Personalized location privacy perception unloading method based on dual-time scale optimization

Abstract

The invention provides a personalized location privacy perception unloading method based on dual-time scale optimization, and belongs to the field of multi-access edge computing. The personalized position privacy protection requirement of the user is constructed in combination with task sensitivity, different privacy preferences of the user in different scenes and task priorities; in the unloading process of the location privacy awareness task, a dual-time-scale optimization framework is utilized to protect personalized location privacy, meanwhile, the calculation cost is reduced, and the unloading effectiveness is maximized; according to the personalized location privacy protection requirement and the unloading proportion and privacy budget requirement of real-time privacy perception, personalized location privacy perception task unloading is completed by using a personalized location privacy perception task unloading mechanism based on double-agent deep reinforcement learning. The method is simple in step and convenient to use, and personalized position privacy protection is realized while the calculation cost is reduced.

Description

Technical Field

[0001] This invention relates to a personalized location privacy-aware offloading method based on dual time scale optimization, belonging to the field of network communication technology. Background Technology

[0002] In multi-access edge computing (MEC) systems, attackers can infer user location by monitoring the size of unloaded tasks. This exposure of location information poses a significant risk of location privacy breaches. Therefore, users exhibit different privacy needs across different application scenarios: task sensitivity, user privacy preferences, and task priority significantly influence personalized location privacy protection requirements. The sensitivity of unloaded tasks varies across applications, necessitating strict privacy protection. User privacy preferences are context-dependent, changing with spatial and temporal conditions; users are typically more concerned about location privacy in private residential areas than in public spaces. Task priority and urgency further influence the trade-off between privacy protection and unloading performance. Latency-sensitive tasks can tolerate lower privacy to meet strict latency requirements, while less urgent tasks can tolerate stronger location privacy. Existing technologies using differential privacy mechanisms can effectively mitigate privacy breaches in MEC systems and also demonstrate good application potential in ensuring task unloading efficiency.

[0003] However, existing technologies primarily consider the trade-off between computational cost and location privacy, without fully considering and comprehensively analyzing users' personalized location privacy protection needs. Users with high privacy protection requirements face the potential risk of location privacy leakage, while users with lower privacy requirements may be subject to unnecessary strong protection measures, thus significantly reducing offloading performance. Furthermore, location privacy-aware task offloading mechanisms that dynamically optimize perturbation regions or privacy budgets to balance location privacy and computational efficiency often lack the flexibility to adapt to dynamic location privacy protection needs, potentially failing to meet specific location privacy requirements and leading to suboptimal trade-offs between privacy protection and computational cost.

[0004] The prior art disclosed in CN113407249A is a task offloading method for location privacy protection, including: S1, generating a circular virtual location space centered on the user's real location, and selecting a virtual location with an initial random probability; S2, using the virtual location to obtain the locations of nearby base stations where edge servers are deployed, and constructing a server selection matrix; S3, selecting a server based on the server selection matrix to send a task offloading request; S4, allocating bandwidth and calculating relevant parameters for processing the task; S5, selecting the task that can be completed within the task tolerance time and saves the most energy and time, and has the greatest benefit, and offloading it to the edge node for execution. This method only considers user location privacy when offloading tasks, lacking consideration for the complexity of user privacy protection, and also does not perform sufficient optimization of task offloading, resulting in poor performance. Summary of the Invention

[0005] To address the shortcomings of existing technologies, a personalized location privacy-aware offloading method based on dual-timescale optimization is proposed. This method fully considers and analyzes the personalized location privacy protection requirements. For dynamic MEC environments and personalized location privacy protection needs, the personalized location privacy-aware task offloading method, which combines a dual-timescale optimization framework with a deep reinforcement learning algorithm, assists user terminals in offloading tasks, achieving maximum offloading utility and minimizing computational costs while protecting personalized location privacy.

[0006] To achieve the above objectives, this invention proposes a personalized location privacy-aware offloading method based on dual-time-scale optimization, the specific steps of which are as follows: S1. In the dynamic environment of the MEC system, personalized location privacy protection requirements for users are constructed by combining task sensitivity, users' different privacy preferences in different scenarios, and task priorities. ; S2. Designed to meet users' personalized location privacy protection needs. A differential privacy-based location perturbation strategy is proposed, which jointly optimizes the perturbation region and privacy budget; in the location privacy-aware task offloading, the design takes into account the user's personalized location privacy protection needs. Personalized location privacy-aware task unloading utility function Consider reducing the level of privacy breaches while reducing computational costs; S3. Utilize a dual-timescale optimization framework to protect personalized location privacy while reducing computational costs and maximizing offload utility; leverage utility functions... By setting two time scales, the perturbation area is optimized to adapt to personalized location privacy protection needs through the long time scale, and the offloading ratio and privacy budget of real-time privacy awareness are optimized through the short time scale, so as to maximize the offloading utility under instantaneous channel and offloading workload conditions. S4. To address the needs for personalized location privacy protection, as well as the offloading ratio and privacy budget requirements for real-time privacy awareness, Markov Decision Processes (MDPs) are modeled, and the personalized location privacy awareness task offloading mechanism DDPLM based on dual-agent deep reinforcement learning is used to complete the offloading of personalized location awareness tasks.

