An adaptive truncated differential privacy protection method for stream-oriented crowd sensing

Abstract

This invention discloses an adaptive truncated differential privacy protection method for streaming crowdsensing, comprising: generating prior estimates for sensing data generated by each mobile terminal, calculating the deviation between the real-time observations of the sensing data and the prior estimates; setting a dynamic reporting trigger threshold, comparing the deviation with the dynamic reporting trigger threshold: if the deviation is greater than the dynamic reporting trigger threshold, it is determined that the sensing data has undergone a significant change; otherwise, it is determined that the sensing data is in a period of stable change; adding noise to the sensing data that has undergone a significant change locally, and uploading the locally noisy sensing data to an edge server; actively truncating uplink communication for sensing data in a period of stable change; the edge server using a shuffling differential privacy mechanism to perform a random permutation operation on the locally noisy sensing data set to obtain an anonymous multisensing dataset; and the cloud server obtaining a global unbiased estimate of the sensing data based on the inferred value set of the sensing data and the anonymous multisensing dataset.

Description

Technical Field

[0001] This invention belongs to the field of streaming swarm intelligence sensing technology, and specifically relates to an adaptive truncated differential privacy protection method for streaming swarm intelligence sensing. Background Technology

[0002] In mobile crowdsourcing sensing (MCS) and IoT applications, massive numbers of terminal devices continuously generate and upload high-frequency streaming sensing data (such as traffic flow and health indicators). To prevent the leakage of sensitive information, Local Differential Privacy (LDP) technology is widely used. It allows terminals to add noise and perturb data locally before uploading, without relying on a trusted third party. However, in the face of continuous high-frequency streaming data environments, existing privacy protection mechanisms have exposed the following significant technical shortcomings:

[0003] 1. Timing redundancy leads to a double waste of communication overhead and privacy budget.

[0004] Existing streaming LDP mechanisms typically assume continuous data reporting at a fixed frequency, neglecting the autocorrelation of the underlying physical environment over time. In real-world scenarios, physical states (such as vehicle speed or ambient temperature on a road segment) often remain stable over short periods. Continuously uploading highly similar redundant sensing data not only consumes significant amounts of limited network uplink bandwidth but also leads to the unnecessary squandering of privacy budgets. Traditional static sampling or fixed threshold truncation strategies struggle to adapt to anomalous changes in complex, non-stationary environments, failing to strike a balance between truncation sensitivity and system robustness.

[0005] 2. The linear accumulation of privacy budgets significantly shortens the system lifecycle.

[0006] In streaming sensing scenarios, data is continuously published and aggregated over time. According to the cascade combination theorem of differential privacy, if a user executes the conventional LDP perturbation mechanism at every moment, the accumulated privacy budget consumption will increase linearly and exponentially over time. Once the accumulated consumption reaches the preset security limit, the terminal must be forcibly stopped from providing services to meet the privacy commitment. This will lead to a severe gap in available data in the later stages of system operation, resulting in an extremely short lifespan.

[0007] 3. The lack of an edge-cloud collaborative compensation mechanism leads to serious deviations in global aggregation.

[0008] To alleviate communication and budget pressures, some solutions attempt to allow terminals to remain silent when data is stable (i.e., physically truncate uploads). However, due to the lack of a global, cross-layer collaborative design, once a large number of terminals enter a silent state, traditional cloud aggregation mechanisms will face severe data loss, leading to significant deviations in global statistical estimates. Summary of the Invention

[0009] The purpose of this invention is to provide an adaptive truncated differential privacy protection method for streaming crowdsensing. By adaptively truncating temporal redundancy and utilizing a collaborative compensation streaming privacy protection method, it can achieve a globally optimal trade-off between privacy security, communication overhead, and long-term statistical utility.

[0010] The technical solution provided by this invention is as follows:

[0011] An adaptive truncated differential privacy protection method for streaming swarm intelligence sensing includes:

[0012] Generate prior estimates for the sensing data generated by each mobile terminal, and calculate the deviation between the real-time observed values ​​of the sensing data and the prior estimates;

[0013] A dynamic reporting trigger threshold is set, and the deviation is compared with the dynamic reporting trigger threshold: if the deviation is greater than the dynamic reporting trigger threshold, it is determined that the sensed data has undergone a significant change; otherwise, it is determined that the sensed data is in a period of stable change.

[0014] Sensing data that undergoes significant changes is noise-added locally and then uploaded to the edge server; uplink communication is actively cut off for sensing data in a period of stable change.

[0015] The edge server uses a shuffle differential privacy mechanism to perform random permutation on the local noisy sensing data set to obtain an anonymous multisensor dataset, and then uploads the anonymous multisensor dataset to the cloud server.

[0016] The cloud server infers from the sensing data during a period of stable change to obtain the inferred values ​​of the sensing data;

[0017] The cloud server infers a global unbiased estimate of the sensing data based on the set of values ​​inferred from the sensing data and the anonymous multisensor dataset.

[0018] Preferably, prior estimates are generated using the perception data produced by each mobile terminal using a Kalman filter.

