An incentive mechanism for personalized privacy protection crowd sensing based on stackelberg game
By employing a personalized privacy protection incentive mechanism influenced by Stackelberg game theory and social relationships, the balance between data accuracy and privacy protection in the crowd-sensing system is resolved, thereby improving data quality and worker participation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- JINAN UNIVERSITY
- Filing Date
- 2024-01-10
- Publication Date
- 2026-07-24
AI Technical Summary
In crowd-sensing systems, existing technologies struggle to balance data accuracy with personalized privacy protection, lack effective incentive mechanisms to attract workers to participate in tasks, and also suffer from privacy leaks and data noise complexity.
A personalized privacy protection incentive mechanism based on Stackelberg game theory is adopted. Through a two-stage game between the platform and workers, combined with the influence of social relationships, a worker privacy budget and data processing are set. The normal distribution noise fitting method is used to denoise the data, so as to achieve a trade-off between data accuracy and privacy protection.
It achieves improved data accuracy and worker participation under personalized privacy protection, meets workers' privacy needs, improves data quality, and motivates more workers to participate in perception tasks.
Smart Images
Figure CN117688605B_ABST
Abstract
Description
Technical Field
[0001] This invention mainly relates to crowd intelligence sensing, and particularly to the field of crowd intelligence sensing incentive mechanisms. Background Technology
[0002] Crowdsensing (CS) technology has been widely applied in the context of the rapid development of built-in sensors, becoming a powerful tool for solving complex and scalable sensing problems. However, in practical MCS systems, sensor data provided by individual workers is often unreliable. Recruiting a sufficient number of workers is essential to obtain accurate sensor data. However, participating in sensing tasks incurs costs for workers (such as computing power, electricity, and bandwidth). Therefore, a reasonable incentive mechanism is needed to attract workers to participate in crowdsensing tasks. Furthermore, since the data submitted by workers is related to their personal information, the uploaded data may lead to a certain degree of privacy leakage, and each worker's perception of privacy differs. Moreover, for the CS platform to obtain sufficiently accurate data, the accuracy of the data submitted by workers must be guaranteed. Therefore, crowdsensing systems need to provide an incentive mechanism that balances accuracy and privacy protection while meeting individual privacy requirements to attract workers to participate in crowdsensing tasks.
[0003] The Stackelberg Game (SG) is a two-stage dynamic game with complete information. Its core idea is that both players are influenced by each other's decisions. In this model, the decision-makers are divided into a leader and followers. The leader makes the first decision, and the followers make their choices based on the leader's decisions. The leader then adjusts its decisions according to the followers' choices, iterating until a Nash equilibrium is reached. However, in personalized privacy scenarios, workers influence each other's decisions but are unaware of each other's choices, making it an incomplete information game. This necessitates using Bayesian game theory with incomplete information in the second stage of the game.
[0004] In crowdsourced sensing scenarios, the most important aspect is information aggregation and ensuring the accuracy of uploaded information. However, due to the need to protect worker privacy, and the varying privacy requirements of workers necessitating personalized privacy protection, the noise in the data collected by the platform becomes highly complex. This requires the use of data recovery methods for noise reduction to extract accurate information. The privacy protection method considered in this invention is to add normally distributed noise to satisfy ε-KL-Local Differential Privacy (ε-KL-LDP). The noise follows a weighted combination of multiple normal distributions, and MOG is used to fit the noise to achieve a denoising effect. Summary of the Invention
[0005] This invention discloses an incentive mechanism for personalized privacy-preserving crowd perception based on Stackelberg game theory. It mainly balances the interests of the platform and users, while satisfying the privacy needs of workers, thereby attracting more workers to participate in crowd perception tasks and ensuring the accuracy of task results.
