A privacy-sensitive federated learning method for differential forgetting based on trajectory clusters
By employing a multi-dimensional privacy sensitivity assessment based on trajectory clusters and a differentiated forgetting strategy, this study addresses the insufficient privacy protection of trajectory data in existing federated learning methods, achieving efficient and accurate privacy forgetting and model performance maintenance.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HUAIYIN INSTITUTE OF TECHNOLOGY
- Filing Date
- 2026-02-26
- Publication Date
- 2026-06-02
AI Technical Summary
Existing federated learning forgetting methods lack specificity, resulting in insufficient consideration of the differences in privacy sensitivity of trajectory data, which may lead to degraded model performance or persistent privacy risks.
Privacy sensitivity assessment based on trajectory clusters is performed by quantifying multiple dimensions such as semantic features, spatiotemporal features and individual identifier correlation to generate a differentiated forgetting intensity coefficient. Cluster-level reverse gradient updates are performed on the client side, and the global model is aggregated on the server side.
It achieves accurate and efficient forgetting of highly privacy-sensitive trajectory clusters, reducing the risk of privacy leakage, while minimizing the negative impact on the overall model performance, thus balancing privacy protection and model usability.
Smart Images
Figure CN122132803A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of federated learning and privacy protection technology, specifically to a federated learning differential forgetting method based on trajectory cluster privacy sensitivity. Background Technology
[0002] With the rapid development of mobile internet and location technology, a large amount of trajectory data has been generated and collected. This trajectory data contains sensitive information such as users' travel habits and activity patterns, and has significant application value in fields such as traffic planning and intelligent recommendation. Federated learning, as a privacy-preserving computing technology, allows participating parties to collaboratively train models while retaining data locally, effectively solving the problems of data silos and data privacy leaks. However, when users request data deletion or forgetting, how to efficiently and accurately "forget" the impact of specific data on the model within the federated learning framework, while minimizing the damage to the overall model performance, is a key problem that urgently needs to be solved.
[0003] Existing federated learning forgetting methods mostly employ a uniform approach, treating the entire model or all data uniformly, lacking specificity. For trajectory data, the privacy sensitivity of different trajectories varies greatly. A uniform forgetting strategy may either lead to excessive forgetting, causing a severe decline in model performance, or insufficient forgetting, failing to effectively protect the privacy of highly sensitive trajectory data. Therefore, a method is needed that can differentiate forgetting based on the privacy sensitivity of trajectory data. Summary of the Invention
[0004] To address the aforementioned technical problems, the present invention aims to overcome the shortcomings of existing federated learning forgetting methods, such as insufficient consideration of the differences in privacy sensitivity of trajectory data, the tendency for excessive forgetting to lead to model performance degradation, or insufficient forgetting resulting in remaining privacy risks. The invention proposes a federated learning differential forgetting method based on trajectory cluster privacy sensitivity. This method uses a "trajectory cluster" as the smallest management and forgetting unit. On the client side, trajectory data is clustered into trajectory clusters. The privacy sensitivity of each trajectory cluster is quantitatively evaluated and hierarchically graded from multiple dimensions, including semantic features, spatiotemporal features, and individual identifier correlation. Simultaneously, during federated training, the impact of each trajectory cluster on model parameters is calculated. The model calculates the contribution of each update and then combines privacy sensitivity with contribution to generate a differentiated forgetting intensity coefficient and forgetting triggering conditions. After triggering, it performs cluster-level backward gradient updates on the target trajectory clusters and uploads them to the server for aggregation to obtain the forgotten global model. This allows for stronger forgetting of high privacy-sensitive trajectory clusters and avoids excessive forgetting of high-contribution trajectory clusters through a suppression mechanism. While satisfying the "right to be forgotten" and reducing the risk of trajectory privacy leakage, it minimizes the impact on the overall performance and usability of the federated learning model. Furthermore, it achieves efficient forgetting processing without full retraining through cluster-level forgetting and aggregation updates, effectively solving the aforementioned problems.