[0019] Preferably, the formula for calculating the dynamic reporting trigger threshold is:

[0020] ;

[0021] in, express Time Mobile Terminal The dynamic reporting trigger threshold, express Time Mobile Terminal The local mean of the residuals during local online iteration. express Time Mobile Terminal The local variance of the residuals in the local online iteration. This is the sensitivity adjustment coefficient.

[0022] Preferably, the local mean of the local online iterative residual is calculated iteratively using the following formula:

[0023] ;

[0024] in, As a smoothing factor, express Time Mobile Terminal The deviation between the real-time observations and prior estimates of the generated sensing data. express Time Mobile Terminal The local mean of the residual is obtained through local online iteration.

[0025] Preferably, the local variance of the local online iterative residual is calculated iteratively using the following formula:

[0026] ;

[0027] in, As a smoothing factor, express Time Mobile Terminal The deviation between the real-time observations and prior estimates of the generated sensing data. express Time Mobile Terminal Local variance of residuals is obtained through local online iteration.

[0028] Preferably, the Laplace mechanism is used to locally add noise to the sensing data that has undergone significant changes.

[0029] Preferably, the method for generating the anonymous multisensor dataset is as follows:

[0030] Edge servers utilize trusted shufflers to collect local noisy sensing data sets at the current moment. ;

[0031] If the number of mobile terminals uploading noise-sensing data at the current moment satisfy: The edge server generates a size of Virtual Fill Dataset And combine the filled dataset with the local noisy sensing data set at the current time. Combining merged datasets ;

[0032] in, Minimum confusion threshold;

[0033] The trusted shuffler from symmetric group of order A permutation function is randomly selected uniformly from the middle. For merging datasets Applying a permutation operation yields an anonymous multisensor dataset. :

[0034] ;

[0035] In the formula, This is data from an anonymous multisensor dataset.

[0036] Preferably, the cloud server infers the perceived data value according to the following formula:

[0037] ;

[0038] in, This represents the inferred value from the perceived data. Indicates mobile terminal Kalman filter Posterior state estimation at time 10:00. express The cloud server obtains mobile terminals at any time Prior state estimation of perceived data, and These are the scalar state transition coefficient and the observation coefficient, respectively.

[0039] The beneficial effects of this invention are:

[0040] The adaptive truncated differential privacy protection method for streaming crowdsensing provided by this invention effectively combines "end-side adaptive truncation", "edge-side shuffling privacy amplification" and "cloud-based silent state inference" to construct a streaming privacy protection framework with cross-layer collaborative compensation capabilities between end, edge and cloud. It can adaptively truncate temporal redundancy and block the linear accumulation of privacy budget, achieving a globally optimal trade-off between privacy security, communication overhead and long-term statistical utility. Attached Figure Description

[0041] Figure 1 This is a graph showing the impact of privacy budget on data reconstruction accuracy in the experimental examples of this invention.

[0042] Figure 2 This is a graph showing the effect of the smoothing factor on the tracking utility in the experimental examples of this invention.

[0043] Figure 3(a) is a comparison of the root mean square error of different methods applied to the IBRL dataset under different privacy budgets in the experimental examples of this invention.

[0044] Figure 3(b) is a comparison of the root mean square error of different methods applied to the GeoLife trajectory dataset under different privacy budgets in the experimental examples of this invention.

[0045] Figure 4(a) is a comparison of the average communication overhead generated by applying different methods to the IBRL dataset under different privacy budgets in the experimental examples of this invention.

[0046] Figure 4(b) is a comparison of the average communication overhead generated by applying different methods to the GeoLife trajectory dataset under different privacy budgets in the experimental examples of this invention.

[0047] Figure 5(a) shows the time-series evolution of the instantaneous reconstruction absolute error of the IBRL dataset when different methods are applied to it in the experimental example of this invention.

[0048] Figure 5(b) shows the time-series evolution of the absolute error of instantaneous reconstruction of the GeoLife trajectory dataset when different methods are applied to it in the experimental example of this invention.

[0049] Figure 6(a) shows a comparison of the cumulative privacy budget consumption of different methods on the IBRL dataset in the experimental examples of this invention.

[0050] Figure 6(b) shows a comparison of the cumulative privacy budget consumption of different methods on the GeoLife trajectory dataset in the experimental examples of this invention. Detailed Implementation

[0051] The present invention will now be described in further detail with reference to the accompanying drawings, so that those skilled in the art can implement it based on the description.

[0052] The relevant methods involved in this invention will first be described below.

[0053] 1. Local differential privacy

[0054] Local differential privacy is a data privacy protection model for untrusted data collectors. Its core idea is that users randomize the raw data locally before uploading it, thereby limiting the risk of individual information leakage without relying on trusted third parties. It is currently widely used in large-scale distributed data collection and analysis scenarios.

[0055] Definition 1 (Local Differential Privacy) Define the domain for user input data. To randomize the output space. Given a random mechanism. If for any two adjacent inputs and any output subset All satisfy:

[0056] ;

[0057] This is called a random mechanism. satisfy -Local differential privacy. Among them, >0 and is a privacy budget parameter.

