reliability constraints derived from random network calculus (SNC). The length of the virtual reliability queue in its next time slot The calculation formula is:

[0049]

[0050] in, The probability of queue overflow for user i is calculated using the following formula:

[0051]

[0052] in, The target service quality index for user i is determined by the Service Level Agreement (SLA) or application scenario requirements. Let be the maximum allowed task processing latency for user i; e is the base of the natural logarithm.

[0053] The probability estimate of queue overflow for user i in time slot t is calculated using the following formula:

[0054]

[0055] in, For the size of the historical observation window, For indicator functions;

[0056] Based on unified state vector Construct a quadratic Lyapunov function:

[0057]

[0058] Calculate single-slot conditional Lyapunov drift:

[0059]

[0060] Solve the single-slot conditional Lyapunov drift plus cost minimization problem:

[0061]

[0062] Where, ∑ t) represents the total task delay, and V is a Lyapunov control parameter used to balance stability and performance. It is calculated as follows:

[0063]

[0064] in, , , The maximum capacity of the queue. For maximum cost, the Lyapunov stability coefficient γ(t) is the negative of the rate of change of the Lyapunov function in adjacent time slots, used to quantify the speed at which the system tends to stability:

[0065]

[0066] , It is a range of The positive adjustment parameter is used to ensure that the denominator is not zero. ;

[0067] Step 2.2: Model and quantify service quality based on random network calculus (SNC);

[0068] Using arrival curve Service Line Modeling the system queue behavior:

[0069]

[0070] in, This represents the upper bound of the cumulative arrival flow from time 0 to t, used to characterize the randomness and burstiness of data arrival; This represents the lower bound of the cumulative service capacity from time 0 to t, used to characterize the uncertainty of the system's processing capacity; The value represents the service quality index, reflecting the stringency of the system's requirements for service quality; the higher the value, the more stringent the requirements. Indicates in time slot The amount of data arriving within the region; Indicates in time slot The internal system provides the service rate for user i on server b, and E[] represents the expected operation, which is used to handle randomness;

[0071] Calculate the service quality index θ(t) to characterize the system's ability to meet service quality requirements:

[0072]

[0073] Where θ(t) is a dynamic service quality index. The larger the value, the lower the system's tolerance for latency and queue overflow, and the higher the service quality requirements. This is the upper bound of the queue overflow probability, i.e., the maximum queue overflow probability allowed by the system, calculated using the following formula: ; The maximum allowed length of the queue is used for normalization, making θ(t) dimensionless and dynamically adjustable.

[0074] Step 2.3: Construct a collaborative queue management mechanism;

[0075] Step 2.3.1, construct a virtual reliability queue:

[0076] Based on the target queue overflow probability obtained in step 2.1 Estimation of queue overflow probability Build a virtual reliability queue The calculation formula is:

[0077]

[0078] Step 2.3.2, Dynamically adjust the service quality index:

[0079] Based on the Lyapunov stability coefficient γ(t) and target queue overflow probability obtained in step 2.1 Estimation of queue overflow probability The difference will be used to dynamically adjust the service quality index in step 2.2. The formula is adjusted as follows:

[0080]

[0081] Step 2.3.3, construct the collaborative optimization objective function:

[0082] Based on the minimization of Lyapunov drift plus cost in step 2.1 and the sum of the squares of the differences between the estimated queue overflow probabilities and the queue overflow probabilities of all objectives, a unified collaborative optimization objective function is constructed:

[0083]

[0084] in, , which is a weighting coefficient used to balance stability control objectives and service quality assurance objectives.

[0085] The specific method for step 3 includes:

[0086] Based on the collaborative queue management mechanism constructed in step 2, the privacy protection problem is modeled as a multi-objective privacy protection optimization problem; four competing objective functions are defined, including: privacy protection strength objective function f1(x), data utility objective function f2(x), computational cost objective function, and communication cost objective function f4(x); where:

[0087] The objective function f1(x) for privacy protection strength is calculated as follows:

[0088]

[0089] in, For differential privacy budgeting, For homomorphic encryption strength, To learn about the level of privacy protection in federal education;

[0090] The objective function f2(x) for data utility is calculated as follows:

[0091]

[0092] The objective function f3(x) for calculating the cost is calculated as follows:

[0093]

[0094] The objective function for communication overhead, f4(x), is calculated as follows:

[0095] ;

[0096] The set of design constraints includes: {computational resource constraints, bandwidth constraints, latency constraints, and privacy strength constraints}.

[0097] Computational resource constraints: Utilizing the computational resource availability extracted in step 1.3 As an upper bound constraint;

[0098] Bandwidth constraints: Utilizing the network bandwidth status extracted in step 1.3 As an upper bound constraint;

[0099] Delay constraints: Real-time requirements extracted from step 1.3 Sure;

[0100] Privacy strength constraints: Data sensitivity extracted from step 1.3 Security threat levels in environmental parameters Jointly determine the lower limit;

[0101] The improved NSGA-II algorithm is used to solve the Pareto optimal solution set of the multi-objective privacy-preserving optimization problem, including the following steps:

[0102] Step 3.1. Initialize the population by generating a uniformly distributed initial solution using the Latin hypercube sampling method;

[0103] Step 3.2. Evaluate individual fitness and calculate the performance index of each candidate solution on four competing objective functions;

[0104] Step 3.3. Perform a fast non-dominated sort and stratify individuals according to Pareto dominance relationships;

[0105] Step 3.4. Calculate the congestion distance;

[0106] Step 3.5. Perform genetic operations to generate a new generation of population through tournament selection, simulated binary crossover, and polynomial mutation;

[0107] Step 3.6. Apply the elite retention strategy;

[0108] Step 3.7. Determine the termination condition; stop the optimization process when the optimal solution is reached.

[0109] Based on the collaborative queue management mechanism constructed in step 2, the final privacy protection strategy is selected from the Pareto solution set: after obtaining multiple candidate solutions that are not mutually exclusive, the optimal solution is selected as the actual privacy protection strategy to be implemented based on the current running state and objective preferences; a weighted scoring method is used for multi-objective decision-making: a weighted scoring function is used to comprehensively evaluate each candidate privacy protection strategy in the Pareto solution set, as shown in the following formula:

[0110]

[0111] in, Privacy protection strategy The overall score; to These represent the normalized values ​​of four objective functions: privacy protection strength, data utility, computational cost, and communication cost; dynamic weight coefficients. to Determined by the real-time status of the system, among which, and It is negatively correlated with the Lyapunov stability coefficient γ(t) to ensure that stability is prioritized when the system is unstable. The calculation method is as follows: ;

[0112] and It is positively correlated with the Service Quality Index θ(t) of the Random Network Calculus (SNC), prioritizing data utility and communication efficiency when service quality requirements are high. The calculation method is as follows: ;

[0113] in, As the benchmark weight, 0 represents the adjustment coefficient.