[0005] This invention is achieved through the following technical solution:
[0006] A privacy-sensitive federated learning differential forgetting method based on trajectory clusters includes the following steps:
[0007] S1: Preprocess the client trajectory data and construct trajectory clusters. ;
[0008] S2: Privacy sensitivity assessment of trajectory clusters; for each trajectory cluster constructed by the client. The client calculates privacy sensitivity from semantic feature dimensions, spatiotemporal feature dimensions, and individual identifier correlation dimensions, and generates a comprehensive privacy sensitivity score for trajectory clusters. ;Including the following steps:
[0009] S2.1: Extract trajectory clusters Calculate the semantic feature dimension sensitivity of the trajectory points or dwell points;
[0010] S2.2: Calculate the temporal concentration and output the trajectory cluster. The spatiotemporal sensitivity value;
[0011] S2.3: A function that measures the degree of correlation between trajectory clusters and individual identifier information. Calculate the sensitivity of the individual identifier association dimension;
[0012] S2.4: Map the comprehensive score to discrete sensitivity levels based on the threshold function, and classify the comprehensive privacy sensitivity score and stratification into levels;
[0013] S3: Federated learning model training and trajectory cluster contribution calculation;
[0014] S4: Generation of differentiated forgetting strategies based on privacy sensitivity and contribution;
[0015] S5: Federated learning model differential forgetting execution.
[0016] Furthermore, the specific operation method of step S1 is as follows:
[0017] Client The locally collected raw trajectory data is cleaned, denoised, interpolated, and standardized to obtain formatted trajectory data.
[0018] ;
[0019] in, For the first The latitude and longitude coordinates of each trajectory point For the corresponding timestamp, For the client The total number of trajectory points.
[0020] Furthermore, to reduce trajectory data noise and enhance semantic consistency, the client can perform a stop point extraction operation: when the trajectory point is within the spatial radius... Duration exceeding the threshold At that time, the trajectory segment is aggregated into a set of stopping points. The client uses a preset trajectory similarity measurement function. Cluster analysis is performed on the trajectories or stop points to aggregate similar trajectories into multiple trajectory clusters:
[0021] ;
[0022] in, The number of trajectory clusters, each trajectory cluster This represents a set of trajectories whose spatial distribution and temporal patterns are similar, typically corresponding to "family location clusters", "work location clusters", "shopping clusters", "commuting route clusters", etc.
[0023] Furthermore, the specific operation steps of step S2.1 include:
[0024] Extract trajectory clusters The trajectory points or dwell points obtained in the data are mapped to obtain a set of POI types, denoted as:
[0025] ;
[0026] In the above formula, P represents the trajectory cluster. The set of interest point types obtained by mapping the trajectory points or dwell points. Let P be the set of POI types, j = 1, 2, ..., |P|; representing the j-th POI type in set P, such as category labels like hospital, residential area, school, shopping mall, etc.; |P| represents the number or cardinality of POI types in set P.
[0027] make Representing a cluster of trajectories Among the POI types The total number of POI occurrences within a trajectory cluster is defined as follows:
[0028] ;
[0029] In the above formula, This represents the j-th type of POI in set P; To count the number of times this type appears in this cluster; This is the sum of the occurrences of all POI types within the cluster;
[0030] Normalizing by the sum of the occurrence counts of all POI types within the cluster, we get:
[0031] ;
[0032] in, According to the preset sensitive POI weight table;
[0033] Assign sensitive weights to each POI type Combined with the corresponding sensitivity weights in the preset sensitive POI weight table The semantic sensitivity value is obtained by summing the "weight × proportion" of each type:
[0034] ;
[0035] in, The semantic sensitivity value for this cluster; weight. The larger the value, the more sensitive the POI type (e.g., hospital, home address, etc.).