[0007] Furthermore, the edge computing system (MEC) includes users and MEC servers, and the distance between the users and MEC servers is [missing information]. The channel power gain is ,index Given the path loss factor, and based on the Shannon-Hartley theorem, the maximum wireless data transmission rate is achieved. for: , in, For wireless offloading of system bandwidth, To unload power, The power spectral density of the background noise; the MEC server through bandwidth and background noise The channel gain is estimated to infer the approximate distance to the user, as shown below: , Specifically, modeling the requirements for personalized location privacy protection: Based on the user's access time to sensitive tasks and frequency To measure task sensitivity : , in, and Obtained from the user's historical uninstallation tasks and behaviors, coefficients and These represent the weights of access time and access frequency respectively when calculating task sensitivity, satisfying... Higher task sensitivity will require stricter location privacy protection; User privacy preferences for: , in, This is a continuous privacy preference indicator, reflecting the degree to which users value privacy; when users' privacy preferences... When the value approaches 1, users tend to implement higher levels of location privacy protection. When the value approaches -1, users are more willing to relax privacy requirements; Priority of computational tasks in MEC systems : , in, The first priority of the task class, It has the highest priority. Based on task sensitivity User privacy preferences and task priority Solving Personalized Location Privacy Protection Needs , The range is A larger value indicates a stronger need for location privacy protection, expressed as: , in, , and These represent the weights corresponding to task sensitivity, user privacy preference, and task priority, respectively.

[0008] Furthermore, a distance perturbation scheme based on differential privacy is used to perturb the true distance between the user and the MEC server: an offloading strategy is generated based on the perturbation distance so that the MEC server cannot infer the true distance between itself and the user through the user's offloading behavior, thereby providing effective location privacy protection for the user during the task offloading process. At the same time, the perturbation area and privacy budget are jointly optimized to effectively achieve the level of location privacy protection required by the user, while the calculation of the offloading ratio is optimized to minimize the computational cost: the optimization goal is to maximize the offloading utility while achieving personalized location privacy protection. 2.1 Disturbance area The real distance of users Distance from disturbance to post-disturbance l1 and l2 represent the disturbance regions, respectively. Design the distance perturbation probability density function for the left and right boundaries. To ensure that the disturbance mechanism satisfies Differential privacy, The distance perturbation probability density function represents the privacy budget allocated by users based on their location privacy protection needs. The expression is defined as follows: , in, As a shorthand expression on the right side of the equation, The width of the disturbance region. Strictly ensure that the disturbed area is within the server's service application range, including This indicates the maximum service application scope of the MEC server; The probability density distribution function of the perturbation is quantified using the KL divergence. Uniform probability density distribution function without location privacy protection The differences between them are used to obtain the level of location privacy leakage. : , 2.2 Analysis of Energy Loss and Computation Delay in MEC System: Offloading user-collected channel state information, based on task size bit performs task offloading, defines To calculate the strength, This represents the uninstallation percentage. The CPU frequency required for a user to process 1 bit of data, and The CPU frequency required for the MEC server to process 1 bit of data; the user adopts a partial offloading scheme, specifically... Bit data is offloaded to the MEC server for processing, and the remaining data is processed by the MEC server. Bit data is processed locally, taking into account energy loss and computational latency during local computation and computation offloading processes. and These represent the energy loss and computation delay during the local computation process, respectively, and are calculated as follows: , , in The energy consumed by the user terminal per CPU cycle; make and Let represent the energy loss and computation delay during the task unloading process, respectively, as shown in the following expressions: , , Due to the adoption of a parallel computing paradigm, the total energy loss is... Total computational delay This can be obtained from the following: , , Utilizing total energy loss Total computational delay The total computational cost of weighted sum solution The expression is as follows: , in These are weighted parameters; 2.3: Optimization Modeling of Personalized Location Privacy Awareness Task Offloading: Designing a Utility Function for Personalized Location Privacy Awareness Task Offloading It balances users' personalized location privacy protection needs, privacy leakage levels, and computing costs: , in It is a weighted parameter that quantifies how much users value computing costs.

