A privacy-preserving, fair social attention allocation method based on edge computing
By combining local perception of edge terminal devices and a social gravity field model with dynamic privacy budget and fairness penalty terms, the problem of high complexity and insufficient privacy protection in social influence assessment in edge computing is solved, and efficient and interpretable social influence assessment and fair resource allocation are achieved on edge terminal devices.
Patent Information
- Application Number
- CN202611284756.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-24
- Publication Date
- 2026-11-17
AI Technical Summary
Existing technologies for evaluating social influence in edge computing scenarios suffer from high computational complexity, insufficient privacy protection, and unfair resource allocation, making them unable to effectively adapt to the computing power limitations and dynamic environmental changes of edge terminal devices.
By locally sensing raw social interaction data on edge terminal devices, privacy-preserving feature extraction and social gravity field model calculation are performed. Combined with dynamic privacy budget and fairness penalty terms, desensitized social state vectors are generated, and joint optimization model solutions are performed on the edge server to achieve resource scheduling.
It enables efficient and interpretable social influence assessment on edge terminal devices, dynamically adjusts the strength of privacy protection, ensures the fairness of resource allocation and the availability of terminal devices, and adapts to highly dynamic mobile scenarios.
Smart Images

Figure CN123118085A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the field of edge computing and social network resource allocation, and particularly relates to a privacy protection type social attention fair allocation method based on edge computing. BACKGROUND
[0002] With the rapid development of mobile Internet and Internet of Things, the scale and complexity of social networks continue to rise, and the amount of interaction data generated by users on social platforms grows explosively. Social attention resources, i.e. network resources used for content distribution, message pushing, bandwidth allocation, and other services in social networks, have become a scarce and critical resource. How to fairly and efficiently allocate limited social attention resources to a large number of users is one of the core technical problems faced by social network operation and mobile edge computing fields. The existing social attention resource allocation technical solutions mainly have the following outstanding problems:
[0003] 1. The allocation of social attention resources now usually needs to first evaluate the influence of each user in the social network, so as to differentially configure resources according to the size of the influence, wherein the evaluation of social influence generally depends on deep learning models such as graph neural network (GNN), which extracts high-order topological features of nodes by performing neighborhood message passing and multi-layer aggregation operations on node embeddings in social graphs. However, the core operation of graph neural network involves matrix multiplication operations between large-scale adjacency matrices and feature matrices and backpropagation parameter updates, and its computational complexity is usually high, and its training process requires a large amount of labeled data and multiple rounds of iterative optimization. In the edge computing scenario, the computing power, memory and energy resources of edge terminal devices (such as smartphones, vehicle-mounted terminals, wearable devices, etc.) are extremely limited, and they cannot bear the training or online inference calculation of graph neural network; in addition, graph neural network is essentially a "black box" model, and its feature extraction process lacks explainability, and the output results of the influence score cannot be effectively audited and traced, making it difficult to meet the requirements of decision transparency in resource allocation scenarios.
[0004] 2. The privacy protection of social interaction data has not been adequately addressed. Calculating social influence requires users' social interaction data (such as communication records, message frequency, and contact relationships) as input. In existing technologies, most solutions adopt an architecture of centralized data upload and unified server-side computation. That is, each terminal device uploads its raw social interaction data to a central or edge server, where the server centrally performs influence assessment and resource allocation calculations. This architecture poses a serious privacy risk: raw social interaction data may be intercepted, leaked, or misused during transmission and storage. Attackers can deduce a user's social graph structure, interaction patterns, social circles, and even personal identity information from the raw data. While some existing solutions introduce certain privacy protection measures, these measures only protect the security of data during transmission. Once the data reaches the server and is decrypted, the server can still fully obtain the user's original social relationship information, failing to fundamentally achieve the privacy protection goal of data being usable but invisible. More importantly, the privacy protection strength in existing solutions is usually a statically configured fixed parameter that cannot be dynamically adjusted according to the security status of the user's current physical and network environment. For example, the level of privacy threat faced by a user in a public Wi-Fi environment is obviously higher than that in a home encrypted network, but static privacy protection solutions apply the same level of protection to both situations, resulting in insufficient protection in high-risk environments and unnecessary sacrifice of data availability in low-risk environments.
[0005] In summary, there is an urgent need for a technical solution that can efficiently complete social influence assessment under the condition of limited computing power of edge terminal devices, achieve dynamic and adaptive privacy protection locally, and ensure the fairness of resource allocation at the global level through an effective fairness constraint mechanism. Summary of the Invention
[0006] One of the objectives of this invention is to provide a privacy-preserving, fair allocation method for social attention based on edge computing, in order to solve the problems of high computational complexity of social influence, insufficient privacy protection, lack of fairness in resource allocation, and weak adaptability to edge scenarios in existing technologies.
[0007] This invention is achieved through the following technical solution: a privacy-preserving social attention fairness allocation method based on edge computing, comprising the following steps: acquiring an original social interaction dataset within a preset time window based on the local perception of the edge terminal device; wherein, the original social interaction dataset includes the interaction frequency, interaction duration, and content interaction intensity between the user of the edge terminal device and other nodes in the social network; performing local privacy feature extraction on the original social interaction dataset to generate a desensitized social state vector; wherein, the local privacy feature extraction includes: constructing a three-dimensional feature matrix containing interaction breadth, interaction depth, and response timeliness based on the original social interaction dataset; performing aggregation operations on the three-dimensional feature matrix to obtain an original social influence score; and employing a Laplace mechanism, based on privacy budget parameters and the... The system generates random noise data based on the global sensitivity of the original social influence score; and superimposes the random noise data onto the original social influence score to generate the desensitized social state vector; the edge server constructs a joint optimization model containing a system total utility objective function and a fairness penalty term based on the received desensitized social state vectors from multiple edge terminal devices, and solves the joint optimization model to generate an attention allocation weight set; wherein, the system total utility objective function is constructed based on the desensitized social state vectors and logarithmic utility functions of each edge terminal device, and the fairness penalty term is constructed based on the Jain fairness index; and the attention allocation weights corresponding to each edge terminal device in the attention allocation weight set are distributed to each edge terminal device, and corresponding resource scheduling operations are performed according to the attention allocation weights.
[0008] Furthermore, the fairness penalty term is configured as a non-linear penalty function, and the value of the fairness penalty term increases exponentially when the Jain fairness index is lower than a preset fairness threshold.
[0009] Furthermore, the Jain fairness index can be calculated using the following formula:
[0010]
[0011] The fairness penalty is calculated using the following formula:
[0012]
[0013] in, As a fairness threshold, This is the penalty sensitivity coefficient;
[0014] Furthermore, the objective function of the joint optimization model is: ,in, Let the total utility objective function be the system. As a fairness penalty item, Here are the fairness weight coefficients; the constraints of the joint optimization model are: ,in, This represents the upper limit of the total amount of resources.
[0015] Furthermore, before acquiring the original social interaction dataset within a preset time window, the method further includes: an adaptive sampling step based on device hardware status, which involves real-time monitoring of the hardware resource status parameters of the edge terminal device, including the remaining battery percentage and processor load index; calculating a hardware resource pressure index based on the remaining battery percentage and processor load index; and determining an adaptive sampling frequency based on the hardware resource pressure index through a negative exponential mapping relationship, and collecting the original social interaction dataset according to the adaptive sampling frequency.
[0016] Furthermore, the hardware resource pressure index can be calculated using the following formula:
[0017] ,
[0018] in, For edge terminal devices exist The hardware resource pressure index at any given time. For edge terminal devices exist Remaining battery level at any given time The preset battery alarm threshold, For edge terminal devices exist The processor core temperature corresponding to the processor load index at any given time. The preset temperature alarm threshold, This is the maximum temperature that the processor can withstand. This is the scaling factor for the electrical pressure. This is the scaling factor for temperature and pressure.
[0019] Furthermore, the adaptive sampling frequency is calculated using the following formula:
[0020] ,in, For the adaptive sampling frequency, The preset maximum sampling frequency, This is the attenuation sensitivity coefficient.
[0021] Furthermore, when the hardware resource pressure index is greater than a preset pressure threshold, during the process of collecting data at an adaptive sampling frequency, only the interaction records of social objects whose interaction frequency ranks in the top preset order with the user of the edge terminal device are retained, and the remaining interaction records are discarded to generate the sparse original social interaction dataset.
[0022] Further, the three-dimensional feature matrix is aggregated to obtain the original social influence score, including: using an edge social gravity field model to aggregate the three-dimensional feature matrix, whereby the three dimensions of the three-dimensional feature matrix are mapped to three physical quantities in a physical gravity model: the interaction breadth is mapped to equivalent mass, which is the number of unique social objects that have at least one interaction record with the user of the edge terminal device within the preset time window; the interaction depth is mapped to equivalent potential field distance, which is the reciprocal of the weighted combination of interaction frequency and interaction duration; the response timeliness is mapped to an exponential time decay factor, which is the negative exponential decay function value with the average message response time difference as the independent variable; and for each neighbor node of the node corresponding to the user of the edge terminal device, the product of equivalent mass and the reciprocal of equivalent potential field distance is calculated, and then multiplied by the corresponding exponential time decay factor, and the calculation results of all neighbor nodes are summed to obtain the original social influence score.
[0023] Furthermore, the original social influence score can be calculated using the following formula:
[0024]
[0025] in, For nodes exist The original social influence score of the moment. For nodes The set of neighboring nodes, For nodes Equivalent quality, For nodes With nodes The equivalent potential field distance between them For nodes With nodes The exponential time decay factor between them.
