A method for identifying and controlling abnormal participants in mobile crowd sensing under local differential privacy constraint

Abstract

The application provides a mobile crowd sensing abnormal participant identification and control method under local differential privacy constraints. Participants report original sensing data disturbed by a local differential privacy mechanism to a server, including the following steps: obtaining sensing data disturbed by a local differential privacy mechanism submitted by multiple participants for multiple sensing tasks; processing to obtain initial aggregation results of each sensing task and calculating the residual of each participant on the corresponding sensing task; extracting multi-dimensional statistical features to form a feature vector representing the behavior pattern of the participant; based on the feature vectors of all participants, modeling the joint probability distribution of the multi-dimensional statistical features, and constructing a joint probability distribution model; identifying abnormal participants based on the distribution deviation degree of the feature vectors of each participant; and dynamically adjusting the final sensing result aggregation strategy of the crowd sensing system to suppress the influence of abnormal participant data on the final aggregation result.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and neural network security technology, specifically relating to a method for identifying and controlling abnormal participants in mobile swarm intelligence perception under local differential privacy constraints. Background Technology

[0002] In the mobile crowdsourcing sensing technology system, relying on the ubiquitous sensing capabilities of massive mobile intelligent terminals, the limitations of deployment costs and coverage of traditional fixed sensor networks can be overcome. This has led to large-scale applications in numerous fields, including smart city traffic status sensing, urban environmental quality monitoring, public infrastructure inspection, community noise pollution collection, and public health emergency situation awareness. It has become one of the core technological supports for refined urban governance and wide-area environmental information collection. These systems aggregate terminal sensing data from a vast amount of ordinary participants to complete large-scale, wide-coverage information collection and analysis. The reliability of the final sensing results depends heavily on the authenticity and validity of the data reported by the participants.

[0003] The open participation nature of mobile crowdsourcing sensing systems also brings significant privacy and security risks. The sensing data reported by participants often includes sensitive and private information such as their geographical location, device identification, and daily habits. Directly uploading raw sensing data poses a serious risk of privacy leakage and greatly reduces the willingness of ordinary users to participate, becoming one of the bottlenecks restricting the large-scale promotion of mobile crowdsourcing sensing systems. To address this issue, existing mainstream commercial systems and academic solutions generally adopt local differential privacy mechanisms to protect sensing data. Unlike traditional centralized differential privacy, which requires privacy perturbation to be performed on the server side, local differential privacy places the privacy protection execution step on the participant's local terminal. The raw sensing data is only randomly perturbed on the terminal before being uploaded to the server. The server has no access to the participant's raw sensing data throughout the process, achieving end-to-end privacy protection from the data source. This effectively eliminates participants' concerns about privacy leakage and is naturally suitable for the multi-participant, weak-trust distributed deployment environment of mobile crowdsourcing sensing systems.

[0004] However, while local differential privacy (LDP) mechanisms offer strong privacy protection, their inherent random perturbation characteristics also provide natural cover for the behavior of anomalous participants in the system, posing significant technical challenges to traditional malicious node detection and anomalous data filtering schemes. To meet the constraints of differential privacy, LDP mechanisms inject large-variance, independently and identically distributed additive noise into the original data. In applications with high privacy requirements, the fluctuation range of the injected noise often far exceeds the inherent fluctuation range of normally perceived data, directly rendering traditional anomaly detection schemes based on absolute numerical deviations completely ineffective. The core assumption of traditional numerical threshold-based schemes is that the deviation between anomalous data and the true value is significantly greater than the natural fluctuation of normal data. However, LDP noise causes the reported data from normal participants to exhibit extremely large numerical fluctuations. Lowering the anomaly threshold will result in a large number of normal participants being misjudged, while relaxing the threshold will allow malicious poisoning behavior to be completely hidden within legitimate noise and missed. Meanwhile, traditional anomaly clustering and detection algorithms based on distance metrics such as Euclidean distance and Manhattan distance face serious distance inflation and dimensionality curse problems in multi-task mobile crowd sensing scenarios. These algorithms rely on the distance difference between samples to distinguish normal and abnormal behavior. However, the noise injected by the LDP mechanism in each sensing task dimension is independent of each other. As the number of sensing tasks participated in by participants increases, the cumulative effect of multi-dimensional noise will cause the multi-task reported data of normal participants to be extremely dispersed in high-dimensional space. The cumulative distance between normal samples may even exceed the distance difference between abnormal poisoning samples and normal samples, completely drowning out the real data offset caused by malicious behavior, and causing the detection scheme based on distance metrics to completely lose its ability to distinguish. More importantly, most existing mainstream anomaly detection schemes rely on prior assumptions about the proportion of anomalous participants and specific attack patterns, and often use fixed static judgment thresholds to achieve anomaly identification. However, mobile crowdsourcing sensing systems are in a dynamically changing environment, and the number of participants, task types, and sensing environment will continue to change. The heavy-tailed distribution characteristics introduced by LDP noise will further exacerbate the dynamic shift of the overall data distribution. Such detection schemes based on static priors not only cannot adapt to the dynamically changing system environment, but are also very easy for malicious participants to bypass by adaptively adjusting their poisoning strategies, making it impossible to achieve continuous and effective control over anomalous behavior.