[0114] The specific method for step 4 includes:

[0115] Step 4.1, Privacy Protection Policy Conversion and Command Issuance;

[0116] The privacy protection strategy selected in step 3, i.e., based on the comprehensive score, will be used to determine the privacy protection strategy. The highest Pareto solution represents the specific configuration instructions executable by the data privacy protection mechanism library. The system sends instructions to the data privacy protection mechanism library to dynamically configure the parameters of each privacy protection technology module, including:

[0117] Set a privacy budget for the differential privacy module and sensitivity parameters ;

[0118] Configure the local iteration count E and batch size B for the federated learning module;

[0119] Choose an encryption scheme and key length K for the homomorphic encryption module;

[0120] Determine the desensitization rules and retention ratio ρ for the data desensitization module;

[0121] Step 4.2: Monitor the effectiveness of the privacy protection strategy in real time;

[0122] By deploying embedded monitoring agents in various privacy protection technology modules, multi-dimensional key performance indicators (KPIs) are collected and evaluated in real time after the implementation of privacy protection policies. The monitored multi-dimensional key performance indicators (KPIs) include:

[0123] Privacy protection effectiveness: Measuring actual privacy budget consumption Compared with expected value Deviation;

[0124] Data utility level: Calculating information retention rate Correlation with features ;

[0125] System resource consumption: Record CPU utilization Memory usage and network traffic ;

[0126] Lyapunov stability index: The stability coefficient γ(t) is calculated based on the Lyapunov function value L(Θ(t)).

[0127] Random Network Calculus (SNC) Service Quality Metrics: Based on output queue overflow probability and the upper limit of latency (t);

[0128] Step 4.3, Parameter Dynamic Adjustment and Optimization Loop;

[0129] Based on the multi-dimensional key performance indicators (KPIs) obtained from step 4.2, an online learning algorithm is used to dynamically optimize the system control parameters, forming an adaptive optimization loop:

[0130] The performance index function for the control parameter V optimized for Lyapunov is calculated as follows:

[0131]

[0132] in, For average drift, For average cost, The variance of the queue length;

[0133] Calculating the SNC service quality index for random networks The performance index function is calculated as follows:

[0134]

[0135] in, This represents the actual average overflow probability. The average service quality index;

[0136] The Lyapunov control parameter V and the SNC service quality index θ are optimized using an online learning algorithm:

[0137]

[0138]

[0139] in, and This is the learning rate.

[0140] This invention also provides a dynamic data privacy protection system based on the Lyapunov-SNC collaborative computing model, comprising:

[0141] The status awareness and feature extraction module is used to collect system operation status data in real time, perform time-series smoothing processing on the collected system operation status data, and extract environmental parameters and data features.

[0142] The Lyapunov-SNC cooperative queue management controller is used to implement parallel execution of Lyapunov stability control and random network calculus (SNC) reliability assessment based on processed system operating status data; and to construct a cooperative queue management mechanism.

[0143] A multi-objective privacy policy optimizer is used to model a multi-objective privacy protection optimization problem based on a cooperative queue management mechanism, solve the Pareto solution set using the NSGA-II algorithm, and dynamically select the privacy protection policy.

[0144] The policy executor and adaptive optimization closed-loop module are used to convert the selected privacy protection policy into execution instructions, monitor system performance indicators in real time, and dynamically adjust parameters based on feedback to form a closed-loop adaptive optimization.

[0145] This invention also provides a dynamic data privacy protection device based on the Lyapunov-SNC collaborative computing model, comprising:

[0146] Memory: A computer program that stores the above-mentioned dynamic data privacy protection method based on the Lyapunov-SNC collaborative computing model, and is a computer-readable device;

[0147] Processor: Used to implement the dynamic data privacy protection method based on the Lyapunov-SNC collaborative computing model when executing the computer program.

[0148] The present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, enables the implementation of the dynamic data privacy protection method based on the Lyapunov-SNC collaborative computing model.

[0149] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0150] (1) Collaborative management architecture of Lyapunov optimization framework and stochastic network calculus (SNC). This invention innovatively constructs a queue management architecture that deeply collaborates with Lyapunov optimization and stochastic network calculus (SNC). This architecture is not a simple splicing, but rather embeds the reliability constraints (such as upper bound of delay and overflow probability boundary) derived from stochastic network calculus (SNC) into the virtual queue of Lyapunov optimization. At the mathematical level, it realizes the theoretical closed loop and unified decision-making of "dynamic stability control" and "static service quality assurance", forming a brand-new "stability-reliability" joint control paradigm.

[0151] (2) Multi-objective decision-making and strategy generation method guided by collaborative mechanism. This invention proposes a multi-objective optimization method under the dual real-time guidance of Lyapunov stability index and random network calculus SNC quality of service index. The method uses the improved NSGA-II algorithm to solve the Pareto optimal solution set of privacy protection strength, data utility, computational cost and communication cost, and selects the final execution strategy from the solution set according to the dynamic weight model (the weight coefficients are adjusted according to real-time stability and service quality requirements), thus realizing the optimal privacy-utility trade-off under multiple dynamic constraints.

[0152] (3) A full-link adaptive optimization system that integrates perception, decision-making, execution, and feedback. This invention protects a complete closed-loop system, the core of which lies in the full-link automated process of "state perception → Lyapunov / SNC collaborative decision-making → multi-objective policy optimization → policy execution and monitoring → parameter feedback adjustment". In particular, the system can dynamically adjust the Lyapunov control parameter V and the SNC service quality index θ based on the feedback of execution effect through online learning algorithms, thereby achieving continuous self-optimization of system performance.

[0153] (4) A quantifiable and verifiable multi-dimensional performance joint assurance system. This invention provides a complete quantitative evaluation and assurance system that can simultaneously output and guarantee the following core indicators: stability coefficients derived from Lyapunov functions and their derivatives, service quality index and latency upper bound calculated by the SNC model of random network calculus, as well as privacy protection strength, data utility, and resource consumption. This system makes system performance no longer a single-dimensional improvement, but a verifiable result of multi-dimensional joint optimization.

[0154] (5) Innovation of a collaborative mechanism from a single method to deep integration. Existing technologies typically employ a single optimization or evaluation method. This invention solves the problem that a single method cannot fully guarantee system performance by constructing a collaborative working mechanism between Lyapunov and Stochastic Network Calculus (SNC). This collaboration is not only a combination of modules but also achieves complementary advantages in technical characteristics: Lyapunov's real-time decision-making capability compensates for the lag in the response of SNC, while SNC's boundary warning capability prevents Lyapunov from sacrificing service quality in pursuit of stability. The two interact deeply at the information and objective levels, forming a new generation of queue management paradigm that combines dynamic control robustness with performance guarantee determinism, representing a breakthrough combination of theory and practice.

[0155] In summary, compared with existing technologies, this invention achieves a closed-loop unification of dynamic stability control and static service quality assurance through the synergistic integration of Lyapunov optimization and Stochastic Network Calculus (SNC). It constructs a multi-objective decision-making and end-to-end adaptive optimization system based on a collaborative mechanism, which significantly improves the system's dynamic adaptability, privacy-utility trade-off capability, and cross-scenario engineering application value while ensuring strict service quality boundaries. Attached Figure Description

[0156] Figure 1 This is a diagram illustrating the overall architecture of the dynamic data privacy protection system of this invention.