[0009] Furthermore, a dual-timescale optimization framework is used to jointly optimize the perturbation region, privacy budget, and offloading ratio, utilizing a utility function. Using the need for personalized location privacy protection as a time-scale segmentation condition, a fixed interval is set for the long time scale. Short timescale intervals Each long time scale includes multiple short time scales; the optimization problem for offloading personalized location privacy-aware tasks that meet the dynamic location privacy protection requirements under the dual-time-scale optimization framework is modeled as follows: , The constraints of the above optimization problem include: , , , , , in Indicates the uninstallation percentage. This represents the privacy budget allocated by users based on their location privacy protection needs. Strictly ensure that the disturbed area is within the server's service application range, including This represents the maximum service application scope of the MEC server. For the first The fixed interval on a long time scale corresponds to the first... Optimized interval for each disturbance region It is the maximum index. and These represent long time intervals. Corresponding disturbance area The left and right boundaries; This indicates the time interval for unloading location privacy-aware tasks, and An index representing the time interval for unloading location privacy-aware tasks. It is the maximum index, for each long time interval. Average uninstall utility under dynamic location privacy requirements By each short time interval sampling The second disturbance distance is calculated, and the corresponding unloading utility value is also calculated. The average is obtained. Indicates a short time interval The calculation of the unloading ratio, Indicates a short time interval Privacy budget.

[0010] Furthermore, in the personalized location privacy-aware task offloading mechanism DDPLM based on dual-agent deep reinforcement learning, the DQN algorithm is used to determine the personalized perturbation region on a long time scale, while the TD3 algorithm is used on a short time scale to obtain the optimal offloading ratio and privacy budget, thereby achieving personalized location privacy protection while minimizing computational costs.

[0011] Furthermore, determining the personalized perturbation region using the DQN algorithm includes the following steps: Each long time interval The state is represented as: , in, For the current long time interval Personalized location privacy protection needs, , Indicates a long time interval in the past The average level of location privacy leakage is valuable historical information that can be used as a state component; the action space for long-term personalized location privacy protection is defined as follows: , in, and These represent long time intervals. Corresponding disturbance area The left and right boundaries; Long-term reward The expression used to evaluate the offloading strategy for personalized location privacy-aware tasks is as follows: , in, Represents each long time interval middle Average unloading utility value over a short time interval; Experience in identifying personalized disturbance regions Stored in the memory pool; experience determined through personalized perturbation regions within the memory pool. Data update DQN network.

[0012] Furthermore, the state space representation using the TD3 algorithm on a short time scale is as follows: , in, For workload, the offload interval Medium disturbance distance It is generated based on a perturbation function set according to the distance; each long time interval is generated by... It consists of several unloading intervals. At the beginning of each unloading interval, the user generates a task and generates a perturbation distance according to the perturbation strategy, thereby obtaining the current short-time scale state. Using the TD3 algorithm on a short timescale to learn the offloading ratio and privacy budget, location privacy-aware task offloading is achieved while reducing computational costs. The action space is defined as follows: ,in, and The unloading intervals are respectively Uninstallation rate and privacy budget; The short-time-scale optimization objective is to maximize offloading utility. The reward expression for the short-time location privacy-aware task offloading strategy is as follows: , Perform the selected action, interact with the environment, and calculate short-term rewards. And obtain the state on the next short time scale. Unloading location privacy-aware tasks Store in the memory pool; use the location privacy-aware task uninstallation experience data in the memory pool to update the TD3 network.

[0013] A computer device includes a processor and a memory, the processor being electrically connected to the memory for storing instructions and data, and the processor being configured to execute a personalized location privacy-aware offloading method based on dual-timescale optimization.

[0014] Beneficial Effects: Addressing the challenges of dynamic MEC environments and personalized location privacy protection needs, this paper utilizes a dual-timescale optimization framework to solve the complex problem of task offloading for personalized location privacy awareness. On a long-term timescale, it optimizes the perturbation region to reflect the long-term changes in location privacy requirements. On a short-term timescale, it dynamically optimizes the offloading ratio and privacy budget based on instantaneous channel states and offloaded workloads. Furthermore, it employs a personalized location privacy awareness task offloading method based on a dual-agent deep reinforcement learning algorithm to assist user terminals in handling computational tasks, achieving efficient task offloading while maintaining personalized location privacy. The paper models personalized location privacy protection requirements from key factors such as task sensitivity, user privacy preferences, and task priority, and proposes a joint optimization paradigm to jointly optimize the perturbation region and privacy budget according to specific privacy requirements. This approach achieves personalized location privacy protection while reducing computational costs. Attached Figure Description

[0015] Figure 1 This is a flowchart of the personalized location privacy-aware offloading method based on dual time scale optimization of the present invention. Detailed Implementation