[0006] According to the application background of the present invention, a personalized privacy-preserving crowd intelligence perception incentive mechanism based on Stackelberg design is provided, comprising the following steps:
[0007] Step 1: The task requester publishes a request for a sensing task to the platform. The task set is defined as follows: The unit value of each task is
[0008] Step 2: The crowd-sensing platform recruits workers to participate in the task. The worker set is defined as follows:
[0009] Step 3: Workers upload their privacy requirements to the platform for task allocation. These requirements represent their maximum acceptable privacy budget, and this set is defined as follows:
[0010] Step 4: The crowd sensing platform uses the Stackelberg game theory method to derive the privacy budget allocated to each worker. and the set of tasks performed by the user Where π j It is the task number, C i It represents the total number of tasks executed by user i, and for each task, the set of participating users. μ j This represents the worker's serial number, with a total of Wt. k A worker participates in the k-th task. In addition, the unit value of the privacy budget is set as Pe, and the platform distributes the corresponding privacy budget and the set of tasks to be performed to the worker.
[0011] 1) Define the worker utility function i∈N. Two friends in a social network often influence each other and exhibit similar behaviors; therefore, this invention uses a symmetric neighbor matrix. However, because the platform is untrustworthy, it assigns ε to each user. d The privacy budget is used to obtain random response perturbations in the neighbor matrix. Regarding privacy costs, this invention uses a widely adopted quadratic cost quantification function, s,a,b>0,s i Worker privacy budget ε i The equivalent monetary value.
[0012] 2) Define the platform's utility function The adjustable parameter for the equivalent monetary value of η is represented by a quadratic function, where c and d are positive parameters used to represent the concavity and convexity of the function, which are used to capture the property of diminishing marginal returns.
[0013] 3) In the two-stage Stackelberg game, the Stackelberg Nash equilibrium (SE) obtained under the constraints of platform budget and worker personalized privacy requirements is R and E, respectively.
[0014] Stage I (Leader Game):
[0015] Stage II (Follower Game):
[0016]
[0017] ε i ≤m i ,
[0018] Step 5: Workers collect sensor data according to their own privacy budget. Add noise to the sensed data locally and then upload the data.
[0019] Step 6: The crowd sensing platform uses MoG to fit the noise of the received data, and then performs noise reduction processing to obtain the sensing data. The system will determine whether workers have uploaded malicious data or the data quality is low based on the received data, and adjust workers' wages accordingly.
[0020] Compared with existing technologies, the advantages of this method are:
[0021] 1. This invention applies Stackelberg game theory to a crowd intelligence perception incentive mechanism, which can achieve a balance between data accuracy and workers' personalized privacy needs.
[0022] 2. This invention takes into account the impact of social relationships in the choice of privacy budget, which is more in line with reality;
[0023] 3. This invention considers multi-task scenarios to better suit existing crowd sensing platforms;
[0024] 4. This invention uses MoG to fit noise to crowd sensing data, which can obtain more accurate sensing data than the truth discovery method. Attached Figure Description
[0025] Figure 1 This is a flowchart of the present invention;
[0026] Figure 2 This invention is a personalized crowd intelligence sensing platform; Detailed Implementation
[0027] like Figure 1 As shown, the specific steps of the technical solution of the present invention are as follows:
[0028] Step 1: The requester sends a sensing task request to the crowdsourcing sensing platform.
[0029] Step 2: The crowd-sensing platform will issue task requests and recruit workers.
[0030] Step 3: Workers upload their maximum privacy budget.
[0031] Step 4: The crowd sensing platform uses the Stackelberg game theory method to derive the privacy budget allocated to each worker. and the set of tasks performed by the user Where π j It is the task number, C i It represents the total number of tasks executed by user i, and for each task, the set of participating users. μ j This represents the worker's serial number, with a total of Wt. k A worker participates in the k-th task. In addition, the unit value of the privacy budget is set as Pe, and the platform distributes the corresponding privacy budget and the set of tasks to be performed to the worker.
[0032] 1) Define the worker utility function i∈N. Two friends in a social network often influence each other and exhibit similar behaviors; therefore, this invention uses a symmetric neighbor matrix. However, because the platform is untrustworthy, it assigns ε to each user. d The privacy budget is used to obtain random response perturbations in the neighbor matrix. Regarding privacy costs, this invention uses a widely adopted quadratic cost quantification function, s,a,b>0,s i Worker privacy budget ε i The equivalent monetary value.