[0157] Figure 2 This is a flowchart illustrating the implementation of the method of the present invention.

[0158] Figure 3 This is a detailed flowchart of the Lyapunov-SNC collaborative queue management mechanism of the present invention.

[0159] Figure 4 This is a flowchart of the multi-objective privacy strategy optimization and NSGA-II algorithm of the present invention.

[0160] Figure 5 This is a flowchart illustrating the closed-loop process of strategy execution and adaptive optimization in this invention.

[0161] Figure 6 This is a flowchart of the Lyapunov virtual queue construction and update process of the present invention. Detailed Implementation

[0162] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0163] This invention proposes a dynamic data privacy protection system based on the Lyapunov-SNC collaborative computing model. The overall architecture and workflow of the system are as follows: Figure 1 and Figure 2 As shown, the system includes: a state awareness and feature extraction module, a Lyapunov-SNC collaborative queue management controller, a multi-objective privacy policy optimizer, and a multi-objective privacy policy optimizer. The system workflow consists of four main stages: state awareness and feature extraction, Lyapunov-SNC collaborative decision-making, multi-objective privacy policy optimization, and policy execution and feedback adjustment, forming a closed-loop adaptive optimization system.

[0164] The core function of the Lyapunov optimization framework:

[0165] The Lyapunov optimization framework is specifically responsible for the dynamic stability control of the system. By constructing Lyapunov functions and minimizing drift and cost, Lyapunov optimization ensures that the system remains stable in long-term operation, preventing the queue from growing indefinitely and the system from crashing. This feature makes up for the shortcomings of existing stochastic network calculus (SNC) schemes in terms of dynamic control capabilities.

[0166] The core role of Stochastic Network Calculus (SNC):

[0167] The Stochastic Network Calculus (SNC) is specifically responsible for the quantitative assessment and assurance of service quality. By constructing a queue overflow probability model and calculating the service quality index, the SNC provides the system with strict service quality boundary guarantees. This feature makes up for the shortcomings of the existing Lyapunov scheme in terms of service quality quantitative assessment capabilities.

[0168] A dynamic data privacy protection method based on the Lyapunov-SNC collaborative computing model includes the following steps:

[0169] Step 1: Collect system operation status data in real time, perform time-series smoothing on the collected system operation status data, and extract environmental parameters and data features;

[0170] The specific method of step 1 includes:

[0171] Step 1.1: Collect and monitor system operation status data in real time;

[0172] Status data for the following queues is continuously collected through a distributed sensor network deployed at edge nodes:

[0173] Data queue length , representing the total amount of user data waiting to be processed in time slot t, in bits or number of tasks;

[0174] Privacy demand queue The quantitative accumulation of privacy protection requirements is calculated as follows:

[0175]

[0176] in, To meet the privacy protection requirements of user i in time slot t, This represents the amount of privacy protection already satisfied;

[0177] Utility demand queue This reflects the data availability maintenance requirements and is calculated as follows:

[0178]

[0179] in, For user i's utility needs in time slot t, The amount of utility already realized is retained.

[0180] Step 1.2: Use a time-series smoothing method to handle queue fluctuations;

[0181] The queue state data collected in step 1.1 is preprocessed using time series analysis to eliminate transient noise interference and improve the accuracy of state estimation, resulting in processed queue state data; the sliding window size W is adaptively adjusted according to the dynamic characteristics of the system.

[0182] Step 1.3: Extract multi-dimensional environmental parameters and data features;

[0183] By using multi-source information fusion technology, system operating environment parameters, including computing resource availability, are collected. Network bandwidth status Security Threat Level and real-time performance requirements This provides environmental context for optimizing decision-making; simultaneously, a lightweight machine learning model is used to analyze the characteristics of the data to be processed, including data sensitivity. Data scale Structural complexity and real-time requirements This provides a basis for selecting privacy protection mechanisms.

[0184] Step 2: Based on the system operating status data processed in Step 1, perform Lyapunov stability control and stochastic network calculus (SNC) reliability assessment in parallel; and construct a collaborative queue management mechanism.

[0185] The specific method for step 2 includes:

[0186] Step 2.1: Construct the Lyapunov optimization framework and solve for the stability control strategy;

[0187] Based on the queue state data processed in step 1.2, the queue state data is mapped to the unified state vector required for Lyapunov optimization. The calculation method is as follows:

[0188]

[0189] in, For user queues;

[0190] Edge server queue Determined by the length of the relevant data queue The aggregation yields the result, calculated as follows:

[0191]

[0192] in, This refers to the weighting factor for user u, which is based on the proportion of computing resources reserved as specified in user u's Service Level Agreement (SLA). or minimum guaranteed bandwidth Dynamically determined; the calculation formula is:

[0193] Based on the proportion of reserved computing resources ,but:

[0194]

[0195] Based on the minimum guaranteed bandwidth ,but:

[0196]

[0197] in, For the set of all users assigned to edge server b, This represents the percentage of computing resources reserved for user u as agreed in the Service Level Agreement (SLA), with a value range of [value range missing]. , This represents the minimum guaranteed bandwidth agreed upon by user u in the Service Level Agreement (SLA).

[0198] The above weighting calculation method ensures that the weight of each user in the aggregation queue is proportional to the resource guarantee agreed upon in its Service Level Agreement (SLA), thereby guaranteeing its service quality in subsequent Lyapunov optimizations.

[0199] Equivalent persistent queues to guarantee worst-case latency The ε-persistent queue length of its next time slot The calculation formula is:

[0200]

[0201] in, As a preset constant, For the service rate of server b, To discard task quantity, For the system's maximum service rate, This is an indicator function; it returns 1 when the condition is true and 0 otherwise.

[0202] Virtual reliability queue for handling reliability constraints derived from random network calculus (SNC). The length of the virtual reliability queue in its next time slot The calculation formula is:

[0203]

[0204] in, The probability of queue overflow for user i is calculated using the following formula:

[0205]

[0206] in, The target service quality index for user i is determined by the Service Level Agreement (SLA) or application scenario requirements. Let be the maximum allowed task processing latency for user i; e is the base of the natural logarithm.