[0026] Furthermore, the equivalent potential distance can be calculated using the following formula:
[0027] ,in, For nodes With nodes Interaction frequency within a preset time window, For nodes With nodes Interaction duration within a preset time window, The weighting coefficient for interaction frequency. The weighting coefficient for interaction duration. To prevent division by zero of positive numbers.
[0028] Furthermore, the exponential time decay factor can be calculated using the following formula:
[0029] ,in, This is the forgetting rate coefficient. For nodes With nodes The average response time difference between message exchanges.
[0030] Furthermore, the privacy budget parameter is a dynamic privacy budget, which is adaptively adjusted based on the environmental sensitivity assessment value of the current environment in which the edge terminal device is located.
[0031] Furthermore, the dynamic privacy budget is calculated using the following formula:
[0032]
[0033] in, For edge terminal devices exist Dynamic privacy budget at all times For the preset baseline privacy budget, As a sensitivity modulator, For edge terminal devices exist Environmental sensitivity assessment value at any given time.
[0034] Furthermore, the desensitized social state vector is generated using the following formula:
[0035]
[0036] in, For desensitized social state vectors, Score the original social influence. To obtain the scale parameter is Random noise data drawn from the Laplace distribution. For global sensitivity.
[0037] Furthermore, the environmental sensitivity assessment value is determined comprehensively based on the location information of the edge terminal device and the encryption protocol status of the currently connected network; wherein, when the edge terminal device is located in a pre-marked sensitive geofence area, or when the currently connected network does not use an encryption protocol with a preset security level, the environmental sensitivity assessment value increases, and the dynamic privacy budget decreases accordingly.
[0038] Furthermore, before constructing the joint optimization model, the edge server further includes: using the Kalman filter algorithm to predict the disconnection probability of each edge terminal device at the next moment based on the historical connection state sequence of each edge terminal device; and calculating a topology stability factor based on the disconnection probability, wherein the topology stability factor is the complement of the disconnection probability; wherein the system total utility objective function also includes the topology stability factor, such that the edge terminal device with a higher disconnection probability obtains a lower attention allocation weight in the attention allocation weight set.
[0039] Furthermore, the overall system utility objective function can be calculated using the following formula:
[0040] ,in, This represents the total number of edge terminal devices. Assign a set of weights to the attention problem to be solved. To allocate to edge terminal devices The attention allocation weights, For edge terminal devices The aforementioned topological stability factor, For edge terminal devices The uploaded de-identified social state vector, It is a logarithmic utility function.
[0041] Furthermore, the joint optimization model is solved using the alternating direction multiplier method, including: introducing a global consensus variable and a dual variable, transforming the joint optimization model into an augmented Lagrangian function; in each iteration, the following alternating updates are performed: each edge terminal device solves its own local optimization subproblem in parallel, updating its own local allocation weights, each local optimization subproblem only involving the desensitized social state vector of the corresponding edge terminal device and the global consensus variable and the dual variable received from the edge server; under the constraint of the upper limit of total resources, the edge server updates the global consensus variable based on the updated local allocation weights of each edge terminal device, combined with the fairness penalty term; and updates the dual variable based on the residual between the updated local allocation weights and the updated global consensus variable; the alternating updates are repeated until a preset convergence condition is met, and the converged global consensus variable or the local allocation weights are used as the attention allocation weight set.
[0042] Furthermore, the augmented Lagrangian function is:
[0043]
[0044] in, The vector formed by assigning local weights to each edge terminal device. Let be the vector formed by the global consensus variables. The vector formed by the dual variables. For fairness weighting coefficients, This is the step size penalty parameter. It is the topological stability factor. Let be the desensitized social state vector. For the fairness penalty item;
[0045] The local optimization subproblems for each edge terminal device are:
[0046]
[0047] The global update of the edge server is as follows:
[0048]
[0049] The update of the dual variable is as follows: , where superscript Indicates the first Round iteration, This represents the upper limit of the total amount of resources.
[0050] Further, the corresponding resource scheduling operation is performed according to the attention allocation weight, including at least one of the following: adjusting the length of the content recommendation cache queue corresponding to each edge terminal device proportionally according to the attention allocation weight; allocating a dedicated transmission time slot for edge terminal devices whose attention allocation weight is higher than the high priority threshold; and scheduling the non-urgent data transmission request of the edge terminal device to a non-congestion period for execution for edge terminal devices whose attention allocation weight is lower than the low priority threshold.
[0051] Furthermore, the edge computing system can be applied to vehicle-to-everything (V2X) scenarios, where the edge terminal device is an in-vehicle terminal unit and the edge server is a roadside unit or a multi-access edge computing server on the base station side.
[0052] Another aspect of the present invention provides a privacy-preserving social attention fairness allocation system based on edge computing, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements any of the privacy-preserving social attention fairness allocation methods based on edge computing as described above.
[0053] Compared with the prior art, the present invention has the following advantages and beneficial effects:
[0054] 1. This invention achieves dynamic adjustment of sampling frequency according to device health status through an adaptive sparse perception model that maps hardware resource pressure index and negative exponent. While ensuring data quality, it effectively extends the available working time of terminal devices. At the feature extraction level, it replaces the traditional graph neural network scheme with an edge social gravity field model, reducing the calculation of social influence score from high-order matrix operations to linear scalar algebra operations. This successfully adapts to the limited computing power of edge terminals while maintaining white-box interpretability.
[0055] 2. At the privacy protection level, this invention achieves adaptive adjustment of privacy protection strength according to environmental sensitivity through a context-aware dynamic privacy budget mechanism and Laplace noise injection. It strengthens protection in highly sensitive environments and preserves data availability in secure environments. All privacy operations are completed locally on the terminal, and the original data does not leave the terminal. At the global decision-making level, through the dual constraints of the marginal diminishing property of the logarithmic utility function and the nonlinear exponential explosion penalty term based on the Jain exponent, the fairness of resource allocation is guaranteed from both the mechanism design and mathematical constraints, effectively breaking the Matthew effect caused by traditional greedy algorithms.
[0056] 3. At the optimization solution level, this invention employs ADMM to decompose the globally coupled problem into local subproblems that can be solved in parallel on each terminal and a global coordination problem that can be executed on the server side. Each iteration only involves communication. This scalar addresses the communication congestion and privacy leakage issues of centralized solutions. Furthermore, by using the topology stability factor predicted by Kalman filtering, it enables the forward-looking deweighting of nodes about to disconnect from the network, effectively reducing resource waste and system jitter caused by topology changes. This approach is applicable to highly dynamic mobile scenarios such as vehicle-to-everything (V2X) networks. Attached Figure Description
[0057] The accompanying drawings, which are included to provide a further understanding of embodiments of the invention and form part of this application, do not constitute a limitation thereof. In the drawings:
[0058] Figure 1 The above is a flowchart of the overall method provided in Embodiment 1 of the present invention.
[0059] Figure 2 This is a schematic diagram of the power dimension comparison curve provided in Embodiment 1 of the present invention.
[0060] Figure 3 This is a heat map of hardware resource pressure index provided in Embodiment 1 of the present invention.
[0061] Figure 4 This is a schematic diagram comparing the computational complexity provided in Embodiment 1 of the present invention.
[0062] Figure 5 This is a schematic diagram illustrating the contribution of social features to influence scoring in Embodiment 1 of the present invention.
[0063] Figure 6 This is a schematic diagram of the adaptive change curve provided in Embodiment 1 of the present invention.
[0064] Figure 7 This is a schematic diagram showing the distribution comparison of desensitization scores provided in Embodiment 1 of the present invention.
[0065] Figure 8 This is a schematic diagram of the convergence process provided in Embodiment 1 of the present invention.
[0066] Figure 9 This is a schematic diagram of the characteristic curves provided in Embodiment 1 of the present invention.
[0067] Figure 10 This is a Pareto trade-off curve provided in Embodiment 1 of the present invention. Detailed Implementation
[0068] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.
[0069] Example 1
[0070] This embodiment discloses a privacy-preserving, fair allocation method for social attention based on edge computing. Figure 1 The overall method flowchart of this embodiment is shown. As can be seen from the figure, this embodiment includes the following steps:
[0071] Step 1: Obtain the original social interaction dataset based on the local perception of edge terminal devices.
[0072] Edge terminal devices refer to intelligent terminals deployed at the network edge that possess certain local computing capabilities and sensor resources. They can independently complete lightweight computing tasks such as data collection and preprocessing without relying on cloud servers. For example, edge terminal devices can be smartphones carried by users, onboard units (OBUs) installed in vehicles, or gateway devices in smart home scenarios.
[0073] Local sensing refers to the process by which edge terminal devices use their onboard or accessible sensors, communication modules, and operating system-level interfaces to collect raw data related to the social interaction behavior of the device's user in real time. This sensing process is executed entirely on the terminal side and does not involve uploading the raw data to any external server or third-party platform.
[0074] The raw social interaction dataset refers to a structured data set collected by edge terminal devices through local sensing mechanisms within a preset time window, reflecting the interactive behavior characteristics between the device user and other nodes in the social network. This dataset includes, but is not limited to, the following three core dimensions of interaction metrics: interaction frequency, interaction duration, and content interaction intensity.