[0033]
[0034] Regarding privacy costs, this invention uses a widely adopted quadratic cost quantification function, s. i ,a i ,b i >0,s i Worker privacy budget ε i The equivalent monetary value.
[0035] 2) Define the platform's utility function The adjustable parameter for the equivalent monetary value of η is represented by a quadratic function, where c and d are positive parameters used to represent the concavity and convexity of the function, which are used to capture the property of diminishing marginal returns.
[0036] 3) In the two-stage Stackelberg game, the Stackelberg Nash equilibrium (SE) obtained under the constraints of platform budget and worker personalized privacy requirements is R and E, respectively.
[0037] Stage I (Leader Game):
[0038] Stage II (Follower Game):
[0039]
[0040] ε i ≤m i ,
[0041] 4) Considering the social network structure in crowdsourcing perception, workers' privacy budgets are influenced by other workers. Existing research demonstrates that similar characteristics among workers, such as education level, can serve as influencing factors in the social network. Therefore, the social network structure needs to be considered in the design of the incentive mechanism. We assume that the influence between workers is bidirectional, and the resulting neighbor matrix V is symmetric, i.e., v ij =v jiFurthermore, in crowd-aware perception, each user can only know their own degree. The degree set in the network is G = {1, ..., f}. max}, f max Let f∈G be the maximum degree in the network, and let P be the probability distribution of G, and ∑ f∈G P(f) = 1. Since each user only knows their own degree and not the degrees of other users, and is unaware of the network structure, this invention uses a "configuration model," which is a random graph model with a given degree sequence. This means the exact degree of each individual vertex in the network is fixed, and the social network structure is constructed accordingly. Because users submit their maximum privacy requirements, as long as the budget requirement is not exceeded, the condition is met. We use degree as the budget selection parameter, setting the maximum privacy budget for the same degree as the minimum privacy budget requirement at that degree to satisfy the condition. Therefore, the privacy budget is transformed into ε(f), and the mean utility for user i can be rewritten as...
[0042]
[0043] Where γ(f) is the social network effect coefficient of a worker of degree f, defined as:
[0044]
[0045] It is the average privacy budget on the network, defined as:
[0046]
[0047] The platform's average utility function is:
[0048]
[0049] This invention assumes that the platform has a limited budget, so the optimization objective is to ensure that the total reward does not exceed the platform's total budget B. Therefore, the platform optimization objective is written as:
[0050]
[0051] The optimal result E is obtained from the KKT conditions. * ;
[0052] 5) Then the R obtained * To optimize user utility using Bayesian equilibrium (BNE), the goal of this invention is to select the optimal privacy level that maximizes user utility. First, the partial derivative of the rewritten user utility is calculated:
[0053] To maximize user utility get Therefore, the optimal privacy budget for workers can be obtained as follows:
[0054]
[0055] Step 5: Workers collect sensor data according to their own privacy budget. i Add noise to the sensed data locally and then upload the data.
[0056] Step 6: The crowd-sensing platform uses MoG algorithm to denoise the received data to obtain the sensing data. The system will determine whether workers have uploaded malicious data or the data quality is low based on the received data, and adjust workers' wages accordingly.
[0057] To verify the effectiveness of this invention, this method was tested using a Matlab simulation platform. The number of tasks M was set to 200, the number of workers N to 100, and the total budget B to 1000. It was assumed that the probability distribution of malicious uploads by workers or quality errors caused by sensors was E(0, δ). e ), where δ e It follows a gamma distribution. Assume the generated real data follows a uniform distribution [10, 30]. Add N(0, δ) n The noise level satisfies ε-KL-LDP. The MoG-RPCA method is used for denoising the uploaded noisy data. In summary, under the constraints of personalized privacy and a fixed budget, a trade-off can be achieved between the platform's maximum utility and user utility. This provides users with a reasonable privacy budget for noise addition, thus ensuring data quality. Furthermore, providing reasonable rewards incentivizes more workers on the platform to participate in perception tasks.