[0058] Definition 2 (Laplace Mechanism) Given a query function ,That -Sensitivity Defined as any adjacent input The maximum output of the function between Norm distance is:

[0059] ;

[0060] The Laplace mechanism achieves privacy protection by adding noise that follows a Laplace distribution to the true value. For the input variable... The mechanism output is Among them, noise .

[0061] In streaming data publishing scenarios, users generate data sequences on a continuous time index and publish perturbed results at each moment through a random mechanism. Under this continuous reporting paradigm, the risk of privacy leakage will exhibit a significant cumulative effect over the time domain.

[0062] Theorem 1 (Sequential Combination Theorem): Suppose that a set of random mechanisms are executed sequentially on the data stream of the same user. , of which Mechanism Each satisfies - Local differential privacy. Then the joint output of the entire sequence. satisfy -LDP.

[0063] According to the sequential combination theorem, in a streaming publishing model, if the user executes at every time step... With a -LDP mechanism, the accumulated privacy budget will grow linearly over time. This characteristic significantly limits the sustainability of the privacy mechanism when the system has a long operating cycle or a high data release frequency. This mathematical property highlights the urgency of suppressing temporal redundancy at the mechanism design level; that is, reasonable triggering strategies must be introduced to reduce ineffective data release operations in order to curb the long-term excessive accumulation of the privacy budget.

[0064] 2. Shuffling Differential Privacy

[0065] To bridge the utility gap between LDP and centralized differential privacy, this invention employs a shuffling model. This model, based on an Encoder-Shuffler-Analyzer (ESA) architecture, utilizes an anonymous permutation mechanism at the edge to amplify privacy under the assumption of decentralized trust.

[0066] Definition 3 (Shuffle Differential Privacy): The shuffle model consists of an encoder deployed on the user side, i.e., a local randomizer. Shufflers deployed on the edge and analyzers deployed in the cloud It consists of three core components. Let... for The original data vector of each user at the current moment. Each user first independently executes the following... -LDP mechanism Generate perturbation data Then the shuffler use Random permutation of order The received vector is obfuscated to sever the association between data and identity, resulting in an unordered multiset:

[0067] ;

[0068] Finally, the analyzer Statistical aggregation is performed solely based on this unordered multiset.

[0069] Lemma 1 (Shuffle Privacy Amplification) If the local encoder satisfy -LDP, then the shuffling mechanism For the analyzer to satisfy - Differential privacy. When and At that time, centralized privacy budget The upper bound is:

[0070] ;

[0071] This lemma shows that the shuffling mechanism introduces a user scale factor. It can allocate privacy budget from Reduce to This allows the system to significantly improve the accuracy of data aggregation while maintaining high global privacy and security.

[0072] 3. Problem Definition

[0073] Suppose a given Mobile crowd sensing system composed of mobile users (mobile terminals) In discrete timestamps On, each user Generate real-time sensing data .make and Representing time respectively The aggregated estimated value and the true value are compared, and the core utility objective of the system is to minimize the total time period. Cumulative mean square error (MSE) within:

[0074] ;

[0075] However, continuous data release leads to an accumulation of privacy risks. As Theorem 1 states, if users simply perform [the following action] at every moment... - The LDP mechanism is used for reporting, so at time... The cumulative privacy budget expenditure will reach When the accumulated consumption exceeds the preset privacy budget... If this happens, the user will stop the service, causing the data availability to plummet to zero in subsequent moments.

[0076] To extend service life within a limited privacy budget, this invention introduces a binary variable. To represent users exist The act of reporting in real time. When When the user executes the shuffle mechanism, they publish data and deduct from the budget; otherwise, the user remains silent, and the cloud will make inferences based on historical states.

[0077] Therefore, the research objective of this invention can be formalized as: under the premise of satisfying global differential privacy constraints, to jointly optimize the reporting decision. This involves a constrained optimization problem that minimizes long-term aggregation error, along with a local prediction mechanism. The mathematical expression is as follows:

[0078] ;

[0079] ;

[0080] In the formula, This represents the estimated variance introduced by the local perturbation mechanism. This represents the estimation bias generated by the prediction model during the local silent period. The nature of this optimization problem requires designing an adaptive triggering strategy. (Within the budget...) Under the constraints, the noise error brought by instant communication and the prediction error brought by silence are dynamically weighed in order to maximize the utility throughout the entire life cycle.

[0081] This invention assumes that the edge shuffler is an honest and curious semi-trusted entity, and that there is no collusion between the edge server and the cloud server. That is, the shuffler faithfully executes the random permutation protocol, but does not leak the permutation mapping relationship of the data to the cloud.

[0082] The adaptive truncation differential privacy protection method for streaming crowd sensing provided by this invention will be described in detail below.

[0083] The framework of the Adaptive Truncated Differential Privacy Preservation Method for Streaming Crowd Sensing (AdaSSP) consists of a user layer, an edge layer, and a cloud service layer. Its operation mechanism includes the following two key stages.

[0084] Phase 1: At the user level, each mobile terminal independently runs a lightweight time series prediction module to generate prior estimates of the current perceived value. Subsequently, the system adaptively judges the stability of the current underlying environment by evaluating the deviation between real-time observations and estimates, combined with a dynamic threshold mechanism. If the perceived data is in a period of stable change, the mobile terminal actively cuts off uplink communication to maintain silence and avoid ineffective consumption of privacy budget; if a significant change in data is detected, the terminal activates the reporting mechanism and utilizes a local randomization mechanism. Add noise to real data to generate sparse, noisy data. And upload it to the edge node.