[0016] The embodiments of the present invention will be further described below with reference to the accompanying drawings:

[0017] like Figure 1As shown, this invention discloses a personalized location privacy-aware offloading method based on dual-timescale optimization. In multi-access edge computing (MEC) systems, since user offloading behavior is closely related to wireless channel conditions, untrusted MEC servers can infer user location information based on the user's offloading behavior, and users exhibit different privacy needs in different application scenarios. This invention specifically analyzes the threat of location privacy leakage and the need for personalized location privacy protection during the computation offloading process, and then proposes a location perturbation strategy based on differential privacy and jointly optimizes the perturbation region and privacy budget to maximize offloading utility. The complexity of this optimization problem lies in three aspects: 1) It is complex to provide users with provable personalized location privacy protection while minimizing computational costs; 2) Adjusting only the privacy budget or perturbation region is insufficient to meet specific privacy needs, and jointly optimizing the privacy budget and perturbation region leads to a complex optimization objective and high dimensionality, which is a nondeterministic polynomial-time problem that is difficult to solve effectively by traditional methods; 3) The personalized location privacy protection needs and offloading workload are time-varying during the task offloading process, so it is necessary to adaptively obtain the optimal personalized location privacy-aware task offloading strategy under the dynamic MEC system. This invention models the personalized location privacy protection requirements based on key factors such as task sensitivity, user privacy preferences, and task priority, and jointly optimizes the perturbation region and privacy budget to meet the location privacy protection requirements. It uses a dual-timescale optimization framework to solve the complex problem of personalized location privacy awareness task offloading optimization. Specifically, users optimize the personalized perturbation region on a long-term timescale to meet long-term privacy requirements, and optimize the offloading ratio and privacy budget on a short-term timescale based on instantaneous channel state and offloaded workload.

[0018] A personalized location privacy-aware offloading method based on dual-time-scale optimization is proposed, with the following specific steps: S1. In the dynamic environment of the MEC system, personalized location privacy protection requirements for users are constructed by combining task sensitivity, users' different privacy preferences in different scenarios, and task priorities. ...

[0019] S2. Designed to meet users' personalized location privacy protection needs. A differential privacy-based location perturbation strategy is proposed, which jointly optimizes the perturbation region and privacy budget; in the location privacy-aware task offloading, the design takes into account the user's personalized location privacy protection needs. Personalized location privacy-aware task unloading utility function Consider reducing the level of privacy breaches while reducing computational costs; Edge computing (MEC) systems consist of users and MEC servers, with the distance between the offloading user and MEC server being [missing information]. The channel power gain is ,index Given the path loss factor, and based on the Shannon-Hartley theorem, the maximum wireless data transmission rate is achieved. for: , in, For wireless offloading of system bandwidth, To unload power, The power spectral density of the background noise; the MEC server through bandwidth and background noise The channel gain is estimated to infer the approximate distance to the user, as shown below: , Specifically, modeling the requirements for personalized location privacy protection: Based on the user's access time to sensitive tasks and frequency To measure task sensitivity : , in, and Obtained from the user's historical uninstallation tasks and behaviors, coefficients and These represent the weights of access time and access frequency respectively when calculating task sensitivity, satisfying... Higher task sensitivity will require stricter location privacy protection; User privacy preferences for: , in, This is a continuous privacy preference indicator, reflecting the degree to which users value privacy; when users' privacy preferences... When the value approaches 1, users tend to implement higher levels of location privacy protection. When the value approaches -1, users are more willing to relax privacy requirements; Priority of computational tasks in MEC systems : , in, The first priority of the task class, It has the highest priority. Based on task sensitivity User privacy preferences and task priority Solving Personalized Location Privacy Protection Needs , The range is A larger value indicates a stronger need for location privacy protection, expressed as: , in, , and These represent the weights corresponding to task sensitivity, user privacy preference, and task priority, respectively.