Claims
1. An incentive method for personalized privacy-preserving crowd-based intelligent perception based on Stackelberg game theory, characterized in that: Step 1: The task requester publishes a request for a sensing task to the platform. The task set is defined as follows: ,in Indicates the first There are 10 tasks, and the unit value of each task is 100. ,in Indicates the first The unit value of each task; Step 2: The crowd-sensing platform recruits workers to participate in the task. The worker set is defined as follows: ,in Indicates the first One worker; Step 3: Workers upload their privacy requirements to the platform for task allocation. These requirements represent their maximum acceptable privacy budget, and this set is defined as follows: ,in Indicates workers The maximum privacy budget accepted; Step 4: The crowd sensing platform uses the Stackelberg game theory method to derive the privacy budget allocated to each worker. and the set of tasks performed by the user in It is allocated to the workers Privacy budget, It is the task number. User The total number of tasks executed, and for each task, the set of participating users. , This indicates the worker's serial number; there are a total of [number missing]. A worker participates in the k-th task. In addition, the unit value of the privacy budget is set as Pe, and the platform distributes the corresponding privacy budget and the set of tasks to be performed to the worker. Step 5: Workers collect sensor data, add noise to the sensor data locally according to their own privacy budget, and then upload the data. , Indicates workers For the task Uploaded data; Step 6: The crowd sensing platform uses MoG to fit the noise of the received data, and then performs noise reduction processing to obtain the sensing data. The system will determine whether workers have uploaded malicious data or the data quality is low based on the received data, and adjust workers' wages accordingly. ,in User Actual salary; The steps to derive the optimal unit value and privacy budget are as follows: 1) Define the worker utility function In social networks, two friends often influence each other and then exhibit similar behaviors, thus a symmetric neighbor matrix is used. However, because the platform is untrustworthy, it assigns each user... The privacy budget is used to obtain random response perturbations in the neighbor matrix. The widely used quadratic cost quantification function is used to quantify privacy costs, where... The actual should be Workers' privacy budget The equivalent monetary value, since subsequent calculations are all based on the worker. The calculations are simplified to... ; 2) Define the platform's utility function. , The adjustable parameter for the equivalent monetary value is represented by a quadratic function, where... , It is a positive parameter used to represent the concavity or convexity of a function, and is used to capture the property of diminishing marginal returns; 3) In the two-stage Stackelberg game, the Stackelberg Nash equilibrium is achieved under the constraints of platform budget and worker personalized privacy requirements. and ; Stage Ⅰ: Stage Ⅱ: 4) Considering the social network structure in crowdsourcing perception, workers' privacy budgets are influenced by other workers. Existing papers have demonstrated that similar characteristics among workers can serve as influencing factors in the social network. Therefore, the social network structure needs to be considered in the design of the incentive mechanism. This assumes that the influence between workers is bidirectional, and the resulting neighbor matrix V is symmetric. Furthermore, in crowd-aware perception, each user can only know their own degree, which is the set of degrees in the network. It is the largest degree in the network, making ,and yes The probability distribution, Since each user only knows their own degree and not the degrees of other users, and is unaware of the network structure, a "configuration model" is used. This is a random graph model with a given degree sequence, where the exact degree of each individual vertex in the network is fixed. This model is used to construct the social network structure. Furthermore, since users submit their maximum privacy requirements, as long as these requirements are not exceeded, the condition is met. Degree is used as the budget selection parameter, setting the maximum privacy budget for the same degree to the minimum privacy budget requirement at that degree to satisfy the condition. Thus, the privacy budget is transformed into... For users The mean utility can be rewritten as : in, It is a degree The social network effect coefficient of workers is defined as: It is the average privacy budget on the network, defined as: The platform's average utility function is: Given the platform's limited budget, the optimization goal is to ensure that total rewards do not exceed the platform's total budget. Therefore, the optimization goal of this platform is written as: The optimal result is obtained from the KKT conditions. ; 5) Then, from the obtained To optimize user utility using Bayesian equilibrium, the goal is to select the optimal privacy level that maximizes user utility. First, the partial derivative of the rewritten user utility is calculated: To maximize user utility, let the partial derivative... get Therefore, the optimal privacy budget for workers can be obtained as follows: 。