[0075] Interaction frequency refers to the cumulative number of communication interactions (including but not limited to message sending / receiving, voice call initiation / answering, video connection establishment, etc.) between a user of the edge terminal device and a specific social object within a preset time window. A higher interaction frequency indicates more frequent social connections between the two nodes.
[0076] Interaction duration refers to the sum of the cumulative durations of all interactive events between a user of an edge terminal device and a specific social object within a preset time window. For example, if a voice call lasts 5 minutes and a video call lasts 10 minutes, then the interaction duration with that object within that time window is 15 minutes. The longer the interaction duration, the deeper the social connection between the two nodes.
[0077] Content interaction intensity refers to a comprehensive quantitative indicator of the richness and responsiveness of the interactive content between a user of an edge terminal device and a specific social object within a preset time window. For example, content interaction intensity can be measured by the average response time difference between message exchanges—the shorter the response time, the higher the responsiveness and the greater the content interaction intensity; conversely, the longer the response time, the lower the responsiveness and the smaller the content interaction intensity.
[0078] A preset time window refers to a fixed-length historical time interval set in advance to define the time range for data collection. For example, a preset time window could be the past 24 hours, the past week, or the past month. This time window slides and updates as the system progresses, ensuring that the collected data always reflects the user's recent social behavior characteristics.
[0079] It should be noted that the acquisition of the original social interaction dataset in this embodiment is carried out in strict compliance with relevant laws and regulations. Specifically, before collecting any user's social interaction data, the application running on the edge terminal device will clearly inform the user through a clear and explicit privacy notification interface about the specific purpose of data collection (i.e., for the fair allocation and optimization of social attention resources), the scope of collection (limited to statistical indicators such as interaction frequency, interaction duration, and content interaction intensity), the data processing method (feature extraction is performed locally and differential privacy protection is applied before uploading anonymized data; the original data does not leave the terminal device), and the data storage period (data is retained only within a preset time window; expired data is automatically cleared after the window is slid), and obtain the user's explicit and separate consent. All data collection activities strictly comply with the legal basis requirements for personal information processing under Article 13 of the Personal Information Protection Law of the People's Republic of China, the provisions on data security guarantee obligations under the Data Security Law of the People's Republic of China, and the provisions on network information security under the Cybersecurity Law of the People's Republic of China. Furthermore, the user has the right to withdraw their authorization consent at any time through the terminal device's settings interface. After withdrawal, the system will immediately stop collecting the user's social interaction data and delete or anonymize the collected data in accordance with legal provisions. Furthermore, the social interaction data collected in this embodiment is strictly limited to statistical interaction indicators directly related to the attention resource allocation task, and does not involve users' communication text content, voice content, image content, personal identification information, biometric information, or other sensitive personal information unrelated to the resource allocation task. The principle of minimum necessary personal information processing is implemented from the source of data collection.
[0080] In addition, in this embodiment, obtaining the original social interaction dataset also includes an adaptive sampling step based on the device hardware status, which is used to achieve refined management of the energy consumption of edge terminal devices while ensuring the quality of data collection.
[0081] Adaptive sampling refers to a mechanism where edge terminal devices dynamically adjust the frequency of collecting social interaction data based on their current hardware resource status. The motivation behind this mechanism is that in edge computing scenarios, the hardware resources of terminal devices (such as battery power and heat dissipation capacity) are extremely limited. Existing technologies generally employ continuous, fixed-frequency data collection strategies. This always-on monitoring mechanism leads to rapid device overheating and power consumption, severely shortening the terminal's usable operating time. Therefore, it is necessary to construct an adaptive sparse sensing model capable of sensing its own hardware status, enabling the data collection frequency to be dynamically adjusted according to the device's current health condition, ensuring data quality while achieving refined management of device energy consumption.
[0082] In this embodiment, the adaptive sampling step may include:
[0083] Real-time monitoring of hardware resource status parameters of edge terminal devices, including remaining power percentage and processor load indicators.
[0084] The remaining battery percentage refers to the ratio of the remaining usable battery power to the full battery capacity of the edge terminal device, reflecting the device's battery life. The processor load metric is a physical quantity reflecting the current workload of the central processing unit (CPU) of the edge terminal device, which can be characterized by CPU utilization or CPU core temperature. It is understood that there is a strong positive correlation between CPU utilization and CPU core temperature—a sustained high CPU utilization will inevitably lead to a rise in CPU core temperature. As a hardware-level physical quantity, CPU core temperature can more directly and reliably reflect the device's thermal management stress and hardware protection risks. Therefore, in this embodiment, CPU core temperature is used as the specific implementation of the processor load metric to more accurately characterize the hardware stress level faced by the device.
[0085] Specifically, we first need to define a unified index that comprehensively reflects the current hardware load pressure of the device. There are two main factors affecting the sustainable working capacity of edge terminals: remaining battery power and processor temperature. When the remaining battery power is below a safe threshold, it means the device is about to run out of power and needs to reduce its workload to extend battery life. When the processor temperature exceeds a safe threshold, it means the device's heat dissipation capacity has reached its limit, and continued high-load operation will trigger hardware protection mechanisms or even cause damage. Therefore, these two factors can be linearly weighted and combined into a single hardware resource pressure index.
[0086] Exemplarily, in this embodiment, the edge terminal exist Hardware resource pressure index at any time It can be calculated using the following formula:
[0087]
[0088] in, For edge terminals exist The remaining battery power at any given time, its value ranges from 1 to 2. , This represents the battery's full charge capacity. The preset power alarm threshold is the first threshold in step S1.2. When the remaining power is lower than this threshold, the energy-saving frequency reduction mechanism is triggered. For edge terminals exist CPU core temperature at any given time; The preset temperature alarm threshold is the equivalent temperature value corresponding to the second threshold in step S1.2. When the CPU temperature exceeds this threshold, it indicates that the device's heat dissipation capacity has reached its bottleneck. This is the maximum temperature limit that the device can withstand. Exceeding this value will trigger hardware-forced frequency reduction or shutdown protection. This is a scaling factor for electricity pressure, used to adjust the weighting of electricity factors in the overall pressure index. This is a scaling factor for temperature and pressure, used to adjust the weight of temperature factors in the overall pressure index.
[0089] It should be noted that in the above formula, The introduction of the truncation function ensures that when the power is sufficient (i.e., When the temperature is normal (i.e., the electrical pressure term is strictly zero), the electrical pressure term is strictly zero. When the temperature and pressure terms are strictly zero, the corresponding pressure term is only linearly activated when the hardware resources actually exceed the safety boundary, thus avoiding unnecessary suppression of the sampling rate when the device is in good condition. The power consumption component uses a threshold difference ratio; when the power consumption is above the threshold, this term contributes zero, and when it is below the threshold, it increases linearly as the power consumption decreases. Similarly, the temperature component contributes zero when the temperature is below the threshold, and increases linearly as the temperature increases when it exceeds the threshold.
[0090] After obtaining the hardware resource stress index, which reflects the real-time health of the equipment, the next step is to establish a mapping relationship between the hardware resource stress index and the actual sampling frequency. Considering that the sampling rate should have a highly sensitive response to pressure changes—that is, it should quickly begin to decrease when the pressure initially increases, rather than abruptly shutting down when the pressure becomes extreme—a negative exponential function can be used to establish this mapping. The negative exponential function has a natural monotonic decay characteristic and soft-landing feature: when the stress index increases from zero, the sampling rate will decay smoothly and rapidly from its maximum value; when the stress index increases further, the sampling rate gradually approaches zero but will never abruptly drop to zero, achieving a seamless degradation effect.
[0091] For example, in this embodiment, the adaptive sampling rate It can be calculated using the following formula:
[0092]
[0093] in, This represents the highest sampling frequency under ideal conditions, i.e., when the device's hardware resources are fully utilized and the pressure index is low. The sampling rate used at that time, such as 10 data records per second; This is the attenuation sensitivity coefficient, used to control the response speed of the sampling rate to changes in the pressure exponential. The larger the value, the faster the sampling rate decreases; It is a natural exponential function.
[0094] Understandably, when the equipment is in normal working condition, Approaching zero Approaching The device collects data at full speed; when the battery level drops below the threshold or the temperature rises to the alarm range, It increases linearly, while The sampling rate decays exponentially, achieving adaptive dynamic degradation of the sampling frequency. While reducing the sampling rate, the system retains only records of high-priority interaction objects, which are the top-ranking social contacts in terms of interaction frequency with the user in historical statistics. Low-priority, occasional interaction records are directly discarded, thus generating a sparse original social interaction dataset. This provides a sparse data foundation for subsequent low-power feature extraction calculations at the edge. Figure 2 This diagram illustrates a comparison of the power consumption dimension between the adaptive sampling rate and the traditional fixed sampling rate in this embodiment. Figure 2The X-axis represents the remaining battery percentage (0%~100%); the Y-axis represents the sampling frequency (times / second). The blue curve represents the negative exponential adaptive sampling rate of this scheme, which smoothly decays as the battery level drops below the threshold. The red horizontal line represents the traditional fixed sampling strategy. It can be seen that the adaptive sampling in this embodiment can actively reduce the frequency when the device's battery is low, avoiding the sharp reduction in battery life caused by always being online and sampling at high frequencies, while maintaining data quality when the battery is sufficient. Figure 3 A heatmap of the hardware resource pressure index is shown in this embodiment. Figure 3 The X-axis represents the remaining battery percentage (0%~100%); the Y-axis represents the CPU core temperature (°C); different colors represent the hardware resource stress index. This two-dimensional heatmap shows the combined trend of stress index as battery power decreases and temperature increases. The max(0,·) truncation effect of the two components of the high-pressure red area and the zero-pressure blue area in the lower right corner is clearly visible.