[0036] Furthermore, the specific operation steps of step S2.2 include:
[0037] To characterize whether trajectory clusters possess stable temporal patterns and spatial concentration, temporal concentration and spatial concentration are calculated separately; the access time of trajectory clusters is divided into several time interval sets. For each time slice Statistical Trajectory Clusters Number of trajectory points appearing within this time slice Let the number of trajectory points Divide by the total number of times the cluster appears in all time slices. The time distribution probability is obtained as follows:
[0038] ;
[0039] Then, summing the squared probabilities of each time slice yields the time concentration:
[0040] ;
[0041] The final output is the temporal concentration of the cluster. ; The larger the value, the more concentrated the trajectory cluster is within a specific time period (e.g., nighttime or weekdays);
[0042] Spatial concentration is calculated using the spatial radius of the trajectory cluster. Let the coordinates of the center of the trajectory cluster be... Then the average radius of space is:
[0043] ;
[0044] In the above formula, For trajectory clusters The average spatial radius; Indicates the number of samples or points within a cluster; (x, y) represents the coordinates of a point within the cluster; (x, y) represents the coordinates of the cluster center; the radius is calculated as the average distance from the point to the center.
[0045] The spatial radius is normalized into a concentration index, and a preset maximum radius is introduced. As a normalization metric, the following is calculated:
[0046] ;
[0047] The final output is spatial concentration; where The larger the value, the more concentrated the trajectory cluster is in a fixed location (such as a residential area or workplace), and the higher the privacy risk.
[0048] The spatiotemporal sensitivity value is obtained by weighted fusion of temporal and spatial concentration.
[0049] ;
[0050] in, It is time concentration. It is spatial concentration. and Weight and satisfy .
[0051] Furthermore, the specific operation steps of step S2.3 include:
[0052] To measure whether a trajectory cluster can be used to identify individuals, the individual identifier association sensitivity value for the trajectory cluster is defined as:
[0053] ;
[0054] in, It is the final output of the individual identifier associated with the cluster's sensitivity value; It is the k-th trajectory cluster of the i-th client; This is a correlation evaluation function used to measure the degree of correlation between trajectory clusters and individual identification information. It can be calculated based on the uniqueness of the trajectory cluster, the probability of re-identification, or the degree of matching with external identity information, and satisfies the following: A higher value indicates that it is easier to identify individual users and that the sensitivity is higher.
[0055] Furthermore, the comprehensive privacy sensitivity score described in step S2.4 is obtained by weighting three parts, and is defined as follows:
[0056] ;
[0057] ;
[0058] in, For semantically sensitive values, It is a time-space sensitive value. To identify associated sensitive values, , , There are three weights;
[0059] The comprehensive score is mapped to discrete sensitivity levels based on a threshold function:
[0060] , ;
[0061] Where 1, 2, and 3 represent low-sensitivity, medium-sensitivity, and high-sensitivity trajectory clusters, respectively, to achieve subsequent hierarchical forgetting strategy control.
[0062] Furthermore, if the trajectory cluster in step S2 contains a set of highly sensitive semantic tags... When high-sensitivity semantic tag set There exists a certain type And the number of times it appears within the cluster When, then the indicator function The semantic sensitivity value increases as follows:
[0063] ;
[0064] in, The penalty coefficient is... For indicator functions, For POI types in the high-sensitivity set, Given the number of occurrences, the final output is the value after adding the penalty term. .
[0065] Furthermore, the specific operation steps of step S3 include:
[0066] In the federated learning framework, the client downloads the global model parameters during the t-th training round. The local model parameters are obtained by training based on local trajectory cluster data.
[0067] ;
[0068] in, These are the global model parameters for round t. These are the local model parameters trained on the client side. It is the local learning rate. It is the client-side loss function. For parameters The gradient is ultimately output as the local model parameters on the client side. .
[0069] Client-side calculation of trajectory cluster contribution:
[0070] ;
[0071] The resulting training batch set; This represents one of the training batches; Indicates in global parameters Next, batch The gradient of the loss function obtained above; Represents the gradient norm; Contribution to the final output trajectory cluster.
[0072] Furthermore, the contribution of the trajectory cluster in step S3 can also be calculated using an influence function:
[0073] ;
[0074] In the form of the influence function Represents the trajectory cluster of client i The contribution; among which Represents the loss function. For the loss function enclosed in parentheses, the values are taken within the parentheses. Indicates the current model parameters. This represents the local dataset of client i, therefore This indicates that when the model parameters are Time-use dataset The calculated loss value; where " "Used to indicate that the dataset is included in the corresponding loss calculation" It belongs to the trajectory cluster The loss term corresponding to the loss calculation performed after removing data samples; symbol This indicates taking the absolute value of the difference between two items.