[0005] In summary, existing anomaly detection schemes for mobile crowd sensing systems cannot simultaneously ensure accuracy in anomaly identification, adaptability to dynamic scenarios, and feasibility for large-scale system deployment under the strong constraints of local differential privacy. There is an urgent need to develop anomaly identification and control schemes that adapt to the inherent characteristics of local differential privacy mechanisms. These schemes should achieve effective identification and system-level control of anomaly behavior without compromising privacy protection mechanisms or relying on the acquisition of raw sensing data, while meeting the practical application requirements of large-scale mobile crowd sensing systems for low computational overhead and high scenario adaptability. Summary of the Invention

[0006] To address the shortcomings and deficiencies of existing technologies, this invention provides a method for identifying and controlling abnormal participants in a mobile swarm intelligence sensing system under local differential privacy constraints. It is mainly used to solve the core technical bottlenecks faced by traditional anomaly detection schemes in mobile swarm intelligence sensing scenarios with local differential privacy protection, such as failure due to absolute numerical deviation, distance expansion in multiple task dimensions, and poor adaptability of static thresholds. This method breaks through the natural cover-up of abnormal behavior by privacy noise. This invention first acquires perception data submitted by multiple participants for multi-sensory tasks, perturbed by a local differential privacy mechanism. It then performs a specially designed initial aggregation process on this data. The initial aggregation result is used only for constructing subsequent residual statistical features and is not directly output as the final perception result of the system. Next, using the initial aggregation results of each perception task as a benchmark, it calculates the residual of each participant on the corresponding perception task, thus depicting the relative deviation of the participant's reported data relative to the group's aggregation behavior. Subsequently, it extracts multi-dimensional statistical features from the historical residuals of each participant across multiple perception tasks, comprehensively characterizing the participant's long-term behavioral patterns from four dimensions: location, scale, asymmetry, and tail characteristics of the residual distribution. Based on the multi-dimensional statistical features of all participants, it performs joint probability distribution modeling to construct a benchmark distribution model for characterizing the behavioral patterns of normal participants. Then, based on the degree of deviation of each participant's multi-dimensional statistical features from the benchmark distribution model, it completes the identification of abnormal participants. Finally, it directly feeds the anomaly identification results back to the final perception result aggregation stage of the collective intelligent perception system. By dynamically adjusting the aggregation strategy, it suppresses the impact of abnormal participant-reported data on the final aggregation result, forming a complete closed loop of anomaly identification and robust system control. This invention is based on the inherent coupling and consistency design of the multidimensional residual statistical characteristics of normal participants under the influence of local differential privacy noise. It is difficult for abnormal participants to simultaneously forge features that conform to normal behavior patterns in multiple statistical dimensions. The entire process does not require obtaining the original perception data of participants, nor does it require pre-setting attack models or the proportion of abnormal participants. Without violating the local differential privacy protection constraints, it effectively identifies abnormal behaviors hidden in privacy noise, which greatly improves the robustness of the mobile crowd perception system and the reliability of the final perception results in the local differential privacy scenario. At the same time, the overall solution has low computational overhead, can adapt to the deployment requirements of mobile crowd perception systems with large-scale participants, and has good scenario adaptability and engineering feasibility.

[0007] The specific technical solution adopted by this invention to solve its technical problem is as follows:

[0008] A method for identifying and controlling abnormal participants in mobile swarm sensing under local differential privacy constraints. The mobile swarm sensing system includes a server and multiple participants. The participants report their raw sensing data to the server after perturbing it using a local differential privacy mechanism. The method includes the following steps:

[0009] Acquire perception data submitted by multiple participants for multiple perception tasks, after being perturbed by a local differential privacy mechanism;

[0010] An initial aggregation process is performed on the sensing data to obtain the initial aggregation results for each sensing task. The initial aggregation results are only used for residual construction and are not output as the final sensing results.

[0011] Based on the initial aggregation results of each perception task, the residual of each participant on the corresponding perception task is calculated. The residual is the difference between the participant's perception data and the initial aggregation results of the corresponding perception task.

[0012] Multidimensional statistical features are extracted from the residuals of each participant on multiple perceptual tasks to form a feature vector characterizing the participants’ behavioral patterns. The multidimensional statistical features at least characterize the location, scale, asymmetry and tail characteristics of the residual distribution.

[0013] Based on the feature vectors of all participants, the joint probability distribution model of the multidimensional statistical features is performed to construct a joint probability distribution model that characterizes the behavior patterns of normal participants.

[0014] Based on the degree of deviation of each participant's feature vector from the joint probability distribution model, abnormal participants are identified.

[0015] Based on the identification results of anomalous participants, the aggregation strategy for the final perception results of the crowd-aware perception system is dynamically adjusted to suppress the influence of anomalous participant data on the final aggregation results.

[0016] Furthermore, the residual of participant i on perceptual task j is ,in The perception data of participant i on perception task j is perturbed by a local differential privacy mechanism. The residuals are the initial aggregation results for the perception task j and are not used to recover the participants' original perception data.