[0095] Step 2: Perform local privacy feature extraction on the original social interaction dataset to generate anonymized social state vectors.
[0096] Local privacy feature extraction refers to a combined process of feature extraction and privacy protection processing on the original social interaction dataset obtained in step 1, performed in the local computing environment of the edge terminal device. The goal of this process is to extract a comprehensive score that quantifies a user's social influence from the original data, and after applying a differential privacy protection mechanism to this score locally, generate a desensitized social state vector that cannot be reversed to reconstruct the original social relationship data. The entire privacy feature extraction process is completed entirely on the terminal side, without exposing the original social interaction data to any external server or third party.
[0097] The desensitized social state vector refers to the social influence feature data after differential privacy protection processing. It retains the macro trend information of the user's social influence in a statistical sense, but it can no longer be reversed to be restored to the user's real social graph structure or specific list of interaction objects. This vector is the input data for the edge server to perform global optimization calculation in the subsequent step 3.
[0098] Specifically, in this embodiment, performing local privacy feature extraction on the original social interaction dataset may include the following steps:
[0099] S2.1: Construct a three-dimensional feature matrix that includes the breadth of interaction, the depth of interaction, and the timeliness of response.
[0100] The three-dimensional feature matrix refers to a structured feature representation extracted from the original social interaction dataset, used to characterize users' social interaction behavior in multiple dimensions. The three dimensions of this matrix correspond to three complementary metrics of social interaction behavior: interaction breadth, interaction depth, and response timeliness.
[0101] Interaction breadth refers to the number of distinct social interactions a user of an edge terminal device has actually engaged with within a preset time window. This metric directly reflects the coverage and activity level of a user's social network—the more diverse the interactions with different individuals, the wider the user's social reach. Interaction breadth is obtained by deduplicating the interaction objects.
[0102] Interaction depth refers to the degree of closeness of interaction between a user of an edge terminal device and each specific social object within a preset time window. It is quantified by a weighted combination of interaction frequency and interaction duration. The higher the interaction frequency and the longer the cumulative interaction duration, the deeper the social connection with that object, and the greater the interaction depth.
[0103] Response timeliness refers to the reciprocal of the average response time difference between the user of the edge terminal device and each social object within a preset time window. The shorter the response time, the more active the relationship and the higher the level of interaction; conversely, the longer the timeliness of the relationship, the lower the current influence contribution of the relationship should be correspondingly reduced.
[0104] Understandably, the three dimensions described above depict a user's social interaction behavior from different perspectives: interaction breadth measures how many people the user knows, interaction depth measures the closeness of the connection with each person, and response timeliness measures whether the activity of the relationship is declining. These three dimensions complement each other, together forming a comprehensive and multi-dimensional portrayal of a user's social influence.
[0105] S2.2: Perform aggregation operations on the three-dimensional feature matrix to obtain the original social influence score.
[0106] It should be noted that traditional social network feature extraction schemes typically employ Graph Neural Networks (GNNs) to perform message passing and neighborhood aggregation on the node embeddings in the social graph to extract high-order topological features. However, GNNs are essentially black-box models; their feature extraction process is not interpretable, and their core operations involve large-scale matrix multiplication between adjacency matrices and feature matrices, as well as backpropagation parameter updates, resulting in a computational complexity typically of O(log n). Level (of which) Let be the number of edges in the graph. (As a feature dimension), which is almost impossible to achieve on edge terminal devices with extremely limited computing power and memory resources. In addition, the training process of GNN requires a large amount of labeled data and multiple rounds of iterative optimization, and its online inference latency is also difficult to meet the stringent real-time requirements of edge computing scenarios.
[0107] Therefore, in this embodiment, a completely white-box, physically intuition-based lightweight aggregation model, namely the marginal social gravity field model, is creatively constructed to replace the traditional graph neural network for efficient calculation of social influence scores. This model maps user relationships in social networks to a gravitational model framework in physical space, decomposing social influence into three independently measurable physical dimensions, each corresponding one-to-one with the three dimensions of the three-dimensional feature matrix constructed in step S2.1. Compared to traditional graph neural network schemes, the marginal social gravity field model has the following significant advantages: First, the completely white-box calculation process ensures that each intermediate variable has a clear physical meaning, and the model's output is interpretable and auditable; second, the entire calculation process involves only scalar-level addition, multiplication, and simple exponential operations, with computational complexity strictly controlled within a certain range. The linear level completely avoids the high-order matrix operations and backpropagation process required by graph neural networks; third, no offline training data or parameter learning process is required, and model parameters (such as weight coefficients) can be directly set based on domain prior knowledge, perfectly adapting to the extremely limited computing power, storage and energy consumption conditions in edge computing scenarios. Figure 4 This diagram illustrates a comparison of the computational complexity of the edge social gravity field model and the graph neural network in this embodiment. Figure 4 The X-axis represents the number of a node's neighbors; the Y-axis represents the relative number of computational operations (on a logarithmic scale). Figure 4 The blue bars / curves in the diagram represent the complexity of this scheme, while the red bars / curves represent the complexity of GNN. It can be seen that as the node size increases, the difference between the two increases by orders of magnitude. In other words, the edge social gravity field model in this embodiment reduces the computational overhead by several orders of magnitude compared to GNN, and has a strong advantage in engineering feasibility under the condition of limited computing power at the edge terminal.
[0108] Specifically, in the marginal social gravitational field model, the above three dimensions are analogously mapped to the three core physical quantities in the physical gravity model:
[0109] First, we define interaction breadth as the equivalent mass of a node. In the model, the greater the mass of an object, the stronger its influence as a source of attraction. Analogously, the more different objects a user actually interacts with within a specific time window, the greater their social quality and the stronger their fundamental influence as a source of influence in the social network. This metric is obtained by deduplicating the interaction objects.
[0110] For example, in this embodiment, the node exist Equivalent quality at any moment It can be calculated using the following formula:
[0111]
[0112] in, It is the set of all nodes in a social network; For nodes With nodes At any moment The frequency of interactions within the statistical time window; The cardinality of the set is the number of elements in the set. This expression represents the relationship between the node and the current time window. The number of unique nodes with at least one interaction record. The more interaction objects there are, the greater the equivalent quality, and the stronger the node's fundamental strength as a source of social attraction.
[0113] Secondly, we define interaction depth as the equivalent potential field distance between nodes. In the gravitational model, the closer two objects are, the stronger the gravitational force. By analogy, the higher the interaction frequency and the longer the cumulative interaction time between two users, the closer their social distance can be considered. Therefore, the reciprocal of the weighted combination of interaction frequency and interaction time can be defined as the equivalent potential field distance—the closer the interaction, the smaller the distance, and the stronger the gravitational potential field.
[0114] For example, in this embodiment, the node With nodes Between Equivalent potential field distance at time t It can be calculated using the following formula:
[0115]
[0116] in, For nodes With nodes Interaction frequency within the statistics window; For nodes With nodes The cumulative interaction time within the statistics window; This is a weighting coefficient for interaction frequency, used to adjust the contribution ratio of interaction frequency in distance calculation; This is a weighting coefficient for interaction duration, used to adjust the contribution ratio of interaction duration in distance calculation; For a very small positive number (e.g.) This is used to prevent division by zero errors when both the interaction frequency and interaction duration are zero.
[0117] It should be noted that in the denominator The larger the value, the closer the interaction between the two users. The smaller the value, the closer the physical distance, thus resulting in a larger gravitational value in subsequent calculations. This reciprocal relationship ensures a positive correlation between the interaction depth and the gravitational potential field strength.
[0118] Then, we define response timeliness as an exponential time decay factor. The activity level of social relationships depends not only on the frequency and duration of interactions but also heavily on the timeliness of those interactions. A relationship that has historically involved frequent interactions but has recently experienced prolonged periods of inactivity should have its current social influence significantly diminished. Therefore, an exponential decay function can be used to model this forgetting effect—the longer the average message response time, the smaller the decay factor, indicating a lower timeliness of the relationship.
[0119] For example, in this embodiment, the node With nodes Between Exponential time decay factor at time It can be calculated using the following formula:
[0120]
[0121] in, It is a natural exponential function; This is the forgetting rate coefficient. The larger the value, the more sensitive the model is to response latency, and the faster the relationship activity decays. For nodes With nodes The average response time difference between messages is measured in units consistent with the time granularity of the preset time window.
[0122] Understandably, when both parties reply to messages very quickly (i.e., When it approaches zero, A value approaching 1 indicates that the relationship is fully active, and its influence contribution is not diminished in any way; when the average response time difference is large... An exponential decay rate approaching zero indicates that the relationship has become inactive, and its contribution to current social influence should be significantly reduced. Compared to the linear decay function, the exponential decay function has more reasonable physical properties, with the fastest decay rate at the beginning of the relationship's inactivity, reflecting the recency effect of social memory.
[0123] Finally, by integrating the above three dimensions, an original social gravitational potential energy is constructed as a comprehensive influence score for nodes. Analogous to the law of universal gravitation, where gravity is directly proportional to mass and inversely proportional to the square of the distance, in the social gravitational field model, nodes... The combined gravitational potential energy of a node over all its neighbors is equal to the sum of its equivalent mass and the potential field strength between each neighbor (i.e., the reciprocal of the distance), modulated by a time decay factor. The larger this combined potential energy value, the higher the node's overall influence in the social network.
