Unmanned aerial vehicle assisted space-air-ground integrated task shunting method based on privacy-time delay collaborative optimization

Abstract

The invention discloses an unmanned aerial vehicle assisted space-air-ground integrated task shunting method based on privacy-time delay collaborative optimization, and belongs to the technical field of communication. According to the method, on the basis of constructing a communication and calculation model, an improved Laplacian position disturbance mechanism is introduced, random disturbance is carried out on the real position of a user, and false position distribution meeting service coverage and signal constraint is generated on the premise that task shunting feasibility is guaranteed. Establishing a probability density function of a disturbance position by performing optimization search on a disturbance radius and eliminating an angle domain which does not meet system constraints; and furthermore, in combination with a received signal indication intensity simulation model, the transmitting power required in an air node communication scene such as an unmanned aerial vehicle at a false position is calculated, so that the leakage risk of a position side channel is reduced. On the basis, a deep neighborhood search cross entropy algorithm is adopted to solve a task shunting decision, and collaborative optimization of privacy protection and system performance is achieved. According to the method, the effectiveness and the calculation efficiency of task shunting can be maintained while the location privacy of the user is guaranteed, and the method is suitable for complex dynamic scenes such as a space-air-ground integrated network containing unmanned aerial vehicle nodes.

Description

Technical Field

[0001] This invention belongs to the field of communications and relates to a task offloading method for an integrated air-space-ground network that integrates location privacy protection. Specifically, it is a UAV-assisted integrated air-space-ground task offloading method based on privacy-latency co-optimization. Background Technology

[0002] In recent years, low-Earth orbit (LEO) satellites and unmanned aerial vehicle (UAV) systems have gained widespread attention in integrated air-space-ground networks as an important supplement to terrestrial networks, thanks to their wide coverage, high mobility, and flexible deployment methods. By working collaboratively with ground edge nodes, satellites and UAVs can provide users with on-demand communication and computing resources, making task offloading a key technology for improving computing power utilization efficiency and service quality. However, in these scenarios, users often need to interact with multiple airborne or ground nodes during task offloading and computation offloading. This process inevitably exposes the user's geographical location and task data content, leading to serious privacy risks. Attackers can still infer the user's true location or movement trajectory by analyzing the user's offloading decisions and link quality. In air-space-ground networks, due to the large coverage area of ​​satellites and the flexible deployment of UAVs, user offloading behavior often exhibits a significant spatial bias, further increasing the possibility of inferring the user's location through task offloading patterns. Therefore, relying solely on traditional secure transmission mechanisms is insufficient to fundamentally prevent location privacy leaks. Existing research has begun to focus on introducing location privacy protection mechanisms into task offloading and edge computing processes. These mechanisms utilize methods such as location perturbation, anonymization, or differential privacy to make it statistically difficult to accurately distinguish a user's true location when participating in computation offloading. However, directly introducing location perturbation into integrated air-space-ground networks still faces several challenges. On the one hand, the coverage and service capabilities of satellite and UAV nodes are subject to significant spatial constraints; excessive location perturbation may render offloading decisions infeasible, leading to computational failures or a significant decline in service quality. On the other hand, user location is highly coupled with network state, making it difficult to strike a balance between privacy protection and system performance with simple random perturbations. Therefore, how to integrate location privacy protection mechanisms into the offloading decision-making process while meeting the feasibility and performance requirements of task offloading, and maintaining computational efficiency while reducing location inference risks, has become a critical issue that urgently needs to be addressed in integrated air-space-ground task offloading scenarios. Summary of the Invention

[0003] To address the shortcomings of existing research, such as the fact that most solutions assume the user's true location is known without considering the risk of location privacy leakage, or only use static privacy perturbation methods unrelated to task decision-making, making it difficult to balance privacy protection and task offloading performance, this invention proposes a privacy-preserving task offloading method in a space-air-ground network. This method aims to optimize task offloading for users in a heterogeneous network composed of satellites, UAVs, and ground nodes while ensuring user location privacy. Based on different user real locations, an improved Laplace mechanism is used to randomly perturb the user's real location, satisfying bit error rate constraints while balancing privacy protection and task offloading performance. For spurious locations after perturbation, a task offloading scheme is developed, and a deep neighborhood search cross-entropy algorithm is designed to achieve efficient task offloading, thereby maintaining the effectiveness of task offloading while ensuring user location privacy.