[0017] Furthermore, the multidimensional statistical features include residual mean, residual root mean square, residual skewness, and residual kurtosis, which respectively characterize the location, scale, asymmetry, and tail characteristics of the residual distribution; the feature vector of participant i for:

[0018]

[0019] Where M represents the number of perceptual tasks that participant i participates in. Let the standard deviation of the residual sequence of participant i be . The residual of participant i on perception task j; the construction of the participant feature vector is based solely on the reported data after differential privacy perturbation and its relative residual statistical characteristics, without relying on the real perceived value, attack model assumptions or prior labels of the participant, and is not used to recover or infer the participant's original perceived data.

[0020] Furthermore, the joint probability distribution modeling adopts a nonparametric probability density estimation method or a semiparametric probability density estimation method.

[0021] Furthermore, based on the joint probability distribution model, the probability density of the participants' feature vectors is evaluated, and an anomaly score that is negatively correlated with the probability density value is constructed. The anomaly score is used to quantify the degree to which a participant deviates from the behavior pattern of normal participants. By analyzing the distribution characteristics of the anomaly score in the participant set, an anomaly judgment threshold is adaptively determined, and anomaly participants are identified based on the anomaly judgment threshold.

[0022] Furthermore, the initial aggregation result is obtained by performing median aggregation on the perceived data after perturbation by the local differential privacy mechanism, or by calculating a weighted average based on historical consistency.

[0023] Furthermore, the dynamic adjustment method of the aggregation strategy includes at least one of the following: reducing the weight of abnormal participants in the final aggregation process, delaying the adoption of data reported by abnormal participants, and restricting the participation of data from abnormal participants in the final aggregation.

[0024] Furthermore, when the number of perceptual tasks participated in by the participants is less than a preset threshold, the statistical features corresponding to residual asymmetry and tail characteristics are stabilized, or robust statistics are used to replace at least some of the higher-order statistical features.

[0025] Furthermore, before extracting multidimensional statistical features, the residual data is truncated or shortened, and the extracted feature vectors are standardized to eliminate scale differences between different statistical features.

[0026] And, a mobile crowd sensing system for implementing the method of any one of claims 1-9, comprising:

[0027] The data acquisition module is used to acquire perception data submitted by multiple participants for multiple perception tasks, after being perturbed by a local differential privacy mechanism.

[0028] The initial aggregation module is used to perform initial aggregation processing on the perceived data and output the initial aggregation result for residual construction only;

[0029] The residual construction module is used to calculate the residuals of each participant on the corresponding perceptual task based on the initial aggregation results of each perceptual task.

[0030] The feature extraction module is used to extract multidimensional statistical features from the residuals of each participant on multiple perceptual tasks to form a feature vector that characterizes the behavior pattern of the participants. The multidimensional statistical features at least characterize the location, scale, asymmetry and tail characteristics of the residual distribution.

[0031] The distribution modeling module is used to perform joint probability distribution modeling on the multidimensional statistical features based on the feature vectors of all participants, and to construct a joint probability distribution model that characterizes the behavioral patterns of normal participants.

[0032] The anomaly detection module is used to identify anomalous participants based on the degree of deviation of each participant's feature vector from the distribution of the joint probability distribution model.

[0033] The aggregation control module is used to dynamically adjust the aggregation strategy of the final perception result of the crowd intelligence perception system based on the identification results of abnormal participants, so as to suppress the influence of abnormal participant data on the final aggregation result.

[0034] And a computer device including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the method described above.

[0035] A non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described above.

[0036] Compared to existing technologies, this invention and its preferred solution effectively overcome the technical bottlenecks faced by traditional anomaly detection schemes in local differential privacy protection scenarios. It eliminates reliance on prior assumptions about absolute numerical deviations, spatial distance metrics, attack patterns, or anomaly proportions. It can effectively identify anomalous participants even when privacy noise provides natural cover for abnormal behavior, significantly reducing the risk of misjudgment and missed detection that is common in traditional solutions in this scenario. The entire process is based on reported data perturbed by local differential privacy, without requiring the acquisition of participants' original perception data or the recovery or inference of the original perception data. It completes the identification and control of anomalous behavior without violating local differential privacy protection constraints or affecting participant privacy and security, effectively resolving the core contradiction between privacy protection and data reliability assurance in crowdsourced sensing systems. It possesses extremely high... With strong scene adaptability and versatility, it does not require customized design for specific attack modes and has good adaptability to dynamically changing crowd sensing environments and participant sets of different sizes. Its optimal solution can also be adapted and optimized for scenarios with low participation participants and fluctuating data distribution, effectively resisting adaptive attack strategies and avoiding the problem of easy failure of traditional static threshold solutions. The overall solution has simple computational logic, without the need for complex multi-round iterative optimization or adversarial game solving, resulting in low computational overhead and easy deployment in large-scale mobile crowd sensing systems, with excellent engineering feasibility. At the same time, it constructs a complete closed loop of anomaly recognition and system aggregation control, which can directly suppress the impact of abnormal data on the final sensing results by dynamically adjusting the aggregation strategy, thereby improving the operational robustness of the mobile crowd sensing system and the reliability of the final output sensing results at the system level. Attached Figure Description

[0037] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0038] Figure 1 This is a flowchart illustrating the implementation of an embodiment of the present invention. Detailed Implementation

[0039] To make the features and advantages of the present invention more apparent and understandable, specific embodiments are described below in detail:

[0040] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used in this specification have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0041] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0042] Under the constraint of local differential privacy noise, although anomalous participants can forge behavioral characteristics similar to normal participants in a single statistical dimension, it is difficult for them to maintain statistical consistency across multiple residual statistical feature dimensions simultaneously. This statistical consistency constraint is characterized by non-simultaneous forgery under the masking condition of local differential privacy noise. The aforementioned statistical consistency constraint does not depend on the specific form of statistical quantities, but rather originates from the inherent coupling relationship between multidimensional residual statistical features under the influence of local differential privacy noise.