[0075] Furthermore, the generation of the differentiated forgetting strategy in step S4 includes the calculation of forgetting trigger conditions and forgetting intensity coefficients, specifically as follows:
[0076] S4.1: The indicator function for the forgetting trigger condition is:
[0077] ;
[0078] In the above formula, Representing a cluster of trajectories Whether to trigger the forgetting indicator result, 0 or 1; ( ) is an indicator function. For data retention duration, The retention threshold, To forget the number of requests, The threshold for the number of requests, symbol This represents a logical "OR" operation; based on the above formula, the final output is an indication of whether forgetting has been triggered. ;
[0079] The monotonic relationship between threshold and sensitivity indicates that when the overall sensitivity score... When increased, retention time threshold The threshold for the number of requests should be reduced. It should also be reduced, ultimately resulting in the constraint that "the threshold decreases as sensitivity increases";
[0080] S4.2: The Forgetting Intensity Coefficient (DUR) is based on the comprehensive privacy sensitivity score. Contribution of trajectory clusters Generate a differential forgetting intensity coefficient for each trajectory cluster:
[0081] ;
[0082] in: Representing a cluster of trajectories The forgetting intensity coefficient; Sensitivity level (1 / 2 / 3); The sensitivity level mapping function satisfies a monotonically increasing relationship: The monotonicity constraint of the mapping function is: Sensitivity rating Changes satisfy In other words, the higher the sensitivity score, The larger the output, the more you will get in the end. The constraint that "increases with increasing sensitivity"; Output of the sensitivity level mapping function; This is the contribution suppression coefficient, used to avoid excessive forgetting of high-contribution trajectory clusters, which could lead to a significant decrease in model performance. As a contribution level.
[0083] Furthermore, the forgetting triggering condition mentioned in step S4.1 includes data retention time and satisfies the following negative correlation: the negative correlation between the data retention threshold and sensitivity is expressed as follows: ;in, Retention duration threshold, Based on the maximum threshold, For coefficients, The final output, based on a comprehensive sensitivity score, is a retention threshold related to sensitivity.
[0084] Furthermore, the operation method for performing differentiated forgetting in step S5 is as follows:
[0085] When the trajectory cluster is satisfied When the forgetting condition is triggered, the client performs a cluster-level forgetting update on the local model;
[0086] Define the trajectory cluster loss function as follows:
[0087] ;
[0088] in, These are model parameters; For trajectory cluster data, The loss function is used to determine the final result, which is "only the clusters..." Defined cluster-level loss function ;
[0089] The gradient for forgetting the trajectory cluster is the reverse gradient:
[0090] ;
[0091] Representing a cluster of trajectories The forgetting gradient is updated so that the final output is the forgetting gradient of the cluster; This is the forgetting intensity coefficient; This represents the gradient of the cluster-level loss with respect to the parameters; the negative sign indicates a reverse update.
[0092] The client forgetting update amount is obtained by summing the forgetting update gradients for all trajectory clusters on the client side. ;
[0093] This represents the total number of forgotten updates for client i, and the final output is the number of forgotten updates uploaded by the client. The number of trajectory clusters for the client; The forgetting gradients for each cluster are summed, and the final output is the amount of forgetting updates uploaded by the client. ;
[0094] The client uploads the forgotten updates to the server, and the server processes the set of clients that initiated the forgotten requests. Perform an aggregate update to obtain the forgotten global model:
[0095] ;
[0096] in, This represents the global model after forgetting, and the final output is the parameters of the global model after forgetting. Represents the global model before forgetting; This represents the set of clients that initiated the forget request; This indicates the number of forgotten updates uploaded by client i; The client aggregates the weights, and the final output is the global model after forgetting. ;
[0097] Higher sensitivity levels correspond to larger trajectory clusters. This allows for stronger reverse updates, prioritizing the forgetting of highly privacy-sensitive trajectories; while the forgetting intensity of trajectory clusters with lower sensitivity levels is weaker, thus avoiding excessive degradation of model utility and achieving a balance between privacy protection and model performance.