[0004] The technical solution adopted in this invention is: a drone-assisted air-ground integrated mission offloading method based on privacy-latency collaborative optimization, comprising the following steps:

[0005] (1) Construct a system model, jointly establish a communication model and a latency model, and on this basis, construct a task routing optimization problem to minimize user latency and privacy leakage; the optimization problem is:

[0006]

[0007] This indicates a false location after user disturbance. Represents the user's subtask Should I unload it onto the drone? , Represents the user's subtask Whether the drone unloads the contents onto the satellite, It is the simulated power calculated based on the spurious position. Subtasks Task processing latency, Represents a set of subtasks. Indicators of location privacy leaks , This indicates the real locations of two different users. Indicates privacy budget, express and The Euclidean distance between them Indicates the amount of data in the subtask. It is the capacity of the drone queue. This indicates the coverage radius of the drone u. This indicates the radius of the range where the signal strength received by the server cannot be imitated. This represents the distance between the spurious location and the drone's location u. Indicates maximum power limit. Indicates task The probability of successful uninstallation. This represents the threshold for the probability of successful uninstallation.

[0008] Constraint C1 is the geographical indistinguishability requirement; constraint C2 indicates that user tasks can only be transmitted to satellites by drones; constraint C3 indicates drone queue capacity limit; constraints C4 and C5 indicate that the user's false location cannot fall within the range that the server's received signal strength cannot be imitated, i.e., the user's false location must fall within the coverage area of ​​at least one server; C6 is the imitation power constraint; and constraint C7 is that the total probability of successful offloading should be greater than a pre-specified reliability threshold.

[0009] (2) Introduce a disturbance radius search mechanism to determine the optimal disturbance radius that satisfies the system bit error rate constraint and minimizes system latency and privacy leakage;

[0010] (3) Based on the user's real location, the angle domain that cannot support task unloading is eliminated. On this basis, an improved Laplace position perturbation model is constructed to obtain the false position probability density function in the limited angle domain.

[0011] (4) For the generated false location, calculate the transmission power required for the user to simulate the strength of the real received signal indication, and use the deep neighborhood search cross-entropy algorithm to determine the final task diversion scheme.

[0012] Step (2) specifically includes:

[0013] Centered on the user's real location, with a radius of The circular region is defined as the position disturbance region, and is set as follows: ,in Indicates the maximum distance from the user to the edge server coverage boundary; using , and This indicates the location privacy leakage, uninstallation effectiveness, and uninstallation success probability when using a fake location for uninstallation, and calculates the perturbation domain. The radius is Location privacy leaks Uninstallation effect and the probability of successful uninstallation Expectations:

[0014]

[0015] in The perturbation domain is False positions are generated at times The probability of;

[0016] The utility of the perturbation domain for:

[0017]

[0018] in It is a customized privacy factor that reflects the personalized location privacy requirements of mobile users;

[0019] The optimization problem of finding the optimal perturbation radius is expressed as:

[0020]

[0021] The above optimization problem is solved using a search algorithm to obtain the optimal perturbation radius.

[0022] In step (3), an improved Laplace position perturbation model is constructed to obtain the false position probability density function in the restricted angle domain, including two cases: 1. When the user's real position falls within the server's non-imitable range; 2. When the user's real position is outside the server's non-imitable range.

[0023] Scenario 1: When the user's real location falls within the server's unassimilable range, a minimum constraint is applied to r. Based on the Laplace mechanism, the sampling angle follows a uniform distribution, and the sampling length follows a Laplace distribution. The probability density function for generating false locations is as follows:

[0024]

[0025] in It is the truncation normalization factor. denoted by , where r represents the minimum disturbance radius and r represents the variable sampling radius. Indicates the sampling angle of the variable.

[0026] Scenario 2: When the user's actual location is outside the range that the server cannot mimic, the sampling angle range will be limited. We need to eliminate infeasible angle ranges. According to the Laplace mechanism, the sampling angle follows a uniform distribution, and the sampling length follows a Laplace distribution, resulting in the probability density function for generating false locations:

[0027]

[0028] in It is the length of the feasible angular domain, and the truncation normalization factor. .

[0029] The calculation of the transmit power required by the user to simulate the strength of a real received signal in step (4) specifically includes:

[0030] The strength of the received signal when the user sends data to the drone u at location q. Considering ,but Then it is necessary to simulate the received signal strength. Needs and Equal, that is Therefore, it can be calculated that:

[0031]

[0032] Indicates the user's real location With drone location The distance between them Indicates a fake user location With drone location The distance between them This indicates that the spurious position mimics power.

[0033] After generating a false location, to further enhance location privacy protection, the wireless signal characteristics need to be consistent with that false location. To this end, based on the geometric distance between the false location and the edge server, and combined with a free-space path loss model, the transmit power used to mimic the corresponding received signal strength is calculated. By adaptively adjusting the transmit power, the received signal strength observed by the server is kept consistent with the signal strength corresponding to the false location, thereby preventing attackers from using physical layer signal characteristics to deduce the user's true location and improving the overall effectiveness of privacy protection.