[0020] A distance perturbation scheme based on differential privacy is used to perturb the true distance between the user and the MEC server: an offloading strategy is generated based on the perturbation distance so that the MEC server cannot infer the true distance between itself and the user through the user's offloading behavior, thereby providing effective location privacy protection for the user during the task offloading process. At the same time, the perturbation area and privacy budget are jointly optimized to effectively achieve the level of location privacy protection required by the user, while the calculation of the offloading ratio is optimized to minimize the computational cost: the optimization goal is to maximize the offloading utility while achieving personalized location privacy protection. 2.1 Disturbance area The real distance of users Distance from disturbance to post-disturbance l1 and l2 represent the disturbance regions, respectively. Design the distance perturbation probability density function for the left and right boundaries. To ensure that the disturbance mechanism satisfies Differential privacy, The distance perturbation probability density function represents the privacy budget allocated by users based on their location privacy protection needs. The expression is defined as follows: , in, As a shorthand expression on the right side of the equation, The width of the disturbance region. Strictly ensure that the disturbed area is within the server's service application range, including This indicates the maximum service application scope of the MEC server; The probability density distribution function of the perturbation is quantified using the KL divergence. Uniform probability density distribution function without location privacy protection The differences between them are used to obtain the level of location privacy leakage. : , 2.2 Analysis of Energy Loss and Computation Delay in MEC System: Offloading user-collected channel state information, based on task size bit performs task offloading, defines To calculate the strength, This represents the uninstallation percentage. The CPU frequency required for a user to process 1 bit of data, and The CPU frequency required for the MEC server to process 1 bit of data; the user adopts a partial offloading scheme, specifically... Bit data is offloaded to the MEC server for processing, and the remaining data is processed by the MEC server. Bit data is processed locally, taking into account energy loss and computational latency during local computation and computation offloading processes. and These represent the energy loss and computation delay during the local computation process, respectively, and are calculated as follows: , , in The energy consumed by the user terminal per CPU cycle; make and Let represent the energy loss and computation delay during the task unloading process, respectively, as shown in the following expressions: , , Due to the adoption of a parallel computing paradigm, the total energy loss is... Total computational delay This can be obtained from the following: , , Utilizing total energy loss Total computational delay The total computational cost of weighted sum solution The expression is as follows: , in These are weighted parameters; 2.3: Optimization Modeling of Personalized Location Privacy Awareness Task Offloading: Designing a Utility Function for Personalized Location Privacy Awareness Task Offloading It balances users' personalized location privacy protection needs, privacy leakage levels, and computing costs: , in It is a weighted parameter that quantifies how much users value computing costs.

[0021] S3. Utilize a dual-timescale optimization framework to protect personalized location privacy while reducing computational costs and maximizing offload utility; leverage utility functions... By setting two time scales, the perturbation area is optimized to adapt to personalized location privacy protection needs through the long time scale, and the offloading ratio and privacy budget of real-time privacy awareness are optimized through the short time scale, so as to maximize the offloading utility under instantaneous channel and offloading workload conditions. A dual-timescale optimization framework is used to jointly optimize the perturbation region, privacy budget, and offloading ratio, utilizing a utility function. Using the need for personalized location privacy protection as a time-scale segmentation condition, a fixed interval is set for the long time scale. Short timescale intervals Each long time scale includes multiple short time scales; the optimization problem for offloading personalized location privacy-aware tasks that meet the dynamic location privacy protection requirements under the dual-time-scale optimization framework is modeled as follows: , The constraints of the above optimization problem include: , , , , , in Indicates the uninstallation percentage. This represents the privacy budget allocated by users based on their location privacy protection needs. Strictly ensure that the disturbed area is within the server's service application range, including This represents the maximum service application scope of the MEC server. For the first The fixed interval on a long time scale corresponds to the first... Optimized interval for each disturbance region It is the maximum index. and These represent long time intervals. Corresponding disturbance area The left and right boundaries; This indicates the time interval for unloading location privacy-aware tasks, and An index representing the time interval for unloading location privacy-aware tasks. It is the maximum index, for each long time interval. Average uninstall utility under dynamic location privacy requirements By each short time interval sampling The second disturbance distance is calculated, and the corresponding unloading utility value is also calculated. The average is obtained. Indicates a short time interval The calculation of the unloading ratio, Indicates a short time interval Privacy budget.

[0022] S4. To address the needs for personalized location privacy protection, as well as the offloading ratio and privacy budget requirements for real-time privacy awareness, Markov Decision Processes (MDPs) are modeled, and the personalized location privacy awareness task offloading mechanism DDPLM based on dual-agent deep reinforcement learning is used to complete the offloading of personalized location awareness tasks.

[0023] In the personalized location privacy-aware task offloading mechanism DDPLM based on dual-agent deep reinforcement learning, the DQN algorithm is used to determine the personalized perturbation region on a long time scale, while the TD3 algorithm is used on a short time scale to obtain the optimal offloading ratio and privacy budget, thereby achieving personalized location privacy protection while minimizing computational costs.