[0085] In Phase 1, to suppress the privacy budget consumption of redundant data in streaming sensing scenarios from the source, this invention establishes a state-space model based on a one-dimensional linear Gaussian system and designs a lightweight adaptive predictive Kalman filter suitable for mobile terminals to generate prior estimates of the current sensed values. This mechanism utilizes the temporal autocorrelation of sensed data to dynamically determine data reporting behavior by quantifying the deviation between prior estimates and real-time observations. The specific implementation process of Phase One is as follows.

[0086] (1) Prediction-driven reporting triggering mechanism

[0087] Assume that the underlying physical state of mobile terminal i follows a discrete-time dynamic system. To adapt to the computing power constraints of the mobile terminal, the standard Kalman filter is degenerated into a one-dimensional scalar system. Define the time interval... Hidden true state Compared with the actual observation value of the sensor The evolution process is as follows:

[0088] ;

[0089] In the formula, and These are the scalar state transition coefficient and the observation coefficient, respectively. and These represent independent process noise and observation noise, respectively.

[0090] At any time The local LAP-KF module independently performs time updates. This is based on the posterior state estimate from the previous time step. To extrapolate the predicted values ​​from prior observations at the current moment. As the default inferred value for global aggregation by the cloud server when the user is currently in a silent state.

[0091] ;

[0092] The Kalman filter's prediction logic is executed twice in the AdaSSP architecture: as a local predictor on the mobile terminal side to trigger decisions, and as a shadow prediction model in the cloud side for missing data completion. Except for the initial input, the two are completely decoupled on the physical link to ensure the independence of cloud inference during communication truncation.

[0093] (2) Determination of the dynamic threshold of exponentially weighted moving average

[0094] Define time Real-time prediction residuals The absolute deviation between the actual observed value and the prior predicted value:

[0095] ;

[0096] Because streaming sensing environments are typically non-stationary, traditional static thresholding strategies struggle to balance anomaly detection sensitivity with system robustness. This invention introduces an exponentially weighted moving average algorithm to construct an adaptive threshold update mechanism. The system iterates locally online using the local mean of the residuals. With local variance :

[0097] ;

[0098] ;

[0099] Therefore, time Adaptive reporting trigger threshold for:

[0100] ;

[0101] in, As a smoothing factor, it controls the model's memory weights for historical residual fluctuations; This is the sensitivity adjustment coefficient, used to adjust the tolerance of the dynamic envelope to sudden abnormal data.

[0102] (3) Adaptive reporting triggering and status update

[0103] Based on the dynamic threshold obtained above This invention establishes an adaptive data reporting triggering mechanism. Binary decision variables. The judgment criteria are as follows:

[0104] ;

[0105] in, To predict residuals in real time The absolute value of represents the deviation between the real-time observation and the prior estimate.

[0106] When predicting residuals At that time, determine the current ( (Time) Mobile Terminal The generated sensing data is in a stable period; at this time, the local triggering mechanism determines... The terminal device actively cuts off the uplink communication link to maintain silence, using only the actual observations. Locally, Kalman measurement updates y are implicitly performed to calibrate the posterior state. In this state, the mobile terminal No data was leaked, and the privacy budget was strictly consumed at zero. When a sudden change in the underlying environment occurs and prior predictions fail, the local triggering mechanism makes a determination. mobile terminal Inject differential privacy noise into real observations to generate perturbed data (locally noisy sensing data). Then, the data upload and edge shuffling process is activated, and the local predictor is reset simultaneously.

[0107] Phase Two: Trusted Shuffler Deployed at the Edge Layer It acts as a privacy firewall between users and untrusted cloud servers, responsible for collecting a set of sparse data that triggers reporting at any given moment. The data is then randomly permuted to sever the link between the data value and the physical user's identity, thus amplifying privacy. The shuffled anonymous multiset is then sent to the cloud, where it combines the prior predictions of silent users to complete the unbiased aggregation of the global state.

[0108] In Phase Two, AdaSSP introduces a trusted shuffling mechanism at the edge, with the cloud server combining prior inferences from silent users to complete global aggregation. The specific implementation process of Phase Two is as follows.

[0109] (1) Local privacy disturbances and edge shuffling mechanism

[0110] make Indicates at time The set of active mobile terminals that triggered the reporting, the number of active mobile terminals is denoted as . For any active mobile terminal Using the Laplace mechanism to analyze real-world perception data generated by mobile terminals Inject independent noise to generate noisy data .

[0111] ;

[0112] In the formula, To perceive the global sensitivity of the data, The local privacy budget allocated for each data upload.

[0113] After this operation, the mobile device updates its remaining privacy budget:

[0114] ;

[0115] If the updated remaining budget satisfies The mobile terminal will be forced into a silent state during subsequent detection.