[0207] The probability estimate of queue overflow for user i in time slot t is calculated using the following formula:

[0208]

[0209] in, For the size of the historical observation window, For indicator functions;

[0210] Based on unified state vector Construct a quadratic Lyapunov function:

[0211]

[0212] Calculate single-slot conditional Lyapunov drift: used to predict the trend of system stability changes;

[0213]

[0214] Solve the single-slot conditional Lyapunov drift plus cost minimization problem: determine the system stability control strategy;

[0215]

[0216] Where, ∑ t) represents the total task delay, and V is a Lyapunov control parameter used to balance stability and performance. It is calculated as follows:

[0217]

[0218] in, , , The maximum capacity of the queue. For maximum cost, the Lyapunov stability coefficient γ(t) is the negative of the rate of change of the Lyapunov function in adjacent time slots, used to quantify the speed at which the system tends to stability:

[0219]

[0220] , It is a range of The positive adjustment parameter is used to ensure that the denominator is not zero. ;

[0221] Step 2.2: Model and quantify service quality based on random network calculus (SNC);

[0222] Using arrival curve Service Line Modeling the system queue behavior:

[0223]

[0224] in, This represents the upper bound of the cumulative arrival flow from time 0 to t, used to characterize the randomness and burstiness of data arrival; This represents the lower bound of the cumulative service capacity from time 0 to t, used to characterize the uncertainty of the system's processing capacity; The value represents the service quality index, reflecting the stringency of the system's requirements for service quality; the higher the value, the more stringent the requirements. Indicates in time slot The amount of data arriving within the region; Indicates in time slot The internal system provides the service rate for user i on server b, and E[] represents the expected operation, which is used to handle randomness;

[0225] Calculate the service quality index θ(t) to characterize the system's ability to meet service quality requirements:

[0226]

[0227] Where θ(t) is a dynamic service quality index. The larger the value, the lower the system's tolerance for latency and queue overflow, and the higher the service quality requirements. This is the upper bound of the queue overflow probability, i.e., the maximum queue overflow probability allowed by the system, calculated using the following formula: ; The maximum allowed length of the queue is used for normalization, making θ(t) dimensionless and dynamically adjustable.

[0228] Step 2.3: Construct a collaborative queue management mechanism;

[0229] Step 2.3.1, construct a virtual reliability queue:

[0230] Based on the target queue overflow probability obtained in step 2.1 Estimation of queue overflow probability Build a virtual reliability queue The calculation formula is:

[0231]

[0232] Step 2.3.2, Dynamically adjust the service quality index:

[0233] Based on the Lyapunov stability coefficient γ(t) and target queue overflow probability obtained in step 2.1 Estimation of queue overflow probability The difference will be used to dynamically adjust the service quality index in step 2.2. The formula is adjusted as follows:

[0234]

[0235] Step 2.3.3, construct the collaborative optimization objective function:

[0236] Based on the minimization of Lyapunov drift plus cost in step 2.1 and the sum of the squares of the differences between the estimated queue overflow probabilities and the queue overflow probabilities of all objectives, a unified collaborative optimization objective function is constructed:

[0237]

[0238] in, , which is a weighting coefficient used to balance stability control objectives and service quality assurance objectives.

[0239] Step 3: Based on the collaborative queue management mechanism built in Step 2, model a multi-objective privacy protection optimization problem, use the NSGA-II algorithm to solve the Pareto solution set of the multi-objective privacy protection optimization problem, and dynamically select a privacy protection strategy.

[0240] The specific method for step 3 includes:

[0241] Based on the collaborative queue management mechanism constructed in step 2, the privacy protection problem is modeled as a multi-objective privacy protection optimization problem; four competing objective functions are defined, including: privacy protection strength objective function f1(x), data utility objective function f2(x), computational cost objective function, and communication cost objective function f4(x); where:

[0242] The objective function f1(x) for privacy protection strength is calculated as follows:

[0243]

[0244] in, For differential privacy budgeting, For homomorphic encryption strength, To learn about the level of privacy protection in federal education;

[0245] The objective function f2(x) for data utility is calculated as follows:

[0246]

[0247] The objective function f3(x) for calculating the cost is calculated as follows:

[0248]

[0249] The objective function for communication overhead, f4(x), is calculated as follows:

[0250] ;

[0251] Design a set of constraints to ensure that the optimization results meet the actual operation requirements: {computational resource constraints, bandwidth constraints, latency constraints, privacy strength constraints};

[0252] Computational resource constraints: Utilizing the computational resource availability extracted in step 1.3 As an upper bound constraint;

[0253] Bandwidth constraints: Utilizing the network bandwidth status extracted in step 1.3 As an upper bound constraint;

[0254] Delay constraints: Real-time requirements extracted from step 1.3 Sure;

[0255] Privacy strength constraints: Data sensitivity extracted from step 1.3 Security threat levels in environmental parameters Jointly determine the lower limit;

[0256] An improved NSGA-II algorithm is used to solve the Pareto optimal solution set of a multi-objective privacy-preserving optimization problem. The multi-objective privacy-preserving optimization problem is defined as having four conflicting objectives, and its solution is not a single optimal solution, but a set of compromise solutions that cannot be mutually improved, called the Pareto optimal solution set. The steps include:

[0257] Step 3.1. Initialize the population by generating a uniformly distributed initial solution using the Latin hypercube sampling method;

[0258] Step 3.2. Evaluate individual fitness and calculate the performance index of each candidate solution on four competing objective functions;

[0259] Step 3.3. Perform a fast non-dominated sort and stratify individuals according to Pareto dominance relationships;

[0260] Step 3.4. Calculate the crowding distance to ensure the uniformity of the solution set distribution in the target space;

[0261] Step 3.5. Perform genetic operations to generate a new generation of population through tournament selection, simulated binary crossover, and polynomial mutation;

[0262] Step 3.6. Apply an elite preservation strategy to prevent superior individuals from being lost during the evolutionary process;

[0263] Step 3.7. Determine the termination condition; stop the optimization process when the optimal solution is reached.

[0264] Based on the collaborative queue management mechanism constructed in step 2, the final privacy protection strategy is selected from the Pareto solution set: After obtaining multiple candidate solutions that are not mutually exclusive, the optimal solution is selected as the actual privacy protection strategy to be implemented based on the current running state and objective preferences; a weighted scoring method is used for multi-objective decision-making: to achieve a quantitative trade-off between different objectives, a weighted scoring function is used to comprehensively evaluate each candidate privacy protection strategy in the Pareto solution set, as shown in the following formula:

[0265]

[0266] in, Privacy protection strategy The overall score; to These represent the normalized values ​​of four objective functions: privacy protection strength, data utility, computational cost, and communication cost; dynamic weight coefficients. to Determined by the real-time status of the system, among which, and It is negatively correlated with the Lyapunov stability coefficient γ(t) to ensure that stability is prioritized when the system is unstable. The calculation method is as follows: ;

[0267] and It is positively correlated with the Service Quality Index θ(t) of the Random Network Calculus (SNC), prioritizing data utility and communication efficiency when service quality requirements are high. The calculation method is as follows: .

[0268] in, As the benchmark weight, 0 represents the adjustment coefficient.

[0269] Step 4: Convert the privacy protection strategy selected in Step 3 into execution instructions, monitor system performance indicators in real time, and dynamically adjust parameters based on feedback to form a closed-loop adaptive optimization.