[0043] The aforementioned statistical consistency constraints and inherent coupling relationships are specifically manifested as follows: the reported data of normal participants are only superimposed with independent and identically distributed random noise injected by the local differential privacy mechanism. Their multi-task residual sequences conform to the inherent laws of independent and identically distributed noise in terms of statistical characteristics, that is, the residual mean approaches 0, the fluctuation amplitude matches the privacy budget, the distribution shape approaches a symmetrical distribution, and the tail characteristics are consistent with the noise distribution. The statistical characteristics of the four dimensions have a fixed inherent coupling relationship. However, in order to carry out poisoning attacks, abnormal participants need to introduce a systematic bias into the reported data. This bias will cause the residual distribution to show synchronous anomalies in at least two statistical dimensions, making it impossible to fit the coupling relationship of the four-dimensional statistical characteristics of normal participants at the same time, and thus be identified in the joint distribution modeling.

[0044] Based on the above-mentioned technical understanding, this invention achieves effective identification of abnormal participants by jointly modeling the multidimensional residual statistical characteristics of participants, and uses the identification results for the aggregated control of the crowd perception system.

[0045] The purpose of this invention is to propose a method for identifying and controlling anomalous participants in mobile crowd sensing under local differential privacy constraints. This method can effectively identify and control anomalous participants hidden in privacy noise under local differential privacy protection, thereby improving the robustness of the crowd sensing system and the reliability of the sensing results. The anomaly determination in this invention is not based on a fixed threshold or a single statistic, but rather on analyzing the overall distribution of anomalous scores in the participant set to characterize the relative anomalousness of each participant within the current group, thus achieving anomalous participant identification in multi-participant crowd sensing scenarios.

[0046] To achieve the above objectives, this invention provides a method for identifying and controlling anomalous participants based on joint modeling of multidimensional residual statistical features. Applied to a locally differential privacy-preserving crowd-aware sensing system, the method aims to improve both the accuracy of anomalous participant identification and control and the system's computational efficiency. It processes the participant set and its sensing data, including a residual statistical feature construction stage, a joint distribution modeling stage based on multidimensional features, and an anomalous participant identification and system-level control stage based on density assessment. In this embodiment, kernel density estimation is used as one of the joint distribution modeling methods; however, other nonparametric or semiparametric modeling methods can also be used in practical applications.

[0047] The residual statistical feature construction stage first acquires differential privacy perturbation reporting data from multiple participants across multiple perception tasks, and then calculates the initial aggregation results for each perception task based on the reported data. Unlike traditional anomaly identification and control methods that directly construct biases based on original observations or true values, this invention introduces a residual construction mechanism based on the initial aggregation results, addressing the noise amplification characteristics in local differential privacy scenarios. The initial aggregation results are not the final perception results, but rather serve as a reference benchmark for characterizing the relative behavioral shifts of participants under differential privacy noise interference, thereby preventing further amplification of privacy noise during the residual construction process. Multidimensional statistical features are extracted from the residual data to characterize the participants' data distribution characteristics, thus depicting their behavioral features throughout the overall perception process.

[0048] In the kernel density modeling stage based on multidimensional features, the multidimensional statistical features corresponding to each participant are combined to form a feature vector. A kernel density estimation model is then constructed based on the feature vectors of all participants to characterize the probability density distribution of normal participants in the multidimensional feature space, thus forming a behavioral baseline model for normal participants. The multidimensional statistical features used are not intended to characterize the anomalies of single perceived values, but rather to model the distribution expansion phenomenon caused by local differential privacy noise, based on the overall distribution characteristics of participants in the multi-task residual sequence. Specifically, positional and scale-related statistical features are used to characterize the central tendency and dispersion after the introduction of privacy noise, while asymmetric and tail-related statistical features are used to capture the systematic shifts in the distribution pattern of abnormal participants. These shifts are difficult to distinguish in a single task but are cumulative in multi-task statistics. Without departing from the technical concept of this invention, the joint distribution modeling method can be equivalently replaced according to system requirements.

[0049] In the density-based abnormal participant identification and system-level control stage, anomaly assessment is performed on participants based on the density value of their corresponding feature vectors under the kernel density estimation model or the anomaly score obtained from it. Participants with anomaly scores higher than a preset threshold are detected as abnormal participants and removed during subsequent perception result aggregation to obtain robust perception results. This forms an aggregation control closed loop based on anomaly identification results, enabling continuous regulation of the operational status of the crowd-based intelligent perception system.