[0034] This invention employs an optimized task allocation scheme based on the cross-entropy method of deep neighborhood search. First, two independent neural networks encode task features (such as task size and generation location) and server features (such as server location, computing power, and the size of remaining task data in the queue), resulting in task vector representations and server vector representations. These two types of vectors are then input into a third pairing scoring network, which outputs a non-negative score for each task assigned to each server. This score is then normalized to form a prior probability distribution for task-server matching, used to initialize the probability table of the cross-entropy method. During the solution process, multiple allocation solutions are constructed by sampling tasks one by one within the feasible server set based on the current probability table. An objective function is calculated for each solution, and the best subset of solutions is selected as the elite solution set. The probability table is then updated based on the frequency of task-server matching in the elite solution set. The scoring table output by the pairing network is introduced as a regularization term to smooth the probability table update, gradually concentrating the sampling distribution towards high-quality allocation schemes. Meanwhile, after obtaining the allocation solution in each sampling, a neighborhood search is introduced for local improvement, and the improved solution quality is used to participate in elite selection and probability update, thereby taking into account both global distribution search capability and local refinement capability, and improving the allocation quality and stability of the final task distribution.

[0035] Furthermore, the deep neighborhood search cross-entropy algorithm uses a combination of two optimization objectives to guide the learning of the neural network, namely... ,in It is a normalization factor; the neural network training objective is defined as:

[0036]

[0037] in Represents the expectation function, This represents the target weight. This represents the total processing time for task m. Represents the parameters of the neural network. Indicates the scoring table Induced sampling distribution, Represents a neighborhood search mechanism;

[0038] An unbiased estimation of the desired objective is performed using a policy gradient method based on REINFORCE.

[0039]

[0040] Where M represents the number of solutions sampled from the current policy during a single parameter update. Indicates a given task split instance Below, based on the parameters of the neural network The determined sampling distribution The generated k-th task assignment solution, where b is the baseline term. Indicates the training objective.

[0041] The present invention also provides a communication system capable of executing the above-described method for offloading UAV-assisted air-ground integrated missions based on privacy-latency co-optimization.

[0042] The symbols involved in this invention are summarized as follows:

[0043]

[0044] The beneficial effects of this invention are:

[0045] This invention addresses the task offloading scenario in integrated air-space-ground networks, proposing a joint method that integrates location privacy protection and task offloading optimization. By introducing an improved Laplace mechanism to randomly perturb the user's true location, the risk of attackers inferring the user's true location based on offloading decisions, link quality, and received signal strength is statistically reduced, thereby achieving provable location privacy protection.

[0046] Meanwhile, this invention fully considers the coverage constraints of satellites, UAVs, and ground edge nodes during the location perturbation process. By searching for the optimal perturbation radius and eliminating infeasible angle domains, it ensures that the generated false locations meet the feasibility requirements of task offloading, avoiding offloading failures or significant service quality degradation due to excessive perturbation. Furthermore, by adjusting the transmit power to mimic the received signal indication strength corresponding to the false location, the wireless physical layer characteristics are kept consistent with the location perturbation results, further enhancing the robustness and concealment of the privacy protection mechanism in real-world communication environments.

[0047] Furthermore, this invention combines a deep neighborhood search cross-entropy algorithm to optimize task routing decisions under perturbation conditions. While protecting user privacy, it effectively reduces system latency and improves computational resource utilization efficiency, achieving a good balance between location privacy protection and task routing performance. This method exhibits good adaptability and scalability, making it suitable for integrated air-space-ground network scenarios involving multiple satellites and UAVs. Attached Figure Description

[0048] Figure 1 A system model for a privacy protection task offloading scheme in a space-air-ground network.

[0049] Figure 2 It is the infeasible angular domain geometry outside the range that RSSI cannot imitate.

[0050] Figure 3 It is the infeasible angular domain geometry within the range that cannot be imitated by the user in RSSI.

[0051] Figure 4 This is a schematic diagram of the deep neighborhood search cross-entropy algorithm.

[0052] Figure 5 This diagram illustrates the random sampling of 100 points using the Laplace mechanism improved by this invention.

[0053] Figure 6 This represents the perturbation domain effect with different perturbation domain radii under different privacy budgets.

[0054] Figure 7 This represents the perturbation domain effect with different perturbation domain radii under different preference factor budgets.

[0055] Figure 8 This represents the training loss of the deep neighborhood search cross-entropy algorithm of this invention.

[0056] Figure 9 This refers to the latency under different task computation intensities.

[0057] Figure 10 This refers to the latency under different numbers of tasks.

[0058] Figure 11 This relates to the impact of different privacy budgets on privacy leaks.

[0059] Figure 12 This relates to the impact of different preference factors on privacy leaks. Detailed Implementation

[0060] To make the technical solution and advantages of the present invention clearer, the counting scheme of the present invention will be described in detail below.

[0061] This invention provides a privacy protection scheme for task offloading in an integrated air-space-ground network, the method comprising:

[0062] Step 1: Build a system model, jointly establish a communication model and a latency model, and on this basis, plan the task routing optimization problem with system latency and privacy as the core.

[0063] Step 1): Construct a system model, jointly establish a communication model and a latency model, and on this basis, plan a task routing optimization problem with system latency and privacy as the core.