[0270] The specific method for step 4 includes:

[0271] Step 4.1, Privacy Protection Policy Conversion and Command Issuance;

[0272] The privacy protection strategy selected in step 3, i.e., based on the comprehensive score, will be used to determine the privacy protection strategy. The highest Pareto solution represents the specific configuration instructions executable by the data privacy protection mechanism library. The system sends instructions to the data privacy protection mechanism library to dynamically configure the parameters of each privacy protection technology module, including:

[0273] Set a privacy budget for the differential privacy module and sensitivity parameters ;

[0274] Configure the local iteration count E and batch size B for the federated learning module;

[0275] Choose an encryption scheme and key length K for the homomorphic encryption module;

[0276] Determine the desensitization rules and retention ratio ρ for the data desensitization module;

[0277] Step 4.2: Monitor the effectiveness of the privacy protection strategy in real time;

[0278] By deploying embedded monitoring agents in various privacy protection technology modules, multi-dimensional key performance indicators (KPIs) are collected and evaluated in real time after the implementation of privacy protection policies, providing a data foundation for parameter adjustment. The monitored multi-dimensional KPIs include:

[0279] Privacy protection effectiveness: Measuring actual privacy budget consumption Compared with expected value Deviation;

[0280] Data utility level: Calculating information retention rate Correlation with features ;

[0281] System resource consumption: Record CPU utilization Memory usage and network traffic ;

[0282] Lyapunov stability index: The stability coefficient γ(t) is calculated based on the Lyapunov function value L(Θ(t)).

[0283] Random Network Calculus (SNC) Service Quality Metrics: Based on output queue overflow probability and the upper limit of latency (t);

[0284] Step 4.3, Parameter Dynamic Adjustment and Optimization Loop;

[0285] Based on the multi-dimensional key performance indicators (KPIs) obtained from step 4.2, online learning algorithms such as gradient descent are used to dynamically optimize the system control parameters, forming an adaptive optimization loop:

[0286] The performance index function for the control parameter V optimized for Lyapunov is calculated as follows:

[0287]

[0288] in, For average drift, For average cost, The variance of the queue length;

[0289] Calculating the SNC service quality index for random networks The performance index function is calculated as follows:

[0290]

[0291] in, This represents the actual average overflow probability. The average service quality index;

[0292] The gradient descent method is used to optimize the Lyapunov control parameter V and the SNC service quality index θ is calculated using a stochastic network.

[0293]

[0294]

[0295] in, and This is the learning rate.

[0296] This invention also provides a dynamic data privacy protection system based on the Lyapunov-SNC collaborative computing model, comprising:

[0297] The state awareness and feature extraction module is used to collect system operating status data in real time in step 1, perform time-series smoothing on the collected system operating status data, and extract environmental parameters and data features.

[0298] The Lyapunov-SNC collaborative queue management controller is used to implement Lyapunov stability control and random network calculus SNC reliability assessment in parallel based on the system operating status data processed in step 1 in step 2; and to construct a collaborative queue management mechanism.

[0299] A multi-objective privacy policy optimizer is used to implement the collaborative queue management mechanism built in step 2 in step 3, model the multi-objective privacy protection optimization problem, solve the Pareto solution set using the NSGA-II algorithm, and dynamically select the privacy protection policy.

[0300] The policy executor and adaptive optimization closed-loop module are used to convert the privacy protection policy selected in step 3 into execution instructions in step 4, monitor system performance indicators in real time, and dynamically adjust parameters based on feedback to form a closed-loop adaptive optimization.

[0301] This invention also provides a dynamic data privacy protection device based on the Lyapunov-SNC collaborative computing model, comprising:

[0302] Memory: A computer program that stores the above-mentioned dynamic data privacy protection method based on the Lyapunov-SNC collaborative computing model, and is a computer-readable device;

[0303] Processor: Used to implement the dynamic data privacy protection method based on the Lyapunov-SNC collaborative computing model when executing the computer program.

[0304] The present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, enables the implementation of the dynamic data privacy protection method based on the Lyapunov-SNC collaborative computing model.

[0305] like Figure 1 The diagram shown illustrates the overall architecture of the dynamic data privacy protection system of this invention. It illustrates the four core modules of the system and their interactions, employing a closed-loop control design. The system begins with the state awareness and feature extraction module, achieving joint control of stability and reliability through the Lyapunov-SNC collaborative queue management controller. Then, a multi-objective privacy policy optimizer selects the optimal privacy protection policy. Finally, the policy executor and adaptive optimization closed-loop module execute the policy and collect performance feedback, forming a complete adaptive optimization loop. This architecture ensures that the system can continuously self-optimize in dynamic environments while maintaining a balance between data privacy and system performance.

[0306] like Figure 2 The diagram shows a detailed implementation flowchart of the dynamic privacy protection system in this invention. This flowchart elaborates on the four main implementation stages of the system and their internal logic. The process begins with system state awareness and feature extraction, then deeply integrates stability control and service quality assessment through the Lyapunov-SNC collaborative queue management mechanism. Next, the NSGA-II multi-objective optimization algorithm solves for the Pareto optimal solution set of the privacy protection strategy. Finally, the strategy is executed, and parameters are adaptively adjusted based on performance feedback. The process includes key decision steps to determine whether to continue system operation, ensuring continuous optimization in a dynamic environment.

[0307] like Figure 3 The diagram shows a detailed flowchart of the Lyapunov-SNC collaborative queue management mechanism in this invention. This flowchart elaborates on the internal implementation logic of the collaborative work between the Lyapunov optimization framework and the Stochastic Network Calculus (SNC). The process begins by receiving system state data, quantifying the system's stability state by constructing Lyapunov functions, evaluating service quality indicators through SNC modeling, embedding reliability constraints derived from SNC into the Lyapunov virtual queue, and dynamically adjusting SNC parameters based on Lyapunov stability guidelines. Finally, a collaborative queue management strategy is output. This mechanism ensures that the system can simultaneously guarantee stability and service quality during long-term operation.

[0308] like Figure 4The diagram shows a detailed flowchart of the multi-objective privacy strategy optimization and the NSGA-II algorithm in this invention. This flowchart elaborates on the internal implementation logic of the NSGA-II multi-objective optimization algorithm. The process begins with initializing the population, determining the quality of individuals through evaluation of fitness, fast non-dominated sorting, and crowding calculation. Then, a new generation of the population is generated through selection, crossover, and mutation operations. An elite retention strategy is applied to prevent the loss of superior individuals. Finally, when the termination condition is met, the Pareto optimal solution set is output. This algorithm effectively balances multiple competing objectives such as privacy protection strength, data utility, computational overhead, and communication overhead, providing various trade-offs for privacy protection strategy selection.

[0309] like Figure 5 The diagram shown is a detailed flowchart of the policy execution and adaptive optimization closed loop in this invention. This flowchart elaborates on the internal implementation logic of policy execution and parameter adaptive adjustment. The process begins with receiving the optimal privacy protection policy, configuring various privacy protection technology modules through policy conversion and instruction issuance, monitoring privacy protection effectiveness, data utility level, system resource consumption, stability, and service indicators in real time, dynamically adjusting Lyapunov control parameters and the SNC service quality index based on monitoring data, and sending feedback data back to the state awareness module, forming a complete "perception-decision-execution-optimization" closed-loop control process, enabling the system to continuously self-optimize.