[0116] Subsequently, the trusted shuffler of the edge server Collect a noisy dataset of all active users at the current moment. To ensure privacy amplification limits during shuffling in extremely stable scenarios, the shuffler pre-defines a minimum obfuscation threshold derived from the global privacy cap. If the current number of active users The edge server will proactively generate a size of Virtual Fill Dataset .in, Each element in the dataset is independently sampled from the same Laplace distribution as the actual perturbation data. At this point, the total data size actually involved in the reshuffling is... The shuffler from symmetric group of order A permutation function is randomly selected uniformly from the middle. For the merged set Apply a permutation operation to generate an anonymous multiset and output it to the cloud:

[0117] .

[0118] This operation can conceal the mapping relationship between specific noisy data and user identifiers, effectively amplifying privacy during the shuffling process while maintaining the overall statistical distribution characteristics.

[0119] (2) Cloud-based state reconstruction and global aggregation

[0120] Due to the prediction truncation mechanism, Only one moment One active user uploaded noisy data, the rest... A silent user (mobile terminal) did not interact with data because its perceived data was in a period of steady change. According to the classic event-triggered state estimation theory in cybernetics, to supplement the missing statistical information, the cloud needs to synchronously maintain the shadow prediction model for each silent user. That is, the cloud generates the prior inference value for the current moment based on the posterior state of the previous moment. .

[0121] Finally, by combining the inferred values ​​from the anonymous multiset received in the cloud with the perceived data generated by the silent mobile terminal, a globally unbiased estimator is calculated. :

[0122] ;

[0123] in, The inferred value is the perceived data of the silent mobile terminal. This represents the total amount of noisy data uploaded by active users who triggered the reporting mechanism at the current moment. The process for obtaining this value is as follows: Active users... Laplace noise is added to real observations using a local differential privacy mechanism to generate [the data]. The anonymized multiset is then obfuscated by a trusted shuffler at the edge node, and finally obtained by the cloud server through linear summation of the received anonymized multiset.

[0124] Since the independent random noise introduced by the Laplace mechanism and the expected value of the virtual padding data are both zero, and the prior inferences of silent users are generated based on the time update equation of the Kalman filter, the linear summation and aggregation operation in the above formula strictly guarantees the global statistics while filtering out individual privacy features. The unbiasedness of .

[0125] The Adaptive Truncated Differential Privacy Protection Method (AdaSSP) for Streaming Crowd Sensing provided by this invention integrates the above-mentioned local prediction, privacy perturbation, edge shuffling and cloud aggregation processes, and the complete execution steps are as follows.

[0126] Step S1 (System Initialization):

[0127] Obtain system input parameters, including the mobile user set. Total time period Initial privacy budget total System parameters Smoothing factor With spreading factor For sets Any user in Initialize its available remaining privacy budget The posterior state of the local Kalman filter and initial adaptive threshold .

[0128] Step S2 (Edge-side local prediction and adaptive trigger determination): at any time during the time period Initialize the set of active users at the current moment. .gather Each user in The following local computations are performed in parallel:

[0129] S21 (Bias Assessment): The terminal calculates the prior state estimate at the current moment. and combined with actual observations Calculate the predicted residual Simultaneously, the adaptive threshold is updated based on the state of the previous time step. .

[0130] S22 (Dual Trigger Decision): The terminal determines whether condition one (predicted residual greater than the threshold) is met simultaneously. Condition 2 (the remaining budget at the previous time step is sufficient to cover a single instance of noise addition) and condition 3 (that is, If so, then the decision variable will be triggered. Set to 1 to move the user to the active user set. Perturbation data is generated using local differential privacy mechanisms. Report to higher authorities, deduct the corresponding budget, and update. If not: then the decision variable will be triggered. Set to 0, the terminal actively cuts off the uplink and remains silent; maintain the remaining budget unchanged, i.e. .

[0131] S23 (Implicit Update): Regardless of the triggering result, the terminal uses the current true observation value. Implicitly update the posterior state of the local Kalman filter This is for the purpose of simulation in the next moment.

[0132] Step S3 (Edge-side Trusted Shuffle and Security Compensation): Edge nodes collect perturbation data uploaded by all active users at the current moment to construct a perturbation dataset. And record the total number of currently active users. .

[0133] S31 (Virtual Fill Mechanism): Edge nodes determine the number of currently active users. Is it less than the preset minimum safety threshold? If so, then the generation scale is... Virtual Fill Dataset and update the perturbation dataset. .

[0134] S32 (Anonymous Permutation): A Trusted Shuffler Within Edge Nodes For the updated dataset Perform uniform random permutation Completely shuffle the data order and output anonymous multisets. Send to the cloud.

[0135] Step S4 (Cloud-based Joint Inference and Unbiased Aggregation): The cloud server receives the anonymous multiset. Then, synchronously retrieve all items in a silent state, i.e. The user's prior inference value. A linear summation is performed on the anonymous multiset and silent inference value to output the global unbiased statistical estimate at the current time step. .

[0136] Step S5 (Time Rolling): Time advances forward, and steps S2 to S4 are executed repeatedly until the total time period is reached. Finally, the cloud outputs a complete sequence of global statistical estimates. .