[0064] This invention constructs a system model for studying task offloading methods in an air-space-ground network environment. The system consists of a satellite constellation and several unmanned aerial vehicles (UAVs). composition, This refers to the number of drones; both satellites and drones can serve as edge servers to provide computing services to users. (Location) A user generates a computing task that needs to be offloaded. To achieve fast task processing, the user tends to divide the computing task into multiple subtasks and distribute them to multiple edge servers for parallel processing. Let... Represents a set of subtasks. It is the number of subtasks. ( () indicates the data size of the subtask. Binary variable. Indicates the user's task Should I unload it onto the drone? ,like Indicates uninstallation, otherwise Binary variables Indicates the user's task Whether the unloading is done by the drone to the satellite, if Indicates uninstallation, otherwise .

[0065] The wireless channel between the user and the drone is characterized using a free-space path loss model. Based on this, and incorporating Shannon's theorem, and assuming the system uses Orthogonal Frequency Division Multiple Access (OFDMA), the user's data transmission rate is... It can be modeled as:

[0066]

[0067] in This is the user's available bandwidth. It is the user's transmit power. For channel noise, It's cell interference from other drones. It is the user and the drone The path gain between the user and the edge server is assumed to be affected by path loss, log-normal shadowing, and Rayleigh fading to account for factors such as shadowing and fading in the channel. Therefore, It can be modeled as:

[0068]

[0069] in It is a random variable representing the combined channel power gain under log-normal shadowing and Rayleigh fading, and its expectation is... . Indicates user to drone distance, It is the path loss factor.

[0070] The same data transmission rate between the drone and the satellite constellation. It can be modeled as:

[0071]

[0072] in This is the user's available bandwidth. It is the drone's transmission power. For channel noise, It is a drone Path gain with satellites.

[0073] Wireless transmission between user equipment and the drone carries a certain risk of bit errors. To characterize the impact of the wireless link on task offloading reliability, this invention uses bit error rate (BER) as a metric for link transmission reliability. Specifically, when the user is constantly loading the task... Unload to drone When the corresponding wireless transmission bit error rate is denoted as :

[0074] ,

[0075] It is a complementary error function. It is the signal-to-noise ratio when task m is unloaded onto the drone.

[0076] In summary, the probability that a single bit of data is correctly received during the offloading and transmission process is: Assuming that each bit transmission process is independent, the task The probability of successful uninstallation can be expressed as:

[0077]

[0078] The user will subtask The transmission time to the drone can be determined by Represented and calculated as:

[0079]

[0080] After transmission, the server performs calculations on the task, which takes time. It can be calculated as follows:

[0081]

[0082] in It is a drone computing power It is a subtask The computational intensity (CPU cycle / bit).

[0083] Similarly, drones will sub-tasks Transmission time to satellite can be determined by Represented and calculated as: After transmission, the server performs calculations on the task, which consumes computation time. It can be calculated as follows: , It refers to satellite computing power.

[0084] In summary, subtasks The task processing latency can be expressed as:

[0085]

[0086] and These represent the queuing delays for sub-tasks in the drone and satellite queues, respectively. Since the visibility time of a single satellite is limited, the satellites to which tasks may be unloaded may differ. This indicates the time it takes for mission data to be transmitted back to visible satellites.

[0087] Since its inception, differential privacy has been widely used in privacy protection research for data publishing due to its rigorous mathematical definition and sound theoretical properties. Differential privacy introduces a stochastic mechanism to provide provable privacy security for user data, thereby effectively defending against adversaries with side-information attack capabilities. Differential privacy refers to a stochastic mechanism... supply Differential secrecy, if for any dataset that is different on at most one element and and any output subset O,

[0088]

[0089] in This indicates the privacy budget. A smaller budget ensures stronger privacy protection.

[0090] Geographic indistinguishability, as an extension of differential privacy, is widely used in location privacy protection, with random mechanisms... For any location , and output position have:

[0091]

[0092] in express and The Euclidean distance between them. The geographical indistinguishability constraint mechanism applies to adjacent or nearby real locations. and Under these conditions, the output probabilities for the same false location should remain similar. Given the user's sensitive distance... and the maximum value of privacy leaks Location privacy leakage indicators can be defined as follows:

[0093]

[0094] This represents the decay factor of the privacy leakage indicator.

[0095] The optimization objective of this invention is to minimize user latency and privacy leakage. The problem is described as follows:

[0096]

[0097] Constraint C1 is the geographical indistinguishability requirement, constraint C2 indicates that user tasks can only be transmitted to satellites by drones, and constraint C3 indicates the drone queue capacity limit. This refers to the drone queue capacity. Constraints C4 and C5 indicate that a user's false location cannot fall within the range that the server's received signal strength cannot be imitated; that is, a user's false location must fall within the coverage area of ​​at least one server. This represents the distance between the spurious location and the drone's location u. This indicates the coverage radius of the drone u. Does this indicate the radius of the range where the strength of the signal received by the server cannot be imitated? C6 is the imitation power constraint, and C7 is the constraint that the total probability of successful offloading should be greater than a pre-specified reliability threshold.