[0124] For example, in this embodiment, the node exist The original social gravitational potential energy at any given moment, i.e., the original social influence score. It can be calculated using the following formula:
[0125]
[0126] in, For nodes The set of neighboring nodes, that is, the set of nodes with which the node is located. All nodes that have at least one interaction record within the current time window; For nodes The equivalent mass is obtained from the aforementioned interaction breadth calculation; For nodes With nodes The equivalent potential distance between them is calculated from the aforementioned interaction depth; This is the corresponding exponential time decay factor, calculated from the aforementioned response timeliness.
[0127] It should be noted that, In fact equals Therefore, the entire formula only involves scalar addition, multiplication, and simple exponentiation, and the computational complexity is strictly controlled within a certain range. The linear level. Compared to graph neural networks. With a matrix operation complexity of several orders of magnitude, this edge social gravity field model reduces computational overhead by several orders of magnitude while maintaining the ability to effectively measure social influence, perfectly adapting to the extremely limited computing power and stringent power consumption constraints of terminal devices in edge computing scenarios.
[0128] S2.3: Uses the Laplace mechanism to generate random noise data based on privacy budget parameters.
[0129] The original social influence score of each node was obtained through the edge social gravity field model. Subsequently, this score contains the user's true social topology characteristics. If this score is directly uploaded in plaintext to an edge server for subsequent resource allocation optimization, there is a serious risk of privacy exposure—attackers could potentially deduce the user's social graph structure, interaction patterns, and even social circles from the influence score. Therefore, before uploading, a differential privacy protection mechanism needs to be applied locally to the influence score, ensuring that the uploaded data cannot be statistically reversed to reconstruct the user's true social relationships.
[0130] Among them, the Laplace mechanism is the key to achieving... - A classic method of differential privacy is to obscure the original information by adding random noise following a Laplace distribution to the real data. The probability density function is: ,in The scale parameter is related to the noise's scale parameter and the data's global sensitivity. Privacy budget parameters Jointly determined, specifically, scale parameters .
[0131] Global sensitivity In the social gravitational potential energy calculation formula, this refers to the maximum change in the output result caused by a change in the data of a single user. This sensitivity reflects the maximum impact of data changes on the calculation results and is a key parameter for determining the intensity of noise injection.
[0132] Privacy Budget Parameters It is a core parameter in differential privacy theory. The smaller the value, the stronger the privacy protection, because more noise needs to be injected to more fully mask the original information; The larger the value, the weaker the privacy protection. At this point, the data availability is higher, but the risk of privacy exposure also increases accordingly.
[0133] In this embodiment, privacy budget parameters It is dynamically adjustable, and its adjustment mechanism includes adaptive adjustment based on the sensitivity of the social context environment in which the edge terminal device is located.
[0134] Among them, social context sensitivity refers to a comprehensive quantitative assessment of the privacy sensitivity of the physical and network environments in which edge terminal devices are currently located. This assessment value reflects the level of privacy threats faced by user social data in the current environment.
[0135] Specifically, the core idea of this adaptive adjustment mechanism is that the strength of privacy protection should not be fixed, but should be dynamically adjusted according to the sensitivity of the user's current physical environment. When the system detects that the terminal device is in a public sensitive area (such as hospitals, government agencies, schools, or other places involving sensitive personal information) or connected to an untrusted Wi-Fi network (such as unencrypted public Wi-Fi hotspots, commercial networks with unknown credit ratings, etc.), the environmental sensitivity assessment value increases, and the system automatically reduces the privacy budget parameter. This increases the amplitude of injected noise, thereby enhancing privacy protection; when the device is in a secure environment (such as a user's home network environment, a corporate intranet, or other trusted environment), the environmental sensitivity assessment value decreases. By approximating the baseline value, more data authenticity is preserved, thereby improving the accuracy of subsequent resource allocation decisions.
[0136] In this embodiment, the environmental sensitivity assessment function can comprehensively assess the sensitivity level of the current environment based on contextual information such as the device's GPS location information (determining whether the user is within a pre-marked sensitive geofence area) and the encryption protocol status of the currently connected Wi-Fi network (determining whether the network transmission channel uses security encryption standards such as WPA3), and normalize it to... Interval.
[0137] For example, in this embodiment, dynamic privacy budget It can be calculated using the following formula:
[0138]
[0139] in, The baseline privacy budget represents the maximum privacy budget value allowed in a completely secure environment; it is a constant parameter pre-configured by the system. This is a sensitivity adjustment factor used to control the magnitude of the impact of environmental sensitivity on the privacy budget. The larger the value, the more significant the effect of changes in environmental sensitivity on the privacy budget; For edge terminals exist Environmental sensitivity assessment value at any time This indicates a completely secure environment (such as an encrypted Wi-Fi network in a user's home). This indicates a highly sensitive environment (such as connecting to unencrypted public Wi-Fi within a healthcare facility).
[0140] It is understandable that when the device is in a highly sensitive environment (such as...) )hour, Automatically reduce to This means that stronger privacy protection noise will be injected subsequently to address the higher level of privacy threats in this environment; when the device is in a secure environment (such as... )hour, Approaching This preserves more data authenticity, enabling subsequent global optimization decisions to make more accurate resource allocations based on higher-quality input data.
[0141] S2.4: Superimpose random noise data onto the original social influence score to generate an anonymized social state vector.
[0142] Once the dynamic privacy budget is determined, Laplace noise can be applied locally to the original social gravitational potential score to generate anonymized social state values.
[0143] For example, in this embodiment, desensitized social status values are used. It can be calculated using the following formula:
[0144]
[0145] in, The nodes calculated in step S2.2 The original social gravitational potential energy score; To obtain the scale parameter is Random noise samples drawn from the Laplace distribution; Global sensitivity to the formula for calculating social gravitational potential energy; This is the dynamic privacy budget calculated in step S2.3.
[0146] It should be noted that the variance of Laplace noise is Therefore when When environmental sensitivity decreases due to increased sensitivity, the noise variance increases significantly, making it mathematically impossible to reconstruct the user's true social graph from the anonymized data uploaded to the edge server. Conversely, when Maintaining a large value in a secure environment results in less injected noise and a smaller deviation between the anonymized data and the original score, which is beneficial to the accuracy of subsequent global optimization. This mechanism ensures that all privacy protection operations are completed entirely on the local terminal, and the original social interaction data and original influence score never leave the terminal device, without exposing the original data to any third party, fundamentally achieving the privacy protection goal of data being usable but invisible. Figure 5 This diagram illustrates the decomposition of the contribution of social features to the influence score in this embodiment. Figure 5The data shows the dynamic rise and fall of each component at different social stages. For example, when the relationship becomes silent, the decay factor drops sharply, causing the total score to decline. This illustrates the necessity of complementary cooperation of three-dimensional features (breadth × depth × timeliness), as a single dimension cannot fully depict the dynamic evolution of social influence. Figure 6 The dynamic privacy budget in this embodiment is illustrated. A schematic diagram of the adaptive change curve of environmental sensitivity. Figure 7 This diagram illustrates a comparison of the desensitization score distribution of the differential privacy protection effect under different privacy budgets in this embodiment. Figure 7 The X-axis represents the social influence score; the Y-axis represents the probability density. Figure 7 As can be seen from this, this embodiment provides sufficient information obfuscation strength in sensitive scenarios, satisfying the mathematical guarantee of differential privacy.
[0147] After the processing described in S2.1 to S2.4, each edge terminal device de-identifies its social status value. The data is uploaded to the edge server and used as input data for the global attention allocation calculation in subsequent step 300.
[0148] Step 3: The edge server performs a global attention allocation calculation under fairness constraints based on the de-identified social state vectors received from multiple edge terminal devices, and generates a set of attention allocation weights.
[0149] An edge server is a server device deployed at the edge of a network that possesses strong computing and network aggregation capabilities. It can receive anonymized data uploaded from multiple edge terminal devices within its coverage area and perform global optimization calculations. For example, an edge server can be a multi-access edge computing (MEC) server on the base station side or a computing node integrated in a roadside unit (RSU).
[0150] Global attention allocation calculation refers to the process by which the edge server, after receiving the de-identified social state vectors of all terminal devices, calculates the corresponding attention allocation weight for each terminal device by solving a joint optimization model that includes a system overall utility objective and a fairness penalty term. The goal of this process is to maximize the overall system utility while ensuring the fairness of resource allocation and avoiding the Matthew effect of bandwidth monopoly by top users in traditional greedy algorithms.
[0151] Attention allocation weight set refers to the resource allocation weight vector corresponding to each edge terminal device in the network, obtained by the edge server through solving the joint optimization model. Each element in the vector This represents the allocation to the node. Attention resource quota, The larger the value, the more network resources (such as bandwidth, cache space, computing time slots, etc.) the node will obtain in subsequent resource scheduling.
[0152] In this embodiment, the method is further applicable to the high dynamic scenarios of vehicle-to-everything (V2X) networks, and step 300 also includes predictive processing of network topology stability.
[0153] Specifically, in typical mobile edge computing scenarios such as connected vehicles, terminal nodes (i.e., vehicles in motion) may leave the base station's coverage area and disconnect in the next instant. If the resource allocation algorithm fails to anticipate this topology change and still allocates a large amount of bandwidth or cache resources to high-influence nodes that are about to disconnect from the network, it will lead to serious resource waste and system performance fluctuations—these allocated but unusable resources should have been used to serve other nodes still in the network. Therefore, a topology stability prediction and penalty mechanism needs to be introduced to preemptively reduce the weight of nodes that are about to disconnect from the network, making resource allocation decisions more forward-looking.