[0137] Test case

[0138] 1. Experimental setup

[0139] This experimental platform uses a Windows 11 operating system and is equipped with an Intel Core i7-13700KF CPU and 32GB of RAM. All core algorithms and baseline comparisons are implemented using Python 3.9. The experiments are based on real open-source sensor and trajectory datasets, extracting sensor information within a portion of continuous time windows, and performing simulation tests after cleaning, fitting, and resampling. Specifically, the datasets include the following three types:

[0140] The IBRL dataset contains temperature data collected over several months from 54 sensors, exhibiting significant temporal autocorrelation and stationarity, and is used to evaluate the cutoff performance of algorithms during periods of environmental stability.

[0141] GeoLife trajectory dataset: records a large number of users' continuous latitude and longitude coordinate streams, with strong data dynamics and many sudden changes.

[0142] Synthetic dataset Syn: A controlled time series stream generated based on the classic time series autoregressive model AR(1), used for parameter sensitivity analysis and theoretical extreme value testing of the algorithm.

[0143] This paper compares the AdaSSP algorithm with the Exact-Agg algorithm for exact aggregation, the CS-DP algorithm for concurrent shuffling differential privacy, the PS-LDP algorithm for pattern-sensitive local differential privacy, and the AdaScs algorithm for continuous adaptive sampling. Specifically, Exact-Agg is a plaintext aggregation benchmark without any privacy noise, used to calibrate the absolute lower bound of the system's statistical error. CS-DP represents a traditional shuffling architecture lacking a prediction truncation module. PS-LDP serves as an advanced local perturbation scheme for streaming time-series data. AdaScs, on the other hand, is a representative privacy-free adaptive data collection algorithm that dynamically determines whether to sample and upload current data using a plaintext policy by assessing dynamic environmental changes, thereby reducing system communication overhead.

[0144] The system environment and algorithm parameters are configured according to the default values ​​shown in Table 1. Among them, the privacy budget for a single report... The test interval references the latest benchmark settings in the differential privacy domain. Sensitivity adjustment coefficient. Strictly adhering to the classic statistical principle of Laida's criterion, a 99.73% confidence interval is ensured under stationary conditions. Furthermore, privacy budgets are considered. and smoothing factor As a core independent variable, it will be dynamically adjusted in subsequent utility and sensitivity analyses. To ensure a privacy lower bound in extremely stable scenarios, the system references strong privacy theory benchmarks in the differential privacy domain. Given a fault tolerance rate, the global central privacy budget after edge node shuffling and confusion, and the number of users participating in the shuffling, satisfy the following shuffling privacy amplification constraint:

[0145]

[0146] In the formula, For global central privacy budget, For a single reported local privacy budget, For fault tolerance, for The number of active users at any given moment.

[0147] The system sets a global privacy leak limit target. By substituting this into the above constraint formula as a boundary condition and performing rigorous derivation, the minimum shuffle confusion threshold required to guarantee this privacy lower bound can be solved inversely. .

[0148] Table 1 Default Parameter Settings for the Experiment

[0149]

[0150] 2. Parameter settings

[0151] To eliminate the randomness of uncontrollable interference and Laplace random noise in real-world scenarios, privacy budgeting was tested on the synthetic dataset Syn. With smoothing factor Sensitivity and expected value analyses were conducted using two core independent variables, and the results are as follows: Figure 1 Figure 2 As shown in the figure. The threshold spreading coefficient k remains fixed at 3, strictly following the Raida criterion described above.

[0152] like Figure 1 As shown, with As the value increases, the theoretical expectation of RMSE exhibits a significant "L"-shaped convergence trend. When When the value increases from 0.1 to 0.5, the error drops sharply; while when... Once the value exceeds 0.5, the curve flattens out, and the marginal benefit of error reduction diminishes significantly. Analysis indicates that setting the privacy budget for a single reporting session to [value missing]... A value of 0.5 achieves the best compromise between strict privacy protection and high data utility.

[0153] Figure 2 It indicates Impact on tracking utility. Smoothing factor. The confidence weights for historical residual fluctuations and current observations in the time series prediction module were adjusted. Too small a value will cause the system to lag significantly in its response to sudden changes, while too large a value will weaken the mechanism's smoothness and noise resistance. Experimental results show that when... At a value of 0.3, the system achieves the optimal balance between dynamic response and smooth noise reduction. Based on the above utility and sensitivity analysis conclusions, this optimal parameter combination will be consistently used in subsequent benchmark experiments in real-world scenarios. =0.5, =0.3 is used for the default configuration.

[0154] 3. Statistical utility analysis

[0155] Figures 3(a) and 3(b) show the comparison results of the root mean square error of each algorithm under different privacy budgets on two real datasets. As can be seen from the figures, with... As the noise variance injected by the Laplace mechanism increases, it gradually decreases, and the RMSE of each algorithm shows a monotonically decreasing trend, converging towards the noiseless true lower bound of Exact-Agg. AdaScs, lacking privacy protection, derives its error solely from interpolation bias caused by discarded data, and... The values ​​are irrelevant, remaining constant at 0.08 and 0.2 in the two datasets, respectively. CS-DP suffers from excessive noise in a single perturbation under a very small budget, leading to a high initial system error. However, as... The increase in swirl amplifies the error, causing it to converge rapidly. PS-LDP initially exhibits some local noise immunity, but its inherent phase lag in its global smoothing strategy leads to... When the value exceeds 1, the downward trend of its error reaches a significant bottleneck.