[0098] Step 2: Introduce a perturbation radius search mechanism to determine the optimal perturbation radius that satisfies the system bit error rate constraint and minimizes system latency (i.e., privacy leakage).

[0099] This invention is based on the perturbation domain imitation benefit To search for the optimal perturbation radius, for each perturbation radius First, the region is discretized with a certain step size. For each discretized R, we determine the probability density function of its pseudo-position. Then, we sample a series of perturbation positions based on the probability density function and calculate the expected unloading utility and privacy leakage index of this series of perturbation positions, thus obtaining the region imitation utility. Finally, we take the R with the minimum imitation benefit as the optimal perturbation radius.

[0100] Centered on the user's real location, with a radius of The circular region is defined as the location perturbation region. To ensure the feasibility of task offloading, the spoofed location must be within the coverage area of ​​at least one edge server; therefore, a certain threshold is set. ,in This represents the maximum distance from the user to the edge server coverage boundary. A larger perturbation radius helps enhance privacy protection, but it inevitably reduces offloading effectiveness, making the appropriate setting of the perturbation range a key issue in balancing privacy and performance.

[0101] This invention uses , and To represent the location privacy leakage, uninstallation effectiveness, and uninstallation success probability when using a fake location for uninstallation, we can calculate the perturbation domain. The radius is Location privacy leaks Uninstallation effect and the probability of successful uninstallation Expectations:

[0102]

[0103] in The perturbation domain is False positions are generated at times The probability of.

[0104] In summary, the utility of the perturbation domain for:

[0105]

[0106] in It is a customized privacy factor that reflects the personalized location privacy requirements of mobile users. In real-world systems, users have different needs for location privacy protection due to differences in identity and preferences. For example, users involved in confidential work are more concerned about privacy than others. The larger the value, the more mobile users prioritize location privacy. This is achieved through adaptive adjustments. Mobile users can strike a balance between location privacy and computing costs based on their dynamic and personalized privacy needs.

[0107] The optimization problem of finding the optimal perturbation radius can be expressed as:

[0108]

[0109] The search algorithm of this invention solves the above optimization problem and obtains the optimal perturbation radius.

[0110] Step 3: Discuss the user's actual location, eliminate angle domains that cannot support task unloading, and build an improved Laplace position perturbation model to obtain the spurious position probability density function within the restricted angle domain.

[0111] Traditional Laplace distributions, while offering no restrictions on spoof locations for achieving geographic indistinguishability, are unsuitable for our scenario due to two constraints: First, to ensure reasonable task distribution, spoof locations must fall within the coverage area of ​​at least one edge server. Second, to ensure RSSI imitability, spoof locations should not be too close to any edge server. Third, the closer a spoof location is to a server, the greater the power required for imitation; if the spoof location is too close, the power required for imitation will exceed the maximum power limit. Therefore, each server has an unimitable RSSI distance. Our Laplace mechanism will classify the cases into two categories: 1. When the user's real location falls within the server's unimitable range; 2. When the user's real location is outside the server's unimitable range, such as... Figure 2 and Figure 3 As shown.

[0112] Scenario 1: When the user's actual location falls within the server's unmimicry range, for any location within the optimal perturbation range... This would be infeasible because without restricting r, when sampling in the 0 to 360 degree angle domain, false locations might fall within the server's imitation range. Therefore, we need to minimize r. According to the Laplace mechanism, the sampling angle follows a uniform distribution, and the sampling length follows a Laplace distribution. We can obtain the probability density function for generating false locations:

[0113]

[0114] in It is the truncation normalization factor. denoted by , where r represents the minimum disturbance radius and r represents the variable sampling radius. Indicates the sampling angle of the variable.

[0115] Scenario 2: When the user's actual location is outside the range that the server cannot mimic, the sampling angle range will be limited. We need to eliminate infeasible angle ranges. According to the Laplace mechanism, the sampling angle follows a uniform distribution, and the sampling length follows a Laplace distribution. We can obtain the probability density function for generating false locations:

[0116]

[0117] in It is the length of the feasible angle domain, calculated from the geometric relationship between the user and the server. The geometric diagram is provided by... Figure 2 As shown. Truncation normalization factor .

[0118] Step 4: For the generated false locations, calculate the transmit power required for the user to simulate the strength of the real received signal indication, and use the deep neighborhood search cross-entropy algorithm to determine the final task diversion scheme.

[0119] The strength of the received signal when the user sends data to the drone u at location q. Considering ,but Then it is necessary to simulate the received signal strength. Needs and Equal, that is Therefore, it can be calculated that:

[0120]

[0121] Indicates the user's real location With drone location The distance between them Indicates a fake user location With drone location The distance between them This indicates that the spurious position mimics power.