[0050] Preferably, in the residual statistical feature construction stage, the multidimensional statistical features include at least statistics used to describe the location, scale, asymmetry and tail characteristics of the residual distribution, so as to comprehensively reflect the participants' perceptual behavior characteristics from multiple dimensions.

[0051] Preferably, in the kernel density modeling stage based on multidimensional features, a nonparametric probability density estimation method is used to model the participant features, without the need to pre-set the proportion of abnormal participants or the attack model, thereby improving the adaptability of the method to different abnormal behavior patterns.

[0052] Preferably, in the abnormal participation identification and control stage based on density assessment, the abnormal judgment threshold is adaptively determined by analyzing the abnormal score distribution of participants, so as to reduce the risk of misjudging normal participants.

[0053] Compared with existing technologies, this invention introduces joint modeling of multidimensional residual statistical features to form statistical consistency constraints on abnormal behavior under the condition of local differential privacy noise masking. This makes it difficult for abnormal participants to evade identification by forging through a single statistical dimension, thereby achieving reliable identification of abnormal participants under the condition of privacy protection.

[0054] This invention does not require pre-setting the proportion of abnormal participants, attack strategies, or adversarial models. It identifies abnormal participants solely based on the joint distribution characteristics of participants in the statistical features of multi-task residuals, and has good versatility and adaptability.

[0055] This invention directly feeds back the results of abnormal participant identification to the aggregation process of the crowd intelligence perception system, and achieves system-level robust control through weight adjustment or participation restrictions, thereby improving the reliability of the overall perception results without violating local differential privacy constraints.

[0056] This invention does not require multiple rounds of iterative optimization or adversarial game solving, has low computational complexity, is easy to deploy in large-scale mobile swarm intelligence sensing systems, and has good engineering feasibility.

[0057] The implementation of the present invention will be described more completely below with reference to the accompanying drawings and through more specific embodiments. These embodiments are used to illustrate the technical concept and implementation process of the present invention, but do not constitute a limitation on the scope of protection of the present invention.

[0058] This invention addresses the problem of difficulty in identifying abnormal behavior in mobile swarm intelligence sensing systems under local differential privacy protection conditions by proposing a method for identifying abnormal participants and controlling the system at the system level using multidimensional residual statistical consistency constraints.

[0059] This method constructs historical residuals of participants relative to the initial aggregation results based on differential privacy perturbation reporting data of participants on multiple perception tasks, and extracts multidimensional statistical features from the residuals to characterize the distribution characteristics of the residuals. By jointly modeling the distribution of the multidimensional statistical features, the relative anomaly of the participants' behavior is evaluated, thereby identifying abnormal participants. The identification results are then fed back to the aggregation process of the crowd-sensing system to achieve effective control over abnormal participants.

[0060] This invention does not require pre-setting the proportion of abnormal participants or attack models. It can effectively improve the robustness and reliability of the sensing results of the crowd intelligence sensing system under the condition of local differential privacy noise masking. It has the advantages of low computational overhead and applicability to large-scale crowd intelligence sensing systems.

[0061] The implementation framework provided by this invention includes a server and multiple participants. Each participant reports the raw perceived data after perturbing it using a local differential privacy mechanism, such as... Figure 1 As shown, the implementation process follows these steps:

[0062] Step S1: Obtain the collective intelligence perception data submitted by multiple participants through mobile terminals, wherein the collective intelligence perception data is processed locally by a privacy perturbation mechanism and then uploaded to the server;

[0063] Step S2: Perform initial aggregation processing based on the swarm intelligence sensing data. The initial aggregation processing is used to construct residual statistical features rather than as the final sensing result output.

[0064] Step S3: Based on the initial aggregation results, construct the historical residuals of each participant on multiple perceptual tasks;

[0065] Step S4: Extract multidimensional statistical features from historical residuals to characterize participant behavior patterns;

[0066] Step S5: Perform joint probability distribution modeling on the multidimensional statistical features to construct a distribution model for characterizing the behavioral patterns of normal participants;

[0067] Step S6: Identify anomalous participants based on the degree of deviation of the multidimensional statistical characteristics of each participant from the distribution model;

[0068] Step S7: Based on the identification results of abnormal participants, dynamically adjust the aggregation strategy of the crowd perception system, including reducing the weight of abnormal participants in the aggregation process, delaying the adoption of their reported data, or restricting their data participation in the aggregation, so as to reduce the impact of abnormal participants on the final aggregation result.

[0069] The historical residuals are constructed as follows: Let the participants be... In perception tasks Differential privacy perturbation reported data is The initial aggregation result for the corresponding task is Then the participants In the mission The residual on is defined as:

[0070]

[0071] The residual is used to characterize the degree of deviation of a participant from the group's aggregation behavior, rather than to recover the true perceived value.

[0072] The participant feature vector is composed of residual statistical features, expressed as:

[0073]

[0074] in, Indicates participants In perception tasks The residuals on The number of perceptual tasks participated in by the participants. For participants The residual sequence standard deviation is used; each component of the feature vector is used to characterize the overall offset, overall fluctuation, distribution asymmetry, and tail anomaly of the participant's residual distribution. The participant feature vector serves as a representation of participant behavior and is used for subsequent abnormal participant identification and aggregation control. The construction of the participant feature vector is based solely on the reported data after differential privacy perturbation and its relative residual statistical characteristics, without relying on the actual perceived values, attack model assumptions, or prior participant labels, and is not used to recover or infer the participant's original perceived data.