[0310] like Figure 6 The diagram shows a detailed flowchart of the construction and updating of Lyapunov virtual queues in this invention. This flowchart elaborates on the logic of virtual queue construction and updating within the Lyapunov optimization framework. The process begins with input user queue data, then aggregates user queues to construct edge server queues, simultaneously constructs ε-persistent queues to ensure worst-case latency, and virtual reliability queues to handle SNC reliability constraints in random network calculus. Next, a unified state vector is constructed, and finally, the states of various queues are synchronously updated according to the queue update logic, and the update results are output. This process provides a complete system state representation for Lyapunov optimization, ensuring the accuracy and real-time performance of stability control.

[0311] Experimental Analysis

[0312] 1. Experimental conditions

[0313] 1.1 Experimental Environment Configuration

[0314] Hardware platform:

[0315] Edge servers: 4 units, configured with Intel Xeon Gold 6248R (3.0GHz, 24 cores), 256GB RAM; User equipment: 100 units, simulating mobile devices (CPU: 8 cores, RAM: 16GB); Network equipment: Cisco Catalyst 9300 switches, supporting 10Gbps links;

[0316] Software environment:

[0317] Operating System: Ubuntu 20.04 LTS; Virtualization Platform: Docker 20.10, Kubernetes 1.23; Development Framework: Python 3.9, PyTorch 1.12, NS-3 network simulator;

[0318] Dataset:

[0319] Smart grid data: 10,000 user electricity consumption records, including sensitive information such as time, power, and equipment type; medical and health data: 5,000 anonymous patient records, including diagnosis, medication, and examination results; network traffic data: collected from real 5G base stations, containing 1 million data packet records.

[0320] 1.2 Experimental Parameter Settings

[0321]

[0322] 1.3 Comparison Methods

[0323] (1) Single Lyapunov method (Lyapunov-Only): Only Lyapunov optimization is used for dynamic control, without random network calculus SNC reliability guarantee;

[0324] (2) Single Random Network Calculus (SNC) method: Only random network calculus (SNC) is used for service quality assessment, without dynamic stability control;

[0325] (3) Static-Comb method: The Lyapunov algorithm with fixed weights is combined with the random network calculus (SNC) without adaptive adjustment;

[0326] (4) Traditional privacy protection method (Traditional-PP): differential privacy or homomorphic encryption based on fixed parameters;

[0327] (5) The method of the present invention: a dynamic privacy protection method based on Lyapunov-SNC collaboration.

[0328] 2. Experiment Content

[0329] 2.1 Experiment 1: Verification of the effectiveness of the collaborative mechanism

[0330] Objective: To verify the advantages of Lyapunov working in conjunction with Stochastic Network Calculus (SNC);

[0331] Scenario: Under dynamic load conditions, compare the system stability and service quality of different methods;

[0332] index:

[0333] Queue stability coefficient Average time; queue overflow probability Actual measured values; service reliability compliance rate ( (Time proportion).

[0334] 2.2 Experiment 2: Evaluation of Privacy Protection Effectiveness

[0335] Objective: To assess the balance between the strength of privacy protection and data availability;

[0336] Scenario: For different types of sensitive data (power grid, medical, communications);

[0337] index:

[0338] Privacy protection strength: measured by the success rate of member inference attacks; Data availability: model prediction accuracy / data utility retention rate; Privacy-utility trade-off curve (P-UT trade-off curve).

[0339] 2.3 Experiment 3: Dynamic Environment Adaptability Test

[0340] Objective: To test the adaptability of the method under sudden traffic surges and attack threats.

[0341] Scene:

[0342] Phase 1 (slots 0-100): Normal load; Phase 2 (slots 101-200): Sudden traffic surge, load increases by 300%; Phase 3 (slots 201-300): Privacy attack, threat level escalates;

[0343] Metrics: System response time; parameter adjustment speed; service quality retention rate.

[0344] 2.4 Experiment 4: Algorithm Convergence and Complexity Analysis

[0345] Objective: To analyze the convergence and computational complexity of the algorithm.

[0346] Methods: Record the change of the objective function value with the number of iterations; measure the average execution time of a single decision; analyze the computational complexity of each module of the algorithm;

[0347] index:

[0348] Convergence speed (number of iterations required to reach 90% of the optimal solution); decision latency (total time from perception to execution); CPU and memory utilization.

[0349] 3. Experimental Results

[0350] 3.1 Comparison of Collaboration Mechanism Performance and Privacy Protection Effect

[0351] Table 1 Performance Comparison of Coordination Mechanisms

[0352]

[0353] As shown in Table 1, the method of the present invention reduces the queue length by 45.4% compared to Lyapunov-Only and improves the stability by 116% compared to SNC-Only. It can be seen that the method of the present invention achieves significant improvements in both queue length and stability, while approaching the best level of SNC-Only in terms of service quality.

[0354] Table 2 Comparison of Privacy Protection Effects

[0355] Data categories method Attack success rate (%) Data availability (%) Overall score Smart grid data Traditional-PP 15.3 78.2 0.634 Smart grid data Proposed 8.7 85.6 0.785 Healthcare data Traditional-PP 12.8 71.5 0.593 Healthcare data Proposed 6.4 79.3 0.724 Communication traffic data Traditional-PP 18.2 82.4 0.641 Communication traffic data Proposed 10.5 88.1 0.777

[0356] As shown in Table 2, the method of the present invention achieves a better privacy-utility trade-off across various types of data, with an average reduction in attack success rate of 42.5% and an average increase in data availability of 8.7%.

[0357] 3.2 Results of Dynamic Environmental Adaptability

[0358] Adaptive parameter adjustment of response time:

[0359] Normal → Burst Traffic: 3.2±0.5 time slots; Burst Traffic → Attack Status: 2.8±0.4 time slots; Attack Status → Return to Normal: 4.1±0.6 time slots;

[0360] Service quality retention rate:

[0361] Phase 1 (Normal): 96.3%; Phase 2 (Burst Traffic): 92.7% (Other methods: 68.4-85.2%); Phase 3 (Attack Status): 88.5% (Other methods: 52.1-73.8%).

[0362] In summary, the method of the present invention exhibits excellent adaptability in dynamic environments, especially in maintaining 88.5% of service quality even under extreme conditions.

[0363] 3.3 Algorithm Performance Analysis

[0364] Computational complexity analysis:

[0365] Lyapunov optimization module: O(B·I) per iteration; Stochastic Network Calculus (SNC) evaluation module: O(I·logI) per iteration; NSGA-II optimization module: O(P·G·M²), P=100, G=50, M=4.

[0366] Total decision delay: 15.3 ± 2.1 ms (meets real-time requirements)

[0367] Convergence analysis:

[0368] Objective function convergence speed: 15-20 iterations (90% optimal); parameter stability: fluctuation <1% after 50 iterations; Pareto front homogeneity: coverage 0.87 (ideally 1.0).

[0369] 3.4 Key Findings and Strengths Summary

[0370] Significant synergistic gains: Compared with the single method, the Lyapunov-SNC synergy improves stability by 35.8% and service quality by 40.2%.