[0122] First, two independent neural networks are constructed to learn representations of task features and server features, respectively. The task encoding network takes information such as task size and task generation location as input and outputs a task vector representation. The server encoding network takes information such as server location, computing power, and the amount of remaining task data in the queue as input and outputs a server vector representation. Then, the task vector and server vector are input into a third pairing scoring network to model each task-server pair and output the corresponding non-negative score, forming a heuristic scoring table for task-server matching. Figure 4 As shown.

[0123] The Task Encoder employs a three-layer fully connected Multilayer Perceptron (MLP). It takes five dimensions of task features as input: computational requirements, data size, latency constraints, task generation location, and unloading success probability threshold. These features are then non-linearly mapped through hidden layers of 64 and 32 dimensions, ultimately outputting a 16-dimensional task embedding vector to characterize the comprehensive features of the task at both computational and communication levels. The Server Encoder also uses a three-layer fully connected MLP. Its input consists of three dimensions of server-related features: server computational power, resource status, and the distance between the user and the server. The distance feature explicitly incorporates user-server spatial association information. After mapping through hidden layers of 64 and 32 dimensions, it outputs a 16-dimensional server embedding vector to uniformly represent the server's computational power and location-related characteristics. Building upon this, the Pair Scorer network concatenates the task embedding and the server embedding to form a 32-dimensional joint feature representation. This representation is then processed by a three-layer fully connected MLP (with hidden layer dimensions of 64 and 32 respectively) to output the matching score of the task-server pair. This score characterizes the relative suitability of different servers for the same task, providing a heuristic evaluation basis for subsequent task allocation and optimization algorithms.

[0124] This invention uses a combination of two optimization objectives to guide the learning of a neural network, namely... ,in It is the normalization factor. The training objective of the neural network can be defined as follows:

[0125]

[0126] in Represents the expectation function, Indicates the target weight. This represents the total processing time for task m. Represents the parameters of the neural network. Indicates the scoring table Induced sampling distribution. Representing the neighborhood search mechanism, early training from Initially, ensure that a feasible solution is learned; increase the number of solutions later in the training process. The value heuristic is "friendly to local search". Since the allocation decision is a discrete variable, this invention uses a policy gradient method based on REINFORCE to perform an unbiased estimation of the desired objective:

[0127]

[0128] Where M represents the number of solutions sampled from the current policy during a single parameter update. Indicates a given task split instance Below, based on the parameters of the neural network The determined sampling distribution The generated k-th task assignment solution. b is the baseline term, which is usually taken as the average value of the sampled solution loss in the current batch. This indicates training loss.

[0129] After the neural network is trained, an initial probability table is generated to provide a reasonable prior distribution for the cross-entropy algorithm, thus accelerating convergence. During the iteration process, the prior probabilities constructed from the scoring table are introduced into the probability update driven by elite samples as a regularization term to smooth the probability table update, thereby suppressing sampling noise and avoiding premature convergence. Finally, the algorithm outputs the optimal solution for task allocation to the server.

[0130] The proposed algorithm was validated in a simulation environment and compared with three other schemes:

[0131] 1. Three-Network Guided Cross-Entropy Optimization Algorithm (TN-CEM): The proposed solution in this invention.

[0132] 2. Three-Network Guided Optimization Algorithm (TN): A variant of the scheme proposed in this invention.

[0133] 3. Guard Algorithm (GUARD): Generates fake locations through differential privacy, but does not have power mimicry and only selects one server for offloading.

[0134] 4. Baseline Algorithm (NORMAL): Unloading is not performed using fake locations; tasks are evenly distributed across unloadable servers.

[0135] Figures 6-7 This represents the perturbation and utility at different radii under different parameter conditions, and determines the optimal parameters and the optimal radius. Under different task computation intensities ( Figure 9 ) and number of tasks ( Figure 10Under these conditions, the latency of the GUARD scheme is significantly higher than the other three schemes (TN-CEM, TN, and NORMAL), especially in high-load scenarios (such as a computation intensity of 1000 cycles / bit or 50 tasks), where its latency can reach more than 8 seconds, indicating a high performance cost. In contrast, the TN-CEM and TN schemes exhibit similar and lower latency in most cases, outperforming the NORMAL scheme, demonstrating that introducing the task-aware mechanism (TN) and optimized scheduling (CEM) can effectively balance efficiency and resource consumption.

[0136] In terms of privacy protection, Figure 11 (Changes in privacy budget) and Figure 12 (Changes in preference factors) show that as the privacy budget increases or the user's preference factor decreases (i.e., higher privacy requirements), the privacy leakage values ​​of all solutions show a downward trend, which is in line with expectations. However, it is worth noting that the privacy leakage value of the NORMAL solution remains unchanged under different conditions because it lacks a privacy protection mechanism. In contrast, the TN-CEM and TN solutions maintain relatively stable low leakage levels under different parameters, demonstrating stronger privacy adaptability and robustness. In summary, the TN-CEM solution performs best in balancing low latency and controllable privacy leakage.