[0024] Over long timescales, the Deep Q-Network (DQN) algorithm is used to adjust the location perturbation region according to personalized location privacy requirements to adapt to long-term changes in these requirements. Simultaneously, over short timescales, the Twin Delayed Deep Deterministic Policy Gradient (TD3) algorithm is used to optimize the offloading rate and privacy budget based on instantaneous channel states and offloaded workloads. First, the relevant parameters of the DQN and TD3 algorithms in DDPLM are initialized, specifically including the learning rate in the DQN algorithm. Discount Factor Memory pool size, batch size, number of rounds, and exploration step size within a round; learning rate and discount factor in the TD3 algorithm. Exploring noise Soft update parameters Memory pool size, batch size, number of rounds, and exploration step size within each round; initialize the main network parameters in the DQN algorithm. and Target Q network parameters Actor network in TD3 algorithm parameter Target Actor Network parameter Two Critic networks Network parameters Two Target Critic networks Network parameters , and initialize environment parameters. and This involves initializing the level of location privacy breaches and the privacy budget; Each long time interval The state model is as follows: , in, For the current long time interval Personalized location privacy protection needs, , Represents a long time interval in the past Historical information on the average level of location privacy leakage; on a long-term scale, based on personalized location privacy protection needs and the overall level of privacy leakage, the location perturbation region is determined using the DQN algorithm. Therefore, the action space for the long-term personalized location privacy protection problem is defined as follows: , Short-timescale DQN algorithm: Due to the differential privacy-based location perturbation strategy in each long time interval The decision is made once at the beginning, therefore the location perturbation area and specific personalized location privacy needs In each unloading interval The internal remains constant. , Considering the perturbation distance and dynamically changing task workload, the offload rate and privacy budget are determined based on the TD3 algorithm in a short timescale to minimize complexity costs. Since computational costs are affected by workload and perturbation distance, the state space of the TD3 algorithm in a short timescale can be modeled as follows: , in, For workload, the offload interval Medium disturbance distance It was generated based on the distance perturbation function PDF. The design for determining the personalized perturbation region over long timescales played a significant role in guiding the selection of strategies for offloading location privacy-aware tasks on short timescales, using the location perturbation strategy as... Some state elements help users achieve a higher uninstallation efficiency; Each long time interval is composed of The process consists of several unloading intervals. At the beginning of each unloading interval, the user generates a task and generates a perturbation distance according to the perturbation strategy, thereby obtaining the current short-term state. , In short timescales, the TD3 algorithm is used to learn the offloading rate and privacy budget to offload location privacy-aware tasks while reducing computational costs. Therefore, the action space is defined as follows: , in, and The unloading intervals are respectively Uninstall rates and privacy budgets; Based on the current short-timescale state Select a short-timescale location privacy-aware task offloading strategy ; The short-timescale optimization objective is to maximize offloading utility. Therefore, the reward expression for the short-time location privacy-aware task offloading strategy is as follows: , Perform the selected action, interact with the environment, and calculate short-term rewards. And obtain the state on the next short time scale. Unloading location privacy-aware tasks Stored in the memory pool; Update the TD3 network: Extract experience data from the memory pool and update the current state. Input into the Actor network Generate location privacy-aware task offloading strategy , and Target Actor network Based on the next state Output target action Critic Network based on and Calculate Q value , and Target Critic network based on and Calculate Q value To address the policy overfitting problem caused by function approximation error and improve the algorithm's stability and generalization performance, TD3 adds noise to the target action, namely: , in, To explore the noise, it is independently sampled from a truncated normal distribution, and the expression is as follows: , In the above formula, Represents policy noise. Represents the noise clipping threshold; The target Q-value is calculated based on the minimum Q-value of the outputs of the two Target Critic networks. To suppress overestimation, the expression is as follows: , in, Discount factor; Gradient descent is used to update the parameters of the two Critic networks: , Maximize the Q-value to delay updating the Actor network: , Soft update target network: , After updating the network parameters, update the short-timescale state. ; After each unloading interval, the average location privacy leakage level can be calculated. ; Calculating long-term rewards And obtain the state on the next time scale. Long-term reward The expression used to evaluate the offloading strategy for personalized location privacy-aware tasks is as follows: , in, Represents each long time interval middle Average unloading utility value over a short time interval; Experience in identifying personalized disturbance regions Stored in the memory pool; Update the DQN network: In the DQN algorithm, experience is extracted from the memory pool and evaluated using an evaluation neural network. The Q-value is evaluated, and the target neural network is used. The evaluation network is updated based on the sampled data, and the expected reward, i.e., the Q value, is updated as follows:

[0025] in, For learning rate, Discount factor; Loss function of DQN algorithm The definition is as follows: , The gradient descent method is used to optimize the Q value, thereby minimizing the loss value in formula (31) and updating the evaluation neural network parameters. Using evaluation neural network parameters Periodically update the parameters of the target neural network This improves network update efficiency; After updating the network parameters, update the long-term scale state. .