[0156] In comparison, the AdaSSP framework achieved optimal reconstruction accuracy across the entire privacy range. With an extremely low budget, it effectively filtered out noisy and redundant data through an edge-based prediction truncation mechanism, avoiding a surge in initial error from the basic shuffling mechanism. With a moderate increase in budget, the system accurately captured high-value state transitions through adaptive residual truncation and safely amplified them through edge shuffling, overcoming not only the time delay limitation of PS-LDP but also... A value greater than 1 effectively exceeds the interpolation error limit of the plaintext communication baseline AdaScs.

[0157] 4. Communication Overhead Analysis

[0158] Figures 4(a) and 4(b) show the comparison results of the average communication overhead of each algorithm under different privacy budgets. As shown in Figure 4, CS-DP and PS-LDP lack an end-side adaptive truncation mechanism, requiring the system to force full-volume noisy upload at each time step, maintaining 100% full overhead throughout the privacy range. Although the AdaScs algorithm reduces some overhead through a preset frequency reduction strategy, it cannot dynamically adapt to data evolution, and plaintext interaction poses a privacy leakage risk. In contrast, the AdaSSP algorithm provided by this invention exhibits superior time-series awareness and communication reduction capabilities. Under extremely small budgets, due to strong perturbation noise, frequent end-side prediction out-of-bounds errors require the system to maintain a high upload frequency to correct state drift. With the increase of [the specific element], the predictive advantage of higher-order Kalman state spaces is fully realized, and its communication overhead decreases rapidly. When [the specific element] is increased, [the specific element] is reduced. As restrictions continue to ease, AdaSSP will eventually... At 2.0, absolute communication compression of up to 84.5% and 50.5% were achieved, respectively. These results demonstrate that AdaSSP, through cross-layer collaboration of "prediction-truncation," completely overcomes the bandwidth bottleneck inherent in traditional streaming privacy protection.

[0159] 5. Temporal evolution of instantaneous errors

[0160] To evaluate the instantaneous tracking capabilities and robustness of different mechanisms under varying data distribution characteristics, we uniformly configured the datasets on the stable indoor sensor dataset IBRL and the high-dynamic spatial trajectory dataset GeoLife. =0.5, and comparative experiments were conducted on the instantaneous absolute error.

[0161] As shown in Figure 5(a), in a stationary data scenario, CS-DP exhibits drastic random fluctuations in its instantaneous error due to the independent injection of differential privacy noise at each time step. Although PS-LDP and AdaScs mitigate this through smoothing or truncation strategies, their errors remain at a high level. In contrast, the AdaSSP mechanism effectively utilizes the temporal correlation of historical states for filtering and noise reduction, maintaining its instantaneous error at the lowest global level. In the highly dynamic abrupt change scenario shown in Figure 5(b), the spatial trajectory undergoes a drastic jump at t=65. CS-DP experiences a sudden surge in error due to triggering extreme dynamic sensitivity. PS-LDP and AdaScs, on the other hand, expose significant hysteresis error peaks and stepped blind zones, respectively. In contrast, AdaSSP, through its state covariance adaptive expansion mechanism, achieves rapid correction and convergence with only minor perturbations. Therefore, the AdaSSP mechanism of this invention can effectively break the "noise positive feedback" trap in both stationary and abrupt data streams, achieving accurate tracking of the true state, with overall performance significantly superior to the baseline algorithm.

[0162] 6. Comparison of cumulative privacy budget consumption

[0163] Figures 6(a) and 6(b) illustrate the cumulative privacy budget consumption trends of each mechanism in a streaming environment. As shown, CS-DP and PS-LDP, lacking edge-side state awareness, force continuous noise perturbation at each time step, resulting in a strictly linear and steep increase in their budget consumption, easily exhausting the terminal's initial budget in a short time. To measure the overall cost of each mechanism in the same dimension, the experiment equates a single unprotected data transmission of the plaintext mechanism AdaScs to the consumption of a single full privacy budget, i.e., a complete privacy leakage penalty. Although AdaScs can mitigate its consumption in a stepwise manner through threshold truncation, its equivalent cumulative overhead still rises rapidly in highly dynamic and abrupt scenarios due to frequent breaking of the fixed threshold.

[0164] In contrast, the AdaSSP method provided by this invention achieves precise filtering of redundant data through lightweight adaptive residual truncation. In stable scenarios, its budget consumption curve is almost horizontal, achieving extremely low static overhead. Even in highly dynamic and abrupt scenarios, AdaSSP only triggers necessary noisy updates when there are significant jumps in the physical state, completing state correction with minimal step transitions. Experiments show that AdaSSP not only fundamentally breaks through the bottleneck of linear consumption over time in traditional streaming privacy protection, but its overall budget consumption is also consistently lower than the plaintext adaptive baseline AdaScs, significantly reducing privacy overhead by more than 80% over the entire cycle, demonstrating extremely superior long-term budget control capabilities.