[0137] The following is the pseudocode of the algorithm designed in this invention:

[0138]

[0139]

[0140] The above description describes specific embodiments of the present invention and the technical principles employed. Any changes made in accordance with the concept of the present invention that do not exceed the spirit of the specification and drawings should still fall within the protection scope of the present invention.

Claims

1. A method for drone-assisted air-ground integrated mission offloading based on privacy-latency collaborative optimization, characterized in that, Includes the following steps: A system model is constructed, and a communication model and a latency model are jointly established. Based on this, a task routing optimization problem is built to minimize user latency and privacy leakage. The optimization problem is as follows: This indicates a false location after user disturbance. Represents the user's subtask Should I unload it onto the drone? , Represents the user's subtask Whether the drone unloads the contents onto the satellite, The simulated power calculated based on the spurious position, Subtasks Task processing latency, Represents a set of subtasks. Indicators of location privacy leaks , This indicates the real locations of two different users. Indicates privacy budget, express and The Euclidean distance between them Indicates the amount of data in the subtask. It is the capacity of the drone queue. This indicates the coverage radius of the drone u. This indicates the radius of the range where the signal strength received by the server cannot be imitated. This represents the distance between the spurious location and the drone's location u. Indicates maximum power limit. Indicates task The probability of successful uninstallation. This represents the threshold for the probability of successful uninstallation. C1 is the geographical indistinguishability requirement; constraint C2 means that user tasks can only be transmitted to satellites by drones; constraint C3 means the drone queue capacity limit; constraints C4 and C5 mean that the user's false location cannot fall within the range that the server's received signal strength cannot be imitated, that is, the user's false location must fall within the coverage area of ​​at least one server; C6 is the imitation power constraint; and constraint C7 means that the total probability of successful offloading should be greater than the pre-specified reliability threshold. A perturbation radius search mechanism is introduced to determine the optimal perturbation radius that satisfies the system bit error rate constraint and minimizes system latency and privacy leakage; Based on the user's actual location, the angle domain that cannot support task unloading is eliminated. On this basis, an improved Laplace position perturbation model is constructed to obtain the spurious position probability density function in the restricted angle domain. For the generated false locations, the transmit power required for the user to mimic the strength of the real received signal indication is calculated, and the final task offloading scheme is determined using a deep neighborhood search cross-entropy algorithm.

2. The UAV-assisted air-ground integrated mission offloading method based on privacy-latency collaborative optimization according to claim 1, characterized in that: The construction of the communication model specifically includes: the wireless channel between the user and the drone adopts a free-space path loss model, combined with Shannon's theorem, and assumes that the system uses orthogonal frequency division multiple access, and the user's data transmission rate... The model is as follows: in This is the user's available bandwidth. It is the user's transmit power. For channel noise, It's cell interference from other drones. It is the user and the drone Path gain between The model is as follows: in It is a random variable representing the combined channel power gain under log-normal shadowing and Rayleigh fading, and its expectation is... , Indicates user to drone distance, It is the path loss factor; The same data transmission rate between the drone and the satellite constellation. The model is as follows: in This is the user's available bandwidth. It is the drone's transmission power. For channel noise, It is a drone Path gain with satellites, Indicates drone Distance to satellite This represents the channel gain from the UAV to the satellite; When the user assigns subtasks at any time Unload to drone When the corresponding wireless transmission bit error rate is denoted as : ， For complementary error functions, This represents the signal-to-noise ratio when task m is unloaded onto drone u; In summary, the probability that a single bit of data is correctly received during the offloading and transmission process is: Assuming that each bit transmission process is independent, the task The probability of successful uninstallation is expressed as: The latency model is: subtask Task processing latency is expressed as: and These represent the queuing delays of subtasks in the drone and satellite queues, respectively. This indicates the time it takes for mission data to be transmitted back to visible satellites. This indicates that the user will assign subtasks. Transmission time to the drone This indicates the time taken for the drone to perform the task. Indicates that the drone will perform sub-tasks Transmission time to the satellite This indicates the time taken for the satellite to perform calculations on the mission.

3. The UAV-assisted air-ground integrated mission offloading method based on privacy-latency collaborative optimization according to claim 1, characterized in that: The introduced perturbation radius search mechanism determines the optimal perturbation radius that satisfies the system bit error rate constraint while minimizing system latency and privacy leakage. Specifically, this includes: Centered on the user's real location, with a radius of The circular region is defined as the position disturbance region, and is set as follows: ,in Indicates the maximum distance from the user to the edge server coverage boundary; using , and This indicates the location privacy leakage, uninstallation effectiveness, and uninstallation success probability when using a fake location for uninstallation, and calculates the perturbation domain. The radius is Location privacy leaks Uninstallation effect and the probability of successful uninstallation Expectations: in The perturbation domain is False positions are generated at times The probability of; The utility of the perturbation domain for: in It is a customized privacy factor that reflects the personalized location privacy requirements of mobile users; The optimization problem of finding the optimal perturbation radius is expressed as: The above optimization problem is solved using a search algorithm to obtain the optimal perturbation radius.