Claims

1. A personalized location privacy-aware offloading method based on double-time-scale optimization, characterized in that, The specific steps are as follows: S1, in the dynamic environment in the MEC system, combining task sensitivity, different privacy preferences of users in different scenarios and task priority to build personalized location privacy protection requirements of users ; S2. Designed to meet users' personalized location privacy protection needs. A differential privacy-based location perturbation strategy is proposed, which jointly optimizes the perturbation region and privacy budget; in the location privacy-aware task offloading, the design takes into account the user's personalized location privacy protection needs. Personalized location privacy-aware task unloading utility function Consider reducing the level of privacy breaches while reducing computational costs; S3, protect personalized location privacy while reducing computational cost and maximizing offloading utility by using a dual-time-scale optimization framework; use utility function Set dual time scale, optimize disturbance area by long time scale to adapt to personalized location privacy protection demand, optimize real-time privacy-aware offloading ratio and privacy budget by short time scale, maximize offloading utility under instantaneous channel and offloading workload conditions; S4, the personalized location privacy protection requirement and the real-time privacy perception offloading ratio and privacy budget requirement are respectively modeled into Markov decision process MDP, and the personalized location privacy perception task offloading mechanism DDPLM based on double-agent deep reinforcement learning is used to complete the personalized location perception task offloading.

2. The dual-time-scale optimization based personalized location privacy-aware offloading method according to claim 1, characterized in that, The edge computing system MEC includes a user and a MEC server, the distance between the user and the MEC server is , the channel power gain is , the index is a path loss factor, and the maximum wireless data transmission rate realized based on the Shannon-Hartley theorem is : ， wherein, is the system bandwidth offloaded wirelessly, is the offloaded power, is the power spectral density of the background noise; the MEC server estimates the channel gain to infer the approximate distance from the user, denoted as follows: and the background noise ​ ， The personalized location privacy protection requirement is modeled as follows: Measuring task sensitivity by user access time to sensitive tasks and frequency :​ ， where, and are the weights of the access time and the access frequency, respectively, in computing the task sensitivity, satisfying and A higher task sensitivity will require a more stringent location privacy protection.​ User privacy preferences For: ， wherein, is a continuous privacy preference indicator, reflecting the degree of importance of privacy to the user; when the user privacy preference tends to 1, the user tends to perform higher levels of location privacy protection, when tends to -1, the user is more willing to relax privacy requirements; Priority of computing tasks in MEC systems : ， wherein, is the first level of task priority, is the maximum priority; According to the task sensitivity , the user privacy preference and the task priority to solve the personalized location privacy protection requirement , The range is , the larger the value represents the stronger the demand for location privacy protection, and the expression is: ， wherein, , and represent the weights corresponding to task sensitivity, user privacy preference, and task priority, respectively.

3. The dual-time-scale optimization based personalized location privacy-aware offloading method according to claim 2, characterized in that, The real distance between the user and the MEC server is disturbed by using a distance disturbance scheme based on differential privacy: an offloading strategy is generated according to the disturbed distance, so that the MEC server cannot infer the real distance from the user through the user's offloading behavior, thereby providing effective location privacy protection for the user during the task offloading process, while optimizing the disturbance area and the privacy budget to effectively achieve the required level of location privacy protection for the user, and optimizing the calculation offloading ratio to reduce the calculation cost to the greatest extent: the optimization goal is to maximize the offloading utility while achieving personalized location privacy protection; 2.1 Perturbing the region The real distance of a user The post-perturbed distance , and l1 and l2 represent the left and right boundaries of the perturbed region respectively, the distance perturbation probability density function is designed to ensure that the perturbation mechanism satisfies - differential privacy, denotes the privacy budget allocated by the user according to the location privacy protection requirement, the expression of the distance perturbation probability density function is defined as follows: ， wherein, as a shorthand expression for the right side of the equation, is a perturbation region width of the perturbation region, strictly guaranteeing that the perturbation region is within the service application range of the server, wherein denotes the maximum service application range of the MEC server; Quantifying the perturbed probability density distribution function by KL divergence The difference between the uniform probability density distribution function without location privacy protection represents the level of location privacy leakage : ， 2.2 Analyzing the energy consumption and computation latency of MEC system: offloading users collect channel state information, according to the task size bit to perform task offloading, define as the computation intensity, as the offloading ratio, as the CPU frequency required for a user to process 1 bit of data, and as the CPU frequency required for a MEC server to process 1 bit of data; the user adopts a partial offloading scheme, specifically offloading bits of data to the MEC server for processing, and the remaining bits of data are processed locally, considering the energy consumption and computation latency of the local computation and computation offloading processes, let and represent the energy consumption and computation latency of the local computation process, respectively, which are calculated as follows: ， ， wherein is the energy consumed by the user terminal per CPU cycle; Let and denote the energy consumption and computation latency of the task offloading process, respectively, and are expressed as follows: ， ， Due to the parallel computing paradigm, the total energy consumption and the total computing latency can be derived as follows: ， ， The total energy loss is solved by using the weighted sum of the total energy loss and the total computing latency The total computing cost is solved by using the weighted sum of the total energy loss and the total computing latency The expression is as follows: ， wherein is a weighting parameter; 2.3: Modeling of personalized location privacy-aware task offloading optimization: design a personalized location privacy-aware task offloading utility function , taking into account the user's personalized location privacy protection needs, privacy leakage level and computing cost: ， wherein is a weighting parameter quantifying the user's appreciation of the computational cost.