[0075] In a preferred embodiment, the joint probability distribution modeling employs one of the nonparametric probability density estimation methods based on kernel functions, or other nonparametric or semiparametric modeling methods used to characterize the joint distribution of multidimensional statistical features.

[0076] Anomaly scores are constructed based on the evaluation results of participant feature vectors using a distribution model, and are used to quantify the degree to which participants deviate from normal behavioral patterns.

[0077] Multidimensional statistical features are not limited to specific statistical forms; they can be used as long as they can reflect the overall shift, fluctuation, asymmetry, or tail characteristics of the residual distribution.

[0078] By analyzing the distribution characteristics of abnormal scores across the participant set, an anomaly detection threshold is adaptively determined to adapt to crowdsourcing scenarios with varying participant sizes and anomaly rates. The execution order of steps S1 to S7 can be adjusted according to the system implementation without affecting the technical effectiveness of anomaly participant identification and control.

[0079] When the number of perceptual tasks participated in by participants is less than a preset threshold, the residual skewness and residual kurtosis statistical features are stabilized, or robust statistics are used to replace at least some higher-order statistical features to prevent low-participation participants from circumventing anomaly detection. The aforementioned robust statistics refer to statistics that are insensitive to outliers and small sample fluctuations, maintaining the stability of statistical estimates in low-sample-size scenarios. These include, but are not limited to, the median, median absolute deviation, quartile skewness, and quartile kurtosis. Stabilization refers to addressing the problem of excessive variance in higher-order statistical estimates in low-sample-size scenarios by introducing regularization constraints and using biased but smaller variance estimators to improve statistical stability. This includes, but is not limited to, centering the residual sequence, adding a shrinkage regularization term to the skewness / kurtosis estimates, and using Bayesian smoothing estimation. The robust statistics mentioned above refer to statistics that are insensitive to outliers and small sample fluctuations, maintaining the stability of statistical estimates in low-sample-size scenarios. These include, but are not limited to, the median, median absolute deviation, quartile skewness, and quartile kurtosis.

[0080] As a further preferred implementation, when the number of perception tasks M participated in by the participants is less than a preset threshold (e.g., M<10), the following stabilization methods can be used: the median is used to replace the mean of the residuals to calculate the standardized residuals, and the median absolute deviation is used to replace the standard deviation of the residual sequence to reduce the interference of outliers on the statistics under small sample sizes; or the interquartile skewness and interquartile kurtosis are used to replace the conventional third-order and fourth-order moment statistics to avoid the estimation distortion of high-order statistics under low task volume.

[0081] Before extracting residual statistical features, the residual data is truncated or shortened, and the participant feature vectors are standardized to eliminate scale differences between different statistical features and improve the stability of density estimation.

[0082] As a further preferred implementation, the above truncation or shortening processing can be achieved in the following ways: for the residual data of all participants, remove extreme values ​​below the 1st quantile or above the 99th quantile to complete the shortening processing; or for residual data exceeding 3 times the standard deviation of the residual sequence, replace them with boundary values ​​corresponding to 3 times the standard deviation to complete the truncation processing. The standardization processing refers to performing zero-mean and unit-variance standardization on each dimension of the feature vector to eliminate the dimensional differences between different statistical features.

[0083] Based on the above methods, the present invention further provides a mobile crowd sensing system, comprising: a data acquisition module for acquiring sensing data perturbed by local differential privacy; a residual construction module for constructing participant residuals based on initial aggregation results; a feature extraction module for extracting multidimensional statistical features from the residuals; an anomaly identification module for identifying abnormal participants based on a joint distribution model of multidimensional statistical features; and an aggregation control module for adjusting the aggregation strategy according to the anomaly identification results.

[0084] Specifically, in this embodiment, the crowd sensing system includes a set of participants. and a set of perception tasks Each participant uploads the raw perception data after randomly perturbing it according to a local differential privacy mechanism while performing the perception task. The system only performs abnormal participant identification and control based on the data after differential privacy perturbation.

[0085] Phase 1: Residual Statistical Feature Construction Phase

[0086] Step A1: The system collects differential privacy perturbation reporting data from all participants across multiple perception tasks and calculates the initial aggregation results for each perception task using a preset aggregation algorithm. The initial aggregation results are obtained by aggregating the participants' differential privacy perturbation reporting data using the median or by applying a weighted average based on historical consistency. These results are used to construct residual statistical features, rather than being directly output as the final perception result. As a further preferred implementation, the above-mentioned weighted average based on historical consistency refers to an aggregation method that assigns weights based on the degree of consistency between the participant's historical reported data and the group's historical aggregation results. Specifically, this can be achieved as follows: calculate the mean absolute value of each participant's historical residuals, and use the reciprocal of this mean as the participant's aggregation weight. A higher weight indicates a stronger consistency between the participant's historical behavior and the group's overall behavior. Then, perform a weighted average on the reported data of all participants for the current task to obtain the initial aggregation result for that task.