4. The UAV-assisted air-ground integrated mission offloading method based on privacy-latency collaborative optimization according to claim 1, characterized in that: The improved Laplace position perturbation model is constructed to obtain the false position probability density function in the restricted angle domain, including two cases:

1. When the user's real position falls within the server's unmimicry range; 2. The user's actual location is outside the range that the server cannot imitate; Scenario 1: When the user's real location falls within the server's unassimilable range, a minimum constraint is applied to r. Based on the Laplace mechanism, the sampling angle follows a uniform distribution, and the sampling length follows a Laplace distribution. The probability density function for generating false locations is as follows: in It is the truncation normalization factor. The minimum disturbance distance is represented by r, and the variable sampling radius is represented by r. Indicates the sampling angle of the variable; Scenario 2: When the user's actual location is outside the range that the server cannot mimic, the sampling angle range will be limited. We need to eliminate infeasible angle ranges. According to the Laplace mechanism, the sampling angle follows a uniform distribution, and the sampling length follows a Laplace distribution, resulting in the probability density function for generating false locations: in It is the length of the feasible angular domain, and the truncation normalization factor. .

5. The UAV-assisted air-ground integrated mission offloading method based on privacy-latency collaborative optimization according to claim 1, characterized in that: The calculation of the transmit power required by the user to simulate the actual received signal strength specifically includes: The strength of the received signal when the user sends data to the drone u at location q. Considering ,but Then it is necessary to simulate the received signal strength. Needs and Equal, that is Therefore, it can be calculated that: Indicates the user's real location With drone location The distance between them Indicates a fake user location With drone location The distance between them This indicates that the spurious position mimics power.

6. The UAV-assisted air-ground integrated mission offloading method based on privacy-latency collaborative optimization according to claim 1, characterized in that: The deep neighborhood search cross-entropy algorithm includes: first, encoding task features and server features using two independent neural networks to obtain task vector representations and server vector representations; then, inputting these two types of vectors into a third pairing scoring network to output a non-negative score for each task assigned to each server, and normalizing it to form a prior probability distribution for task-server matching, which is used to initialize the probability table of the cross-entropy method; during the solution process, multiple allocation solutions are constructed by sampling each task within the feasible server set according to the current probability table, the objective function is calculated for each set of solutions, and the best set of solutions is selected as the elite solution set. Then, the probability table is updated according to the occurrence frequency of task-server matching in the elite solution set, and the scoring table output by the pairing network is introduced as a regularization term to smooth the update of the probability table, so that the sampling distribution gradually concentrates towards high-quality allocation schemes.

7. The UAV-assisted air-ground integrated mission offloading method based on privacy-latency collaborative optimization according to claim 6, characterized in that: The task encoding network that encodes task features adopts a three-layer fully connected multilayer perceptron. It takes five-dimensional task features, including task computation requirements, data scale, latency constraints, task generation location, and unloading success probability threshold, as input. These features are then passed through hidden layers of 64 and 32 dimensions for nonlinear mapping, and finally output a 16-dimensional task embedding vector to characterize the comprehensive features of the task at the computation and communication levels. The network that encodes the features adopts a three-layer fully connected multilayer perceptron. Its input consists of three-dimensional features: server computing power, resource status, and distance between the user and the server. After being mapped through hidden layers of 64 and 32 dimensions, it outputs a 16-dimensional server embedding vector to uniformly represent the server's computing power and location-related characteristics. The pairing scoring network concatenates the task embedding and the server embedding to form a 32-dimensional joint feature representation, and outputs the matching score of the task-server pair through a three-layer fully connected MLP.

8. The UAV-assisted air-ground integrated mission offloading method based on privacy-latency collaborative optimization according to claim 6 or 7, characterized in that: The deep neighborhood search cross-entropy algorithm uses a combination of two optimization objectives to guide the learning of the neural network, namely... ,in It is a normalization factor; the neural network training objective is defined as: in Represents the expectation of the function. Indicates the target weight. This represents the total processing time for task m. Represents the parameters of the neural network. Indicates the scoring table Induced sampling distribution, Represents a neighborhood search mechanism; An unbiased estimation of the desired objective is performed using a policy gradient method based on REINFORCE. Where M represents the number of solutions sampled from the current policy during a single parameter update. Indicates a given task split instance Below, based on the parameters of the neural network The determined sampling distribution The generated k-th task assignment solution, where b is the baseline term. For training purposes.

9. A communication system, characterized in that: It is capable of executing the UAV-assisted air-ground integrated mission offloading method based on privacy-latency collaborative optimization as described in any one of claims 1-8.

Images