4. The dual-time-scale optimization based personalized location privacy-aware offloading method according to claim 3, characterized in that, Joint optimization of the perturbed region, privacy budget and offloading ratio is performed using a two-time-scale optimization framework, and an utility function is used Personalized location privacy protection requirements are taken as time scale segmentation conditions, and fixed intervals of a long time scale and intervals of a short time scale are set Each long time scale includes multiple short time scales ​ The personalized location privacy perception task offloading optimization problem under the double-time scale optimization framework to meet the dynamic location privacy protection requirement is modeled as: ， The constraints of the above optimization problem include: ， ， ， ， ， wherein denotes the offloading ratio, denotes the privacy budget allocated by the user according to the location privacy protection requirement, strictly guarantee that the perturbation region is within the service application range of the server, wherein denotes the maximum service application range of the MEC server, is the fixed interval of the long time scale, and the physical meaning corresponds to the optimization interval of the perturbation region, is the maximum index, and denote the left boundary and the right boundary of the perturbation region corresponding to the ; denotes the time interval of the location privacy-aware task offloading, and denotes the index of the time interval of the location privacy-aware task offloading, is the maximum index, and each long time interval average offloading utility under dynamic location privacy requirement is obtained by sampling perturbation distances and calculating the corresponding offloading utility value times for each short time interval , and taking the average, denotes the calculated offloading ratio of the short time interval , denotes the privacy budget of the short time interval .

5. The dual-time-scale optimization based personalized location privacy-aware offloading method according to claim 4, characterized in that, In the double-agent deep reinforcement learning personalized location privacy perception task offloading mechanism DDPLM, the long-time scale uses the DQN algorithm to determine the personalized disturbance area, and the short-time scale uses the TD3 algorithm to obtain the optimal offloading ratio and privacy budget, thereby achieving personalized location privacy protection while minimizing the calculation cost to the greatest extent.

6. The dual-time-scale optimization based personalized location privacy-aware offloading method according to claim 5, characterized in that, The steps for determining the personalized disturbance area using the DQN algorithm include: Each long interval is represented by a state as follows: ， wherein, is the current long time interval of personalized location privacy protection requirements, , denotes the past long time interval average location privacy leakage level, which is the valid historical information worth as a state component; the action space of the long time scale personalized location privacy protection problem is defined as follows: ， wherein, and respectively represent long time intervals corresponding to left and right boundaries of the perturbed region respectively. Long timescale reward is used to evaluate the personalized location privacy-aware task offloading policy, which is expressed as follows: ， wherein, represents the average unloading utility value for each long time interval in short time intervals; Determining experience with personalized perturbation regions Storing in a memory pool; determining experience with personalized perturbation regions in the memory pool Data update DQN network.

7. The dual-time-scale optimization based personalized location privacy-aware offloading method according to claim 6, characterized in that, The state space of the short-time scale using the TD3 algorithm is represented as: ， wherein, is the workload, the unload time interval is the perturbation distance is generated according to the distance set perturbation function; each long time interval is composed of unload intervals, at the beginning of each unload interval, the user generates tasks and generates a perturbation distance according to the perturbation strategy, thereby obtaining the current short time scale state; Short time scale uses TD3 algorithm to learn offloading ratio and privacy budget to achieve location privacy-aware task offloading while reducing computing cost, the action space is defined as: wherein, and are the offloading ratio and privacy budget of the offloading interval respectively. The short-time scale optimization goal is to maximize the offloading utility, and the reward expression of the short-time location privacy perception task offloading strategy is as follows: ， Perform the selected action, interact with the environment, and calculate short-term rewards. And obtain the state on the next short time scale. Unloading location privacy-aware tasks Store in the memory pool; use the location privacy-aware task uninstallation experience data in the memory pool to update the TD3 network.

8. A computer device, comprising: The device includes a processor and a memory, the processor is electrically connected with the memory, the memory is used for storing instructions and data, and the processor is used for executing the double-time scale optimization based personalized location privacy perception offloading method in any one of claims 1-7.

Images