[0371] Outstanding adaptability: In dynamic environments, the method of this invention improves the service quality retention rate by 23.4% compared to static methods.

[0372] Privacy-utility balance optimization: Under the same level of privacy protection, data availability is improved by 8.7%; under the same level of data availability, the level of privacy protection is improved by 42.5%.

[0373] Real-time performance requirements are met: the average decision latency is 15.3ms, which is suitable for real-time data processing scenarios.

[0374] Strong cross-scenario generalization ability: It performs stably in three scenarios: smart grid, healthcare, and communication traffic.

[0375] 3.5 Parameter Sensitivity Analysis

[0376]

Claims

1. A dynamic data privacy protection method based on the Lyapunov-SNC collaborative computing model, characterized in that, Includes the following steps: Step 1: Collect system operating status data in real time, perform time-series smoothing on the collected system operating status data, and extract environmental parameters and data features. The specific steps are as follows: Step 1.1: Collect and monitor system operation status data in real time; Status data for the following queues is continuously collected through a distributed sensor network deployed at edge nodes: Data queue length , representing the total amount of user data waiting to be processed in time slot t, in bits or number of tasks; Privacy demand queue The quantitative accumulation of privacy protection requirements is calculated as follows: in, To meet the privacy protection requirements of user i in time slot t, This represents the amount of privacy protection already satisfied; Utility demand queue This reflects the data availability maintenance requirements and is calculated as follows: in, For user i's utility needs in time slot t, To retain the amount of utility already realized; Step 1.2: Use a time-series method to smooth queue fluctuations; The queue status data collected in step 1.1 is preprocessed using time series analysis to eliminate instantaneous noise interference, resulting in processed queue status data; the window size for smoothing is adaptively adjusted according to the dynamic characteristics of the system. Step 1.3: Extract multi-dimensional environmental parameters and data features; By using multi-source information fusion technology, system operating environment parameters, including computing resource availability, are collected. Network bandwidth status Security Threat Level and real-time performance requirements Simultaneously, a lightweight machine learning model is used to analyze the characteristics of the data to be processed, including data sensitivity. Data scale Structural complexity and real-time requirements ; Step 2: Based on the system operating status data processed in Step 1, perform Lyapunov stability control and stochastic network calculus (SNC) reliability assessment in parallel; and construct a collaborative queue management mechanism. Step 3: Based on the collaborative queue management mechanism built in Step 2, model a multi-objective privacy protection optimization problem, use the NSGA-II algorithm to solve the Pareto solution set of the multi-objective privacy protection optimization problem, and dynamically select a privacy protection strategy. Step 4: Convert the privacy protection strategy selected in Step 3 into execution instructions, monitor system performance indicators in real time, and dynamically adjust parameters based on feedback to form a closed-loop adaptive optimization.

2. The dynamic data privacy protection method based on the Lyapunov-SNC collaborative computing model according to claim 1, characterized in that, The specific method for step 2 includes: Step 2.1: Construct the Lyapunov optimization framework and solve for the stability control strategy; Based on the queue state data processed in step 1.2, the queue state data is mapped to the unified state vector required for Lyapunov optimization. The calculation method is as follows: in, For user queues; Edge server queue Determined by the length of the relevant data queue The aggregation yields the result, calculated as follows: in, This refers to the weighting factor for user u, which is based on the proportion of computing resources reserved as specified in user u's Service Level Agreement (SLA). or minimum guaranteed bandwidth Dynamically determined; the calculation formula is: Based on the proportion of reserved computing resources ,but: Based on the minimum guaranteed bandwidth ,but: in, For the set of all users assigned to edge server b, This represents the percentage of computing resources reserved for user u as agreed in the Service Level Agreement (SLA), with a value range of [value range missing]. , This represents the minimum guaranteed bandwidth agreed upon by user u in the Service Level Agreement (SLA). Equivalent persistent queues to guarantee worst-case latency The ε-persistent queue length of its next time slot The calculation formula is: in, As a preset constant, For the service rate of server b, To discard task quantity, For the system's maximum service rate, This is an indicator function; it returns 1 when the condition is true and 0 otherwise. Virtual reliability queue for handling reliability constraints derived from random network calculus (SNC). The length of the virtual reliability queue in its next time slot The calculation formula is: in, The probability of queue overflow for user i is calculated using the following formula: in, The target service quality index for user i is determined by the Service Level Agreement (SLA) or application scenario requirements. Let be the maximum allowed task processing latency for user i; e is the base of the natural logarithm. The probability estimate of queue overflow for user i in time slot t is calculated using the following formula: in, For the size of the historical observation window, For indicator functions; Based on unified state vector Construct a quadratic Lyapunov function: Calculate single-slot conditional Lyapunov drift: Solve the single-slot conditional Lyapunov drift plus cost minimization problem: Where, ∑ t) represents the total task delay, and V is a Lyapunov control parameter used to balance stability and performance. It is calculated as follows: in, , , The maximum capacity of the queue. For maximum cost, the Lyapunov stability coefficient γ(t) is the negative of the rate of change of the Lyapunov function in adjacent time slots, used to quantify the speed at which the system tends to stability: , It is a range of The positive adjustment parameter is used to ensure that the denominator is not zero. ; Step 2.2: Model and quantify service quality based on random network calculus (SNC); Using arrival curve Service Line Modeling the system queue behavior: in, This represents the upper bound of the cumulative arrival flow from time 0 to t, used to characterize the randomness and burstiness of data arrival; This represents the lower bound of the cumulative service capacity from time 0 to t, used to characterize the uncertainty of the system's processing capacity; The value represents the service quality index, reflecting the stringency of the system's requirements for service quality; the higher the value, the more stringent the requirements. Indicates in time slot The amount of data arriving within the region; Indicates in time slot The internal system provides the service rate for user i on server b, and E[] represents the expected operation, which is used to handle randomness; Calculate the service quality index θ(t) to characterize the system's ability to meet service quality requirements: Where θ(t) is a dynamic service quality index. The larger the value, the lower the system's tolerance for latency and queue overflow, and the higher the service quality requirements. This is the upper bound of the queue overflow probability, i.e., the maximum queue overflow probability allowed by the system, calculated using the following formula: ; The maximum allowed length of the queue is used for normalization, making θ(t) dimensionless and dynamically adjustable. Step 2.3: Construct a collaborative queue management mechanism; Step 2.3.1, construct a virtual reliability queue: Based on the target queue overflow probability obtained in step 2.1 Estimation of queue overflow probability Build a virtual reliability queue The calculation formula is: Step 2.3.2, Dynamically adjust the service quality index: Based on the Lyapunov stability coefficient γ(t) and target queue overflow probability obtained in step 2.1 Estimation of queue overflow probability The difference will be used to dynamically adjust the service quality index in step 2.