[0087] Step A2: Using the initial aggregation results as a baseline, calculate the residual data for each participant on each perceptual task. For each participant... In the mission The residual on is defined as: ,in, Indicates participants Differential privacy perturbation reported data, Indicates task The initial aggregation result.

[0088] Step A3: Based on the residual data of participants on multiple tasks, extract multidimensional statistical features to characterize the overall behavioral characteristics of participants.

[0089] Preferably, for participants residual data Calculate the following statistics respectively:

[0090] (1) Mean of residuals:

[0091] ,

[0092] (2) Root mean square of residuals: ,

[0093] (3) Residual skewness: ,

[0094] (4) Residual kurtosis: ,

[0095] in, For participants The standard deviation of the residual data.

[0096] Therefore, each participant is mapped to a multidimensional statistical feature vector:

[0097] The aforementioned multidimensional residual statistical features exhibit an inherent statistical coupling relationship under the influence of local differential privacy noise. The residual distribution of normal participants shows relatively stable consistency across multiple statistical dimensions, while abnormal participants, when attempting to introduce systematic biases to influence the aggregation results, find it difficult to consistently forge multiple statistical features while maintaining statistical acceptability.

[0098] Therefore, by jointly modeling the multidimensional residual statistical characteristics of participants, abnormal behaviors that are difficult to distinguish in a single perception task can be effectively amplified and identified at the multi-task statistical level, thereby achieving reliable identification of abnormal participants.

[0099] Phase Two: Kernel Density Modeling Based on Multidimensional Features

[0100] Step B1: The system collects the multidimensional statistical feature vectors of all participants and constructs a feature vector set. .

[0101] Step B2: Construct a kernel density estimation model based on the feature vector set to characterize the probability density distribution of normal participants in the multidimensional feature space.

[0102] Preferably, the probability density function of the kernel density estimation model is expressed as:

[0103] in, (·) represents a multidimensional kernel function with bandwidth h, where h is a smoothing parameter.

[0104] Phase Three: Identification of Abnormal Participants and System-Level Control Based on Density Assessment

[0105] Step C1: For the feature vector of each participant Calculate its density value under the kernel density estimation model:

[0106] Step C2: Construct anomaly scores based on density values, preferably defined as: , where ε is a very small positive number to prevent numerical overflow.

[0107] Step C3: Perform statistical analysis on the abnormal scores of all participants to determine the abnormality threshold. When a participant's abnormal score exceeds the threshold, the participant is identified as an abnormal participant.

[0108] As a further preferred implementation, the above adaptive determination of the anomaly judgment threshold can be achieved in the following ways: calculate the 95th percentile of the anomaly judgment threshold for all participants, and judge participants whose anomaly scores are higher than this threshold as anomalies; or use the 3σ criterion to judge participants whose anomaly scores are greater than "mean + 3 times standard deviation" as anomalies; or use the box plot method to judge participants whose anomaly scores are higher than the upper quartile + 1.5 times the interquartile range as anomalies.

[0109] Step C4: In the subsequent perception result aggregation process, restrict or adjust the data participation of abnormal participants in the aggregation process, and re-execute the aggregation based only on the remaining participants or the reported data after the weight adjustment to obtain robust perception results.

[0110] As can be seen from the above embodiments, the present invention can identify and control abnormal participants from the perspective of multidimensional statistical distribution under local differential privacy noise masking conditions, without the need to preset the abnormality ratio or attack model, and has the advantages of low computational overhead and large applicability.

[0111] Based on the same inventive concept, this invention also provides a computer device, comprising: one or more processors, and a memory for storing one or more computer programs; the programs include program instructions, and the processor executes the program instructions stored in the memory. The processor may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, used to implement one or more instructions, specifically for loading and executing one or more instructions stored in a computer storage medium to implement the above-described method.

[0112] It should be further explained that, based on the same inventive concept, the present invention also provides a computer storage medium storing a computer program, which, when executed by a processor, performs the above-described method. This storage medium can be any combination of one or more computer-readable media. A computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. A computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In the present invention, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0113] It should be noted that, unless otherwise defined, the technical or scientific terms used in this invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0114] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

[0115] This invention is not limited to the preferred embodiment described above. Anyone inspired by this invention can derive various other forms of methods for identifying and controlling abnormal participants in mobile crowd intelligence under local differential privacy constraints. All equivalent variations and modifications made within the scope of the claims of this invention shall fall within the scope of this invention.

Claims

1. A method for identifying and controlling abnormal participants in mobile swarm sensing under local differential privacy constraints, wherein the mobile swarm sensing system includes a server and multiple participants, and the participants report the original sensing data to the server after perturbing it through a local differential privacy mechanism, characterized in that... Includes the following steps: Acquire perception data submitted by multiple participants for multiple perception tasks, after being perturbed by a local differential privacy mechanism; An initial aggregation process is performed on the sensing data to obtain the initial aggregation results for each sensing task. The initial aggregation results are only used for residual construction and are not output as the final sensing results. Based on the initial aggregation results of each perception task, the residual of each participant on the corresponding perception task is calculated. The residual is the difference between the participant's perception data and the initial aggregation results of the corresponding perception task. Multidimensional statistical features are extracted from the residuals of each participant on multiple perceptual tasks to form a feature vector characterizing the participants’ behavioral patterns. The multidimensional statistical features at least characterize the location, scale, asymmetry and tail characteristics of the residual distribution. Based on the feature vectors of all participants, the joint probability distribution model of the multidimensional statistical features is performed to construct a joint probability distribution model that characterizes the behavior patterns of normal participants. Based on the degree of deviation of each participant's feature vector from the joint probability distribution model, abnormal participants are identified. Based on the identification results of anomalous participants, the aggregation strategy for the final perception results of the crowd-aware perception system is dynamically adjusted to suppress the influence of anomalous participant data on the final aggregation results.