[0154] S3.a: Using the Kalman filter algorithm, predict the network topology change trend at the next moment based on the de-identified social state vectors at historical moments.
[0155] The Kalman filter algorithm is an optimal recursive estimation algorithm based on a linear state-space model. It recursively estimates the current state of the system and predicts its future state by fusing prior predictions and the latest observations from the system model, using the minimum mean square error criterion. In this embodiment, the Kalman filter algorithm is used to predict the probability of each node going offline at the next moment based on the historical connection state sequence of nodes (such as the time series trend of signal strength, historical observations of movement speed and direction, etc.).
[0156] It is understandable that, in this embodiment, a first-order Hidden Markov Model (HMM) can be used to replace the Kalman filter for predicting the probability of network disconnection. The HMM models the state transition probabilities between node connection states (online / about to disconnect / already disconnected), and can also infer the probability distribution of future node connection states based on historical observation sequences.
[0157] S3.b: Introduce a topological stability factor when calculating attention allocation weights.
[0158] The topology stability factor is a quantitative measure of the reliability of edge terminal devices in maintaining network connectivity in the near future. This factor is defined as the complement of the predicted probability of disconnection, so that nodes with more stable connections receive a stability weight closer to 1, while nodes about to disconnect have a stability weight approaching zero.
[0159] For example, in this embodiment, the node exist Topological stability factor at time It can be calculated using the following formula:
[0160]
[0161] in, The nodes predicted by the Kalman filter (or Hidden Markov Model) exist The probability of disconnection at any given moment.
[0162] Understandably, when node connections are stable and the predicted probability of disconnection is extremely low (i.e., )hour, This indicates that the node's resource allocation weight is not affected by any penalty; when the node's signal weakens rapidly and it is predicted to be disconnected from the network (i.e., )hour, In subsequent global optimizations, this node will be significantly downgraded, thus avoiding the allocation of valuable network resources to nodes that will soon be unable to use these resources. This proactive downgrading mechanism effectively solves the problem of resource allocation lag and system jitter caused by sudden node disconnection, which is particularly critical for scenarios with rapidly changing network topologies, such as connected vehicles.
[0163] It should be noted that for non-V2X static or low-mobility scenarios (such as social interactions between smartphone users in a fixed location within a home Wi-Fi environment), network topology changes extremely slowly, and the probability of all nodes dropping out can be approximated as zero. For all nodes, the topology stability factor degenerates to a constant of 1, having no differentiating effect on resource allocation results. Therefore, the introduction of the topology stability factor does not affect the applicability of the method in static scenarios, while significantly enhancing its robustness in highly dynamic scenarios.
[0164] After calculating the topological stability factor, the next step is to construct a joint optimization model under fairness constraints.
[0165] In this embodiment, the specific steps for constructing a joint optimization model that includes a system total utility objective and a fairness penalty term include:
[0166] S3.1: Define the objective function for total system utility .
[0167] Wherein, the system's total utility objective function It refers to a scalar-valued function, which is the sum of the expected benefits of all edge terminal devices in the network after obtaining attention resources, and is used to comprehensively evaluate the utility level of the current resource allocation scheme at the global level.
[0168] Specifically, this function needs to comprehensively consider the anonymized social influence score of each node. Topological stability factor and the resource weight assigned to that node The three factors are integrated into a single, optimizable global scalar index. The relationship between resource allocation weights and utility is modeled using a logarithmic function—this is because the logarithmic function naturally exhibits diminishing marginal utility: when a node has already acquired a significant amount of resources, the utility increment from each additional unit of resource gradually decreases. This property fundamentally encourages the optimizer to distribute resources across more nodes, rather than concentrating them on a few, thus suppressing the tendency for resource monopolization.
[0169] For example, in this embodiment, the system's total utility objective function It can be calculated using the following formula:
[0170]
[0171] in, This represents the total number of edge terminal nodes in the network. Assign a weight vector to the resources to be solved. To be assigned to a node Resource weight; For nodes The topological stability factor (default is 1 in non-V2X scenarios); For nodes De-identified social influence score after differential privacy protection; It is a logarithmic utility function, reflecting diminishing marginal utility.
[0172] It should be noted that the product term This constitutes the overall priority coefficient for each node—reflecting both the node's influence within the social network (as determined by...) The measurement also takes into account the reliability of its network connection (by...). (Measurement) to ensure resources are not wasted on high-influence nodes that are about to be disconnected. Logarithmic function The introduction of this ensures that the rate of utility growth decreases as resources increase, which mathematically tends to produce a more uniform resource allocation result.
[0173] S3.2: Define fairness penalty items .
[0174] Among them, fairness penalty item This is a penalty function used in the joint optimization model to explicitly constrain the fairness of resource allocation. The penalty term is constructed based on the Jain fairness index and is used to quantify the dispersion of resource proportions for each terminal device in the current allocation scheme.
[0175] Relying solely on the diminishing marginal utility function is insufficient to completely resolve the fairness issue. When the combined priority coefficients of different nodes... When there are significant disparities, the optimizer may still allocate the vast majority of resources to a few high-priority nodes, leaving a large number of long-tail users (i.e., users with lower influence scores) unserved, severely damaging the long-term health of the social ecosystem and user retention rates. Therefore, an explicit fairness penalty needs to be introduced to mathematically constrain and enforce the fairness of the allocation results.
[0176] The Jain fairness index is a classic fairness metric widely used in computer networking and resource allocation. Its value range is... The Jain exponent reaches its maximum value of 1 when all nodes receive exactly equal resources, representing absolute fairness; the Jain exponent reaches its minimum value when resources are completely monopolized by a single node. This indicates extreme unfairness.
[0177] For example, in this embodiment, the Jain fairness index can be calculated using the following formula:
[0178]
[0179] In this embodiment, the fairness penalty item The calculation formula is configured as a non-linear penalty function, which applies when the Jain fairness index falls below a preset fairness threshold. At that time, the value of the penalty term grows exponentially to force the optimization algorithm to allocate resources to long-tail users.
[0180] Among them, the preset fairness threshold These are parameters pre-set by the system administrator based on the fairness requirements of the application scenario, reflecting the system's minimum tolerance for fairness in allocation. For example, in a social content recommendation scenario, It can be set to 0.8, requiring the Jain index to be no less than 0.8; in emergency communication scenarios, It can be set to a higher value, such as 0.95, to ensure that resources are distributed very evenly.
[0181] For example, in this embodiment, a non-linear fairness penalty term is used. It can be calculated using the following formula:
[0182]
[0183] in, Jain fairness index for the current resource allocation scheme; This is a preset fairness threshold; The penalty sensitivity coefficient, The larger the Jain index, the more severe the penalty increase when the Jain index is below the threshold.
[0184] It should be noted that when At that time, the current allocation plan meets the fairness requirement. Strictly zero has no effect on the objective function, and the optimizer is free to pursue utility maximization; when hour, Negative value It is a positive value and increases as the fairness gap increases. The function causes the penalty value to grow exponentially, resulting in a huge optimization cost. This, mathematically, generates a very strong gradient signal, forcing the optimizer to distribute resources from the head nodes to the long-tail nodes to restore the fairness level above the threshold. This threshold-gated + exponentially exponential penalty mechanism design avoids unnecessary interference with utility optimization when fairness is satisfied, while providing a powerful corrective force when fairness is critical.
[0185] S3.3: Construct the Lagrange objective function.
[0186] After defining the overall utility objective function and the fairness penalty term, the two are integrated into a unified joint optimization objective.
[0187] For example, in this embodiment, the Lagrange objective function can be expressed as:
[0188]
[0189] in, The objective function is the total utility of the system. This is a non-linear fairness penalty term; This is a hyperparameter for dynamically adjusting fairness weights, used to balance the importance of system utility and fairness. The larger the value, the more the system values fairness, and the more the optimizer will tend to minimize the resource gap between nodes when allocating resources; The smaller the value, the more the system focuses on maximizing utility, allowing for a certain degree of resource allocation.
[0190] Understandably, due to When fairness is not satisfied, a positive value is taken. Therefore, a minus sign is used to connect the terms in the objective function, so that the objective function value decreases when the penalty term increases, thereby guiding the optimizer to search in the direction of reducing the penalty term (i.e. improving fairness).
[0191] S3.4: Set constraints.
[0192] The constraints are designed to ensure that the sum of all assigned weights does not exceed the total available bandwidth or computing resource threshold of the edge server.
[0193] For example, in this embodiment, the constraint can be expressed as:
[0194]
[0195] in, This represents the upper limit of the total resources that can be allocated to the system, corresponding to the total bandwidth or total number of computing time slots currently available to the edge servers. The first constraint ensures that resource allocation does not exceed the system capacity; the second constraint ensures that the allocation weight of each node is non-negative, i.e., there is no unreasonable situation of negative resource allocation.
[0196] In this embodiment, the alternating direction multiplier method (ADMM) is used to solve the joint optimization model in step S3.
[0197] Among them, the alternating direction multiplier method is a solution method suitable for distributed optimization problems. Its core idea is to decompose a globally coupled optimization problem into multiple local subproblems that can be solved in parallel on each terminal and a global coordination problem that is executed on the edge server by introducing auxiliary variables and equality constraints. The two converge to the global optimum through alternating updates of dual variables.