2. The formula is adjusted as follows: Step 2.3.3, construct the collaborative optimization objective function: Based on the minimization of Lyapunov drift plus cost in step 2.1 and the sum of the squares of the differences between the estimated queue overflow probabilities and the queue overflow probabilities of all objectives, a unified collaborative optimization objective function is constructed: in, , which is a weighting coefficient used to balance stability control objectives and service quality assurance objectives.

3. The dynamic data privacy protection method based on the Lyapunov-SNC collaborative computing model according to claim 1, characterized in that, The specific method for step 3 includes: Based on the collaborative queue management mechanism constructed in step 2, the privacy protection problem is modeled as a multi-objective privacy protection optimization problem; four competing objective functions are defined, including: privacy protection strength objective function f1(x), data utility objective function f2(x), computational cost objective function, and communication cost objective function f4(x); where: The objective function f1(x) for privacy protection strength is calculated as follows: in, For differential privacy budgeting, For homomorphic encryption strength, To learn about the level of privacy protection in federal education; The objective function f2(x) for data utility is calculated as follows: The objective function f3(x) for calculating the cost is calculated as follows: The objective function for communication overhead, f4(x), is calculated as follows: ； The set of design constraints includes: {computational resource constraints, bandwidth constraints, latency constraints, and privacy strength constraints}. Computational resource constraints: Utilizing the computational resource availability extracted in step 1.3 As an upper bound constraint; Bandwidth constraints: Utilizing the network bandwidth status extracted in step 1.3 As an upper bound constraint; Delay constraints: Real-time requirements extracted from step 1.3 Sure; Privacy strength constraints: Data sensitivity extracted from step 1.3 Security threat levels in environmental parameters Jointly determine the lower limit; The improved NSGA-II algorithm is used to solve the Pareto optimal solution set of the multi-objective privacy-preserving optimization problem, including the following steps: Step 3.

1. Initialize the population by generating a uniformly distributed initial solution using the Latin hypercube sampling method; Step 3.

2. Evaluate individual fitness and calculate the performance index of each candidate solution on four competing objective functions; Step 3.

3. Perform a fast non-dominated sort and stratify individuals according to Pareto dominance relationships; Step 3.

4. Calculate the congestion distance; Step 3.

5. Perform genetic operations to generate a new generation of population through tournament selection, simulated binary crossover, and polynomial mutation; Step 3.

6. Apply the elite retention strategy; Step 3.

7. Determine the termination condition; stop the optimization process when the optimal solution is reached. Based on the collaborative queue management mechanism constructed in step 2, the final privacy protection strategy is selected from the Pareto solution set: after obtaining multiple candidate solutions that are not mutually exclusive, the optimal solution is selected as the actual privacy protection strategy to be implemented based on the current running state and objective preferences; a weighted scoring method is used for multi-objective decision-making: a weighted scoring function is used to comprehensively evaluate each candidate privacy protection strategy in the Pareto solution set, as shown in the following formula: in, Privacy protection strategy The overall score; to These represent the normalized values ​​of four objective functions: privacy protection strength, data utility, computational cost, and communication cost; dynamic weight coefficients. to Determined by the real-time status of the system, among which, and It is negatively correlated with the Lyapunov stability coefficient γ(t) to ensure that stability is prioritized when the system is unstable. The calculation method is as follows: ; and It is positively correlated with the Service Quality Index θ(t) of the Random Network Calculus (SNC), prioritizing data utility and communication efficiency when service quality requirements are high. The calculation method is as follows: ; in, As the benchmark weight, 0 represents the adjustment coefficient.

4. The dynamic data privacy protection method based on the Lyapunov-SNC collaborative computing model according to claim 1, characterized in that, The specific method for step 4 includes: Step 4.1, Privacy Protection Policy Conversion and Command Issuance; The privacy protection strategy selected in step 3, i.e., based on the comprehensive score, will be used to determine the privacy protection strategy. The highest Pareto solution represents the specific configuration instructions executable by the data privacy protection mechanism library. The system sends instructions to the data privacy protection mechanism library to dynamically configure the parameters of each privacy protection technology module, including: Set a privacy budget for the differential privacy module and sensitivity parameters ; Configure the local iteration count E and batch size B for the federated learning module; Choose an encryption scheme and key length K for the homomorphic encryption module; Determine the desensitization rules and retention ratio ρ for the data desensitization module; Step 4.2: Monitor the effectiveness of the privacy protection strategy in real time; By deploying embedded monitoring agents in various privacy protection technology modules, multi-dimensional key performance indicators (KPIs) are collected and evaluated in real time after the implementation of privacy protection policies. The monitored multi-dimensional key performance indicators (KPIs) include: Privacy protection effectiveness: Measuring actual privacy budget consumption Compared with expected value Deviation; Data utility level: Calculating information retention rate Correlation with features ; System resource consumption: Record CPU utilization Memory usage and network traffic ; Lyapunov stability index: The stability coefficient γ(t) is calculated based on the Lyapunov function value L(Θ(t)). Random Network Calculus (SNC) Service Quality Metrics: Based on output queue overflow probability and the upper limit of latency (t); Step 4.3, Parameter Dynamic Adjustment and Optimization Loop; Based on the multi-dimensional key performance indicators (KPIs) obtained from step 4.2, an online learning algorithm is used to dynamically optimize the system control parameters, forming an adaptive optimization loop: The performance index function for the control parameter V optimized for Lyapunov is calculated as follows: in, For average drift, For average cost, The variance of the queue length; Calculating the SNC service quality index for random networks The performance index function is calculated as follows: in, This represents the actual average overflow probability. The average service quality index; The Lyapunov control parameter V and the SNC service quality index θ are optimized using an online learning algorithm: in, and This is the learning rate.

5. A dynamic data privacy protection system based on the Lyapunov-SNC collaborative computing model, based on the method of claim 1, characterized in that, include: The status awareness and feature extraction module is used to collect system operation status data in real time, perform time-series smoothing processing on the collected system operation status data, and extract environmental parameters and data features. The Lyapunov-SNC cooperative queue management controller is used to implement parallel execution of Lyapunov stability control and random network calculus SNC reliability assessment based on processed system operating status data. Establish a collaborative queue management mechanism; A multi-objective privacy policy optimizer is used to model a multi-objective privacy protection optimization problem based on a cooperative queue management mechanism, solve the Pareto solution set using the NSGA-II algorithm, and dynamically select the privacy protection policy. The policy executor and adaptive optimization closed-loop module are used to convert the selected privacy protection policy into execution instructions, monitor system performance indicators in real time, and dynamically adjust parameters based on feedback to form a closed-loop adaptive optimization.

6. A dynamic data privacy protection device based on the Lyapunov-SNC collaborative computing model, characterized in that, include: Memory: A computer program for a dynamic data privacy protection method based on the Lyapunov-SNC collaborative computing model as described in any one of claims 1-4, and is a computer-readable device; Processor: Used to implement the dynamic data privacy protection method based on the Lyapunov-SNC collaborative computing model as described in any one of claims 1-4 when executing the computer program.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, enables the implementation of the dynamic data privacy protection method based on the Lyapunov-SNC collaborative computing model as described in any one of claims 1-4.

Images