2. The method for identifying and controlling abnormal participants in mobile swarm intelligence sensing under local differential privacy constraints as described in claim 1, characterized in that: The residual of participant i on the perceptual task j is ,in The perception data of participant i on perception task j is perturbed by a local differential privacy mechanism. The residuals are the initial aggregation results for the perception task j and are not used to recover the participants' original perception data.

3. The method for identifying and controlling abnormal participants in mobile swarm intelligence sensing under local differential privacy constraints as described in claim 1, characterized in that: The multidimensional statistical features include residual mean, residual root mean square, residual skewness, and residual kurtosis, which respectively characterize the location, scale, asymmetry, and tail properties of the residual distribution; the feature vector of participant i for: Where M represents the number of perceptual tasks that participant i participates in. Let the standard deviation of the residual sequence of participant i be . The residual of participant i on perception task j; the construction of the participant feature vector is based solely on the reported data after differential privacy perturbation and its relative residual statistical characteristics, without relying on the real perceived value, attack model assumptions or prior labels of the participant, and is not used to recover or infer the participant's original perceived data.

4. The method for identifying and controlling abnormal participants in mobile swarm intelligence sensing under local differential privacy constraints as described in claim 1, characterized in that: The joint probability distribution modeling adopts either nonparametric probability density estimation or semiparametric probability density estimation.

5. The method for identifying and controlling abnormal participants in mobile swarm intelligence sensing under local differential privacy constraints as described in claim 1, characterized in that: Based on the joint probability distribution model, the probability density of the feature vectors of the participants is evaluated, and an anomaly score that is negatively correlated with the probability density value is constructed. The anomaly score is used to quantify the degree to which the participants deviate from the behavior patterns of normal participants. By analyzing the distribution characteristics of the anomaly score in the participant set, an anomaly judgment threshold is adaptively determined, and abnormal participants are identified based on the anomaly judgment threshold.

6. The method for identifying and controlling abnormal participants in mobile swarm intelligence sensing under local differential privacy constraints as described in claim 1, characterized in that: The initial aggregation result is obtained by aggregating the perceived data after it has been perturbed by the local differential privacy mechanism, or by a weighted average based on historical consistency.

7. The method for identifying and controlling abnormal participants in mobile swarm intelligence sensing under local differential privacy constraints as described in claim 1, characterized in that: The dynamic adjustment of the aggregation strategy includes at least one of the following: reducing the weight of abnormal participants in the final aggregation process, delaying the adoption of data reported by abnormal participants, and restricting the participation of data from abnormal participants in the final aggregation.

8. The method for identifying and controlling abnormal participants in mobile swarm intelligence sensing under local differential privacy constraints as described in claim 1, characterized in that: When the number of perceptual tasks participated in by participants is less than a preset threshold, the statistical features corresponding to residual asymmetry and tail characteristics are stabilized, or robust statistics are used to replace at least some of the higher-order statistical features.

9. The method for identifying and controlling abnormal participants in mobile swarm intelligence sensing under local differential privacy constraints as described in claim 1, characterized in that: Before extracting multidimensional statistical features, the residual data is truncated or shortened, and the extracted feature vectors are standardized to eliminate scale differences between different statistical features.

10. A mobile crowd sensing system, characterized in that, To implement the method of any one of claims 1-9, comprising: The data acquisition module is used to acquire perception data submitted by multiple participants for multiple perception tasks, after being perturbed by a local differential privacy mechanism. The initial aggregation module is used to perform initial aggregation processing on the perceived data and output the initial aggregation result for residual construction only; The residual construction module is used to calculate the residuals of each participant on the corresponding perceptual task based on the initial aggregation results of each perceptual task. The feature extraction module is used to extract multidimensional statistical features from the residuals of each participant on multiple perceptual tasks to form a feature vector that characterizes the behavior pattern of the participants. The multidimensional statistical features at least characterize the location, scale, asymmetry and tail characteristics of the residual distribution. The distribution modeling module is used to perform joint probability distribution modeling on the multidimensional statistical features based on the feature vectors of all participants, and to construct a joint probability distribution model that characterizes the behavioral patterns of normal participants. The anomaly detection module is used to identify anomalous participants based on the degree of deviation of each participant's feature vector from the distribution of the joint probability distribution model. The aggregation control module is used to dynamically adjust the aggregation strategy of the final perception result of the crowd intelligence perception system based on the identification results of abnormal participants, so as to suppress the influence of abnormal participant data on the final aggregation result.

Images