[0198] Solving the aforementioned global optimization problem with nonlinear penalties in a centralized manner presents two serious technical obstacles: First, it requires aggregating complete data from all terminals to the server, which contradicts the privacy-preserving architecture design of this solution at the terminal side; second, when the number of terminal nodes in the network... When the problem is large, centralized solutions incur significant communication overhead, which can easily cause uplink congestion. Therefore, this embodiment uses ADMM to achieve decentralized distributed solutions, which offloads the computational load to the terminal and keeps the communication load to a minimum, perfectly matching the end-edge collaborative distributed computing architecture.
[0199] Specifically, this involves introducing auxiliary global variables. And add equality constraints (For all) This transforms the original optimization problem into an equivalent augmented Lagrangian form. In this form, local variables... The optimization only involves nodes Its own utility information can be obtained entirely on the local terminal; global variables The optimization involves fairness penalties and total resource constraints, which are centrally processed on the edge server side.
[0200] For example, in this embodiment, the augmented Lagrange function can be expressed as follows:
[0201]
[0202]
[0203] in, Assign weight vectors to the local resources of each terminal; The global consensus variable vector maintained for edge servers, after being adjusted for fairness constraints, represents a globally coordinated allocation scheme; For the dual variable vector, The dual variable vector is the transpose of the dual variable vector, also known as the Lagrange multiplier. It reflects the cost of the deviation between the local optimum and the global consensus, and in economics, it can be understood as the shadow price of resource allocation. The weighting coefficients for the fairness penalty term correspond to the hyperparameters in the Lagrange objective function in step S3.3; To augment the step size penalty parameter of the Lagrange multiplier, used to control... and The strength of the consistency constraints between them The larger the value, the heavier the penalty for local solutions deviating from the global consensus; It is the square of the L2 norm.
[0204] It should be noted that in the augmented Lagrange function, the first term is the negative value of the utility function (because the original problem is to maximize utility, and the negative sign is taken after transforming it into a minimization problem), the second term is the fairness penalty, and the third term... For standard Lagrange dual terms, the fourth term This is an augmentation term (also known as a quadratic penalty term), which enhances the convexity and numerical stability of the optimization problem and accelerates the convergence of ADMM.
[0205] After constructing the augmented Lagrangian function, ADMM iteratively solves the problem by alternately updating three sets of variables.
[0206] S3.5: Decompose the global optimization problem into multiple sub-problems, each sub-problem corresponding to the local variable update of an edge terminal device.
[0207] In each iteration In, each terminal Solve its own local optimization subproblems independently and in parallel, and update the local resource allocation weights. This subproblem only involves nodes. Its own utility function and the global consensus value received from the server and dual variables It does not require information from any other nodes, so it can complete the calculation independently on the terminal side.
[0208] For example, in this embodiment, the edge terminal local update rule can be as follows:
[0209]
[0210] Among them, superscript Indicates the first Round iteration; This indicates taking the option that minimizes the objective function. Value; constraint Ensure that resource allocation weights are non-negative.
[0211] Understandably, this optimization subproblem is a single-variable problem. The convex optimization problem has an objective function consisting of the negative value (negative logarithmic term) of a concave function and the sum of two quadratic / linear convex functions, making the whole function convex. Therefore, this subproblem has a unique global optimum and can be efficiently solved by simple one-dimensional search, the golden section method, or a closed-form solution based on KKT conditions. The computational cost is extremely low, and it can be executed on edge devices with limited computing power without imposing a significant additional computational burden on the terminal.
[0212] S3.6: The edge server broadcasts global aggregation information based on the dual variable of the current iteration.
[0213] After collecting all terminal updates Then, the edge server performs a global variable update. The edge server operates within the global resource constraints. Next, by combining the fairness penalty term, a unified solution is obtained for the global consensus variable. .
[0214] For example, in this embodiment, the global update rule for the edge server can be as follows:
[0215]
[0216] in, This represents the maximum amount of resources that can be allocated by the system.
[0217] It should be noted that in this global update step, if the Jain index is affected by the independent optimization results of each terminal... Below the fairness threshold ,but This term generates a very large penalty value, forcing the optimizer to solve the problem more efficiently. Resources are redistributed from the head nodes to the long-tail nodes, thus mathematically guaranteeing fairness in resource allocation and effectively breaking the Matthew effect caused by greedy algorithms. After updating global variables, the edge servers will share the updated global consensus value. and dual variables The information is broadcast back to each edge terminal device as reference information for the next round of partial updates.
[0218] S3.7: Based on the global aggregated information and the desensitized social state vector, iteratively update the weights allocated to each edge terminal device until the objective function converges or the preset maximum number of iterations is reached.
[0219] After updating local and global variables, perform a dual variable update to reflect the current state. and The residuals between them.
[0220] For example, in this embodiment, the dual variable update rule can be expressed as follows:
[0221]
[0222] in, The residual between the local solution and the global consensus in the current iteration is multiplied by the step size parameter. This information is then accumulated in the dual variable. The physical meaning of the dual variable is that it records the cumulative deviation between the local solution and the global consensus in each iteration. By feeding this deviation information back to the local and global subproblems in subsequent iterations, it guides the two to gradually narrow the gap and eventually reach a consensus.
[0223] The three steps S3.5 to S3.7 described above constitute one complete iteration of ADMM. The system alternately repeats these three steps until one of the following convergence conditions is met: the original residual and dual residual All are less than the preset convergence tolerance Or the number of iterations reaches the preset maximum number of iterations. .
[0224] It should be noted that throughout the entire ADMM iteration process, the edge server only broadcasts the scalar-level global consensus value to each terminal. and dual variables (total Each terminal only uploads the updated weights at the scalar level to the server. (total (A scalar). The entire communication process does not involve the transmission of any privacy state matrix, raw social data, or social graph structure; the communication volume of each iteration is only a scalar. This allows the computational load to be shifted to the terminal while keeping the communication load to a minimum, perfectly matching the distributed computing architecture of end-edge collaboration and fundamentally solving the communication congestion and privacy leakage problems faced by centralized computing solutions. Figure 8 This diagram illustrates the convergence process of the Jain fairness index and the greedy algorithm for equal allocation in this embodiment. Figure 9This diagram illustrates the characteristic curve of the nonlinear fairness penalty term in this embodiment. Figure 10 The diagram shows the Pareto trade-off between total system utility and fairness. Figure 10 As can be seen, by adjusting the weight coefficients of the fairness penalty term, the system administrator can flexibly compromise between maximizing utility and ensuring fairness. In this embodiment, the Pareto front consistently outperforms both the greedy algorithm and the equal allocation approach. Through the iterative execution of the above three steps, the system eventually converges to an optimal set of attention allocation weights that maximizes global utility while ensuring fairness in resource allocation. .
[0225] Step 4: Distribute the attention allocation weights to each edge terminal device and perform corresponding resource scheduling operations according to the weights.
[0226] Here, attention allocation weight distribution refers to the process by which the edge server distributes the final set of attention allocation weights obtained after ADMM optimization converges. The weight value corresponding to each edge terminal device. The process of sending data to the corresponding terminal devices via the downlink control channel.
[0227] Resource scheduling refers to the process by which edge terminal devices and edge servers dynamically adjust network resource or content caching strategies based on the weights allocated by the received attention. This scheduling operation transforms abstract mathematical optimization results into concrete network resource configuration behaviors and is the final step in the implementation of the entire method.
[0228] In this embodiment, performing the corresponding resource scheduling operation according to the weight specifically includes the following steps:
[0229] S4.1: Adjust the length of the content recommendation cache queue for edge terminal devices in social networks based on attention allocation weights.
[0230] The content recommendation cache queue refers to a first-in, first-out (FIFO) queue maintained on edge terminal devices or their associated edge cache nodes to temporarily store social content (such as posts, updates, videos, etc.) to be pushed to the user. The longer the queue, the more content is pre-cached for the user, resulting in a smoother and lower-latency experience when browsing social content.
[0231] Specifically, the system assigns weights based on attention. The size of the cache queue is adjusted proportionally to the size of the cache queue for each terminal device. Devices with higher weights receive longer cache queues, meaning the system pre-caches more social content for them to ensure a smooth browsing experience and rich content. Devices with lower weights receive shorter cache queues, and their cached resources are reduced accordingly, but they still retain basic content recommendation service capabilities.
[0232] In this embodiment, the length of the cache queue By assigning weights to attention After normalization, multiply by the total available cache space of the system. To determine, that is ,in This is a floor function that ensures the allocated cache space is in integer units.
[0233] S4.2: For devices with a weight higher than the high priority threshold, allocate a dedicated low-latency transmission time slot.
[0234] Among them, the high-priority threshold refers to a pre-set weight value threshold, which determines the weight of attention allocation for a certain edge terminal device. When the threshold is exceeded, the device is marked as a high-priority device and will be granted dedicated network transmission resources.
[0235] A dedicated transmission time slot refers to a transmission time period reserved for a specific device in a multiple access framework such as Time Division Multiple Access (TDMA) or Orthogonal Frequency Division Multiple Access (OFDMA), which is not shared with other devices. Devices allocated a dedicated transmission time slot can exclusively use the transmission channel during that time period, thereby achieving extremely low transmission latency and extremely high transmission reliability, and avoiding resource competition with other devices.