[0165] To address the communication congestion and rapid depletion of privacy budgets caused by high-frequency continuous data reporting in streaming crowdsourcing sensing, this invention provides a time-series prediction-enhanced streaming shuffling privacy protection method (AdaSSP). This method first uses a lightweight adaptive Kalman filter to separate the stationary and fluctuating periods of the data and performs predictive truncation, resolving communication and privacy redundancy caused by high-frequency data acquisition. Then, based on this, cross-layer collaborative privacy amplification is performed, employing an edge-side trusted shuffling method to further anonymize the truncated sparse uploaded data. Finally, an unbiased aggregation and reconstruction technique is applied to the global data distribution to improve the accuracy of data reconstruction. Experiments comparing this method with four other mechanisms—Exact-Agg, CS-DP, PS-LDP, and AdaScs—show that, when dealing with resource-constrained datasets requiring long-term continuous sensing, the AdaSSP method provides by this invention achieves better reconstruction accuracy and lower edge-side runtime overhead while reducing communication costs. Future work considers extending this mechanism to multi-dimensional non-stationary sensing scenarios and data collection with spatiotemporal correlations.

[0166] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. They can be applied to various fields suitable for the present invention. For those skilled in the art, other modifications can be easily made. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details and illustrations shown and described herein.

Claims

1. An adaptive truncated differential privacy protection method for stream-oriented crowd sensing, characterized in that, include: Generate prior estimates for the sensing data generated by each mobile terminal, and calculate the deviation between the real-time observed values ​​of the sensing data and the prior estimates; A dynamic reporting trigger threshold is set, and the deviation is compared with the dynamic reporting trigger threshold: if the deviation is greater than the dynamic reporting trigger threshold, it is determined that the sensed data has undergone a significant change; otherwise, it is determined that the sensed data is in a period of stable change. Sensing data that undergoes significant changes is noise-added locally and then uploaded to the edge server; uplink communication is actively cut off for sensing data in a period of stable change. The edge server uses a shuffle differential privacy mechanism to perform random permutation on the local noisy sensing data set to obtain an anonymous multisensor dataset, and then uploads the anonymous multisensor dataset to the cloud server. The cloud server infers from the sensing data during a period of stable change to obtain the inferred values ​​of the sensing data; The cloud server infers a global unbiased estimate of the sensing data based on the set of values ​​inferred from the sensing data and the anonymous multisensor dataset.

2. The adaptive-truncated differential privacy protection method for stream-oriented mobile crowdsensing according to claim 1, wherein, Prior estimates are generated using the perception data produced by each mobile terminal using a Kalman filter.

3. The adaptive-truncated differential privacy protection method for stream-oriented mobile crowdsensing according to claim 2, wherein, The formula for calculating the dynamic reporting trigger threshold is as follows: ； wherein, denotes the moment mobile terminal a dynamic reporting trigger threshold, denotes the moment mobile terminal a local mean of the locally online iterated residuals, denotes the moment mobile terminal a local variance of the locally online iterated residuals, is a sensitivity adjustment factor.

4. The adaptive truncated differential privacy protection method for streaming swarm intelligence sensing according to claim 3, characterized in that, The local mean of the residuals from the local online iteration is calculated iteratively using the following formula: ； in, As a smoothing factor, express Mobile terminal The deviation between the real-time observations and prior estimates of the generated sensing data. express Time Mobile Terminal The local mean of the residual is obtained through local online iteration.

5. The adaptive truncated differential privacy protection method for streaming swarm intelligence sensing according to claim 3 or 4, characterized in that, The local variance of the local online iterative residual is calculated iteratively using the following formula: ； in, As a smoothing factor, express Time Mobile Terminal The deviation between the real-time observations and prior estimates of the generated sensing data. express Time Mobile Terminal Local variance of residuals is iterated online locally.

6. The adaptive truncated differential privacy protection method for streaming swarm intelligence sensing according to claim 5, characterized in that, The Laplace mechanism is used to add noise locally to the sensing data that has undergone significant changes.

7. The adaptive truncated differential privacy protection method for streaming swarm intelligence sensing according to claim 6, characterized in that, The method for generating anonymous multisensor datasets is as follows: Edge servers utilize trusted shufflers to collect local noisy sensing data sets at the current moment. ; If the number of mobile terminals uploading noise-sensing data at the current moment satisfy: The edge server generates a size of Virtual Fill Dataset And combine the filled dataset with the local noisy sensing data set at the current time. Combining merged datasets ; in, Minimum confusion threshold; The trusted shuffler from symmetric group of order A permutation function is randomly selected uniformly from the middle. For merging datasets Applying a permutation operation yields an anonymous multisensor dataset. : ； In the formula, This is data from an anonymous multisensor dataset.

8. The adaptive truncated differential privacy protection method for streaming swarm intelligence sensing according to claim 7, characterized in that, The cloud server infers the perceived data value based on the following formula: ； in, This represents the inferred value from the perceived data. Indicates mobile terminal Kalman filter Posterior state estimation at time 10:

00. express The cloud server obtains mobile terminals at any time Prior state estimation of perceived data, and These are the scalar state transition coefficient and the observation coefficient, respectively.

Images