[0236] It is understandable that the design intention of allocating dedicated transmission time slots to high-weight devices is that high-weight devices usually correspond to core user nodes with high influence and stable network connections in social networks. The content distribution efficiency and service experience of these nodes have an important leverage effect on the activity and health of the entire social ecosystem. Therefore, providing them with differentiated low-latency dedicated channels has significant system-level value.
[0237] S4.3: For devices with a weight lower than the low priority threshold, schedule their data transmission requests to be executed during non-congestion periods.
[0238] Here, the low-optimal threshold refers to another pre-set weight value threshold, which determines the weight of attention allocation for a certain edge terminal device. When the device is below this threshold, it is marked as a low-priority device.
[0239] Non-congestion periods refer to times when network traffic is low and transmission channel resources are relatively abundant, such as the early morning hours or off-peak hours on weekdays. Scheduling data transmission requests from low-priority devices to non-congestion periods means that non-real-time operations such as content updates and data synchronization on these devices will be delayed until the network is less congested. This frees up more transmission resources for high-priority devices during peak hours, achieving dynamic allocation of network resources over time.
[0240] Understandably, low-priority devices are not completely deprived of service—the system still guarantees their basic social interaction functions (such as sending and receiving messages), only scheduling their non-urgent large-scale data transmission tasks (such as loading social dynamic streams and pre-caching video content) to non-congestion periods. This differentiated scheduling strategy effectively improves the overall network resource utilization efficiency and service quality for high-priority users during peak hours without significantly affecting the basic experience of low-priority users.
[0241] It should be noted that the three sub-steps S4.1 to S4.3 mentioned above are not mutually exclusive, but rather a multi-level resource scheduling strategy that can be executed in combination. Adjusting the cache queue length affects the content storage dimension, allocating dedicated transmission time slots affects the communication transmission dimension, and scheduling non-congestion periods affects the time scheduling dimension—these three aspects respectively provide fine-grained configuration of network resources from the dimensions of storage, transmission, and time, together constituting a complete resource scheduling execution scheme.
[0242] Example 2
[0243] This embodiment discloses a privacy-preserving social attention fairness allocation system based on edge computing.
[0244] Specifically, the privacy-preserving social attention fairness allocation system based on edge computing can be deployed in an edge computing system consisting of multiple edge terminal devices and one edge server. The edge terminal devices can be smartphones, tablets, on-board units (OBUs), wearable devices, smart Bluetooth devices, or other IoT terminals with local computing and network communication capabilities. The edge server can be a multi-access edge computing (MEC) server deployed on the base station side, an edge computing node integrated in a roadside unit (RSU), or an edge server cluster composed of multiple physical servers.
[0245] In this embodiment, the edge computing-based privacy-preserving social attention fairness allocation system can also be deployed in multiple layers of electronic devices. For example, data awareness and privacy protection functions are deployed in each edge terminal device, and the global fairness decision-making function is deployed in the edge server. The edge terminal devices and the edge server work together to implement the edge computing-based privacy-preserving social attention fairness allocation method of this application.
[0246] In this embodiment, when the edge computing system is applied to the vehicle-to-everything (V2X) scenario, the edge terminal device can be an on-board terminal unit installed on a moving vehicle, and the edge server can be a roadside unit (RSU) deployed on the roadside or an MEC server on the base station side. The system is further suitable for highly mobile scenarios with highly dynamic network topologies.
[0247] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A privacy-preserving, fair allocation method for social attention based on edge computing, characterized in that, The method for fair allocation of social attention includes: Based on the local perception of edge terminal devices, the original social interaction dataset is obtained within a preset time window. Local privacy feature extraction is performed on the original social interaction dataset to generate anonymized social state vectors. The local privacy feature extraction includes: Based on the original social interaction dataset, a three-dimensional feature matrix is constructed that includes the breadth of interaction, the depth of interaction, and the timeliness of response. The original social influence score is obtained by performing aggregation operations on the three-dimensional feature matrix. Using the Laplace mechanism, random noise data is generated based on the privacy budget parameter and the global sensitivity of the original social influence score. Random noise data is superimposed onto the original social influence score to generate the desensitized social state vector; The edge server constructs a joint optimization model based on the de-identified social state vectors received from multiple edge terminal devices, which includes a system total utility objective function and a fairness penalty term. Solve the joint optimization model to generate the attention allocation weight set. The overall system utility objective function is constructed based on the de-identified social state vectors and logarithmic utility functions of each edge terminal device. The fairness penalty item is constructed based on the Jain Fairness Index; and The attention allocation weights corresponding to each edge terminal device in the attention allocation weight set are distributed to each edge terminal device, and the corresponding resource scheduling operations are performed according to the attention allocation weights.
2. The privacy-preserving social attention fairness allocation method based on edge computing according to claim 1, characterized in that, Before acquiring the original social interaction dataset within the preset time window, the method further includes: an adaptive sampling step based on device hardware status. Real-time monitoring of the hardware resource status parameters of the edge terminal device, including the remaining power percentage and processor load index; Calculate the hardware resource stress index based on the remaining battery percentage and the processor load index; and Based on the hardware resource pressure index, an adaptive sampling frequency is determined through a negative exponential mapping relationship, and the original social interaction dataset is collected according to the adaptive sampling frequency.
3. The privacy-preserving social attention fairness allocation method based on edge computing according to claim 2, characterized in that, When the hardware resource pressure index exceeds a preset pressure threshold, during the data acquisition process at an adaptive sampling frequency... Only the interaction records of social objects whose interaction frequency ranks among the top preset numbers with the user of the edge terminal device are retained, and the remaining interaction records are discarded to generate the sparse original social interaction dataset.
4. The privacy-preserving social attention fairness allocation method based on edge computing according to claim 1, characterized in that, The aggregation operation on the three-dimensional feature matrix to obtain the original social influence score includes: The three-dimensional feature matrix is aggregated using a marginal social gravitational field model, which maps the three dimensions of the three-dimensional feature matrix to three physical quantities in a physical gravity model: The interaction breadth is mapped to equivalent quality, which is the number of unique social objects that have at least one interaction record with the user of the edge terminal device within the preset time window. The interaction depth is mapped to the equivalent potential field distance, which is the reciprocal of the weighted combination of interaction frequency and interaction duration. The response timeliness is mapped to an exponential time decay factor, which is a negative exponential decay function value with the average message response time difference as the independent variable; and For each neighboring node of the user's node in the edge terminal device, calculate the product of the equivalent mass and the inverse of the equivalent potential distance, multiply it by the corresponding exponential time decay factor, and sum the calculation results of all neighboring nodes to obtain the original social influence score.
5. The privacy-preserving social attention fairness allocation method based on edge computing according to claim 1, characterized in that, The privacy budget parameter is a dynamic privacy budget, which is adaptively adjusted based on the environmental sensitivity assessment value of the current environment in which the edge terminal device is located.
6. The privacy-preserving social attention fairness allocation method based on edge computing according to claim 5, characterized in that, The environmental sensitivity assessment value is determined based on the location information of the edge terminal device and the encryption protocol status of the currently connected network; wherein, When the edge terminal device is located within a pre-marked sensitive geofence area, or when the currently connected network does not use an encryption protocol with a preset security level, the environmental sensitivity assessment value increases, and the dynamic privacy budget decreases accordingly.
7. The privacy-preserving social attention fairness allocation method based on edge computing according to claim 1, characterized in that, Before constructing the joint optimization model, the edge server also includes: Using the Kalman filter algorithm, based on the historical connection state sequence of each edge terminal device, the probability of each edge terminal device dropping the connection at the next moment is predicted; and A topology stability factor is calculated based on the disconnection probability, where the topology stability factor is the complement of the disconnection probability; wherein, The system's overall utility objective function also includes the topology stability factor, which ensures that edge terminal devices with higher disconnection probabilities receive lower attention allocation weights in the attention allocation weight set.
8. The privacy-preserving social attention fairness allocation method based on edge computing according to claim 1, characterized in that, The solution to the joint optimization model employs the alternating direction multiplier method, including: By introducing global consensus variables and dual variables, the joint optimization model is transformed into an augmented Lagrangian function; In each iteration, the following alternating updates are performed: Each edge terminal device solves its own local optimization subproblem in parallel and updates its own local allocation weight. Each local optimization subproblem only involves the de-identified social state vector of the corresponding edge terminal device itself, as well as the global consensus variable and the dual variable received from the edge server. Under the constraint of the total resource limit, the edge server updates the global consensus variable based on the updated local allocation weights of each edge terminal device, in conjunction with the fairness penalty term; and The dual variable is updated based on the residual between the updated local allocation weights and the updated global consensus variable; Repeat the alternating update until the preset convergence condition is met, and use the converged global consensus variable or the local allocation weight as the attention allocation weight set.
9. The privacy-preserving social attention fairness allocation method based on edge computing according to claim 1, characterized in that, Execute corresponding resource scheduling operations based on the attention allocation weights, including at least one of the following: Based on the attention allocation weights, the length of the content recommendation cache queue corresponding to each edge terminal device is adjusted proportionally. For edge terminal devices whose attention allocation weight is higher than the high priority threshold, a dedicated transmission time slot is allocated; and For edge terminal devices whose attention allocation weight is lower than the low priority threshold, non-urgent data transmission requests of the edge terminal device are scheduled to be executed during non-congestion periods.
10. The privacy-preserving social attention fairness allocation method based on edge computing according to claim 1, characterized in that, The attention fairness allocation method is applied to the Internet of Vehicles scenario, and the edge computing device is an in-vehicle terminal unit.