[0098] Furthermore, the server-side uses FedAvg to aggregate forgotten update amounts, and the operation method is as follows: by Aggregated updates;
[0099] in, Represents the global model after forgetting. Represents the global model before forgetting. For the set of clients that initiated the forget request, This represents the amount of local data on client i. The client forgets the update amount, and the final output is the global model after forgetting. .
[0100] Beneficial effects
[0101] The present invention proposes a federated learning differential forgetting method based on trajectory cluster privacy sensitivity, which has the following advantages compared with existing technologies:
[0102] (1) This invention clusters trajectory data into trajectory clusters, evaluates the privacy sensitivity of trajectory clusters from multiple dimensions, and formulates differentiated forgetting strategies based on their contribution to the model, thereby achieving accurate and efficient forgetting of trajectory data with high privacy sensitivity.
[0103] (2) This invention achieves both cluster-level controllability and differentiated adjustability of the "forgotten object" and the "forgotten intensity"; by clustering trajectories with similar spatiotemporal patterns and semantic attributes, trajectories are grouped into trajectory clusters, so that each type of sensitive behavior pattern (such as home location, medical trajectory, etc.) can be stably mapped to a specific cluster, thereby refining the forgetting operation from the granularity of "all data / all model" to the granularity of "target cluster", naturally realizing "only acting on the part of data that needs to be forgotten", avoiding irrelevant data being mistakenly affected, and reflecting the precision of forgetting. Different forgetting strategies are adopted for trajectory clusters with different privacy sensitivities, which can focus on protecting highly sensitive trajectory data, effectively reduce the risk of privacy leakage, meet users' right to forget their personal sensitive trajectory data, and achieve precise privacy protection.
[0104] (3) This invention calculates the privacy sensitivity of each trajectory cluster from the dimensions of semantics, spatiotemporal and individual identifier correlation and classifies them into different levels, so that high-sensitivity clusters can obtain higher forgetting priority and stronger forgetting intensity at the strategy level. This ensures that forgetting resources are concentrated on the clusters with the greatest privacy risks and improves the targeting of forgetting effect. Furthermore, increasing the forgetting intensity only according to sensitivity may lead to a significant decrease in model utility. Therefore, this invention introduces a trajectory cluster contribution measure to characterize the influence of the cluster on the update of model parameters. In the forgetting intensity coefficient, the contribution suppression term is used to automatically "cool down" high-contribution clusters. This avoids performance collapse caused by excessive reverse update of the clusters corresponding to key decision features without weakening the priority forgetting principle of high-sensitivity clusters, and achieves a dynamic balance between privacy protection and model performance.
[0105] (4) At the execution level, the present invention calculates the reverse forgetting gradient based on the cluster-level loss, and uploads the forgetting update amount of each cluster to the server after the client is aggregated and updated to obtain the global model after forgetting. This allows the model parameters to be directionally corrected in the direction of "eliminating the influence of the target cluster". The influence of a specific cluster can be quickly removed without full retraining, thus achieving efficient forgetting.
[0106] (5) This invention considers the contribution of trajectory clusters to the model. When performing the forgetting operation, the forgetting intensity of trajectory clusters with high contribution is appropriately reduced. This can minimize the negative impact on the overall performance of the federated learning model while protecting privacy, thus balancing privacy protection and model usability. Differentiated processing is achieved to improve model usability. Attached Figure Description
[0107] Figure 1 This is a schematic diagram of the overall process of the present invention.
[0108] Figure 2 This is a schematic diagram of the overall architecture of the experimental environment constructed in the experiment of this invention.
[0109] Figure 3This is a schematic diagram illustrating the change in global accuracy with each round in the experiment of this invention.
[0110] Figure 4 This is a schematic diagram illustrating the change in the accuracy of non-target clusters with each round in the experiment of this invention. Detailed Implementation
[0111] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments. The described embodiments are merely some embodiments of the present invention, and not all embodiments. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the design concept of the present invention should fall within the protection scope of the present invention.
[0112] Example 1:
[0113] like Figure 1 As shown, a federated learning differential forgetting method based on trajectory cluster privacy sensitivity includes the following steps:
[0114] S1: Preprocess the client trajectory data and construct trajectory clusters. The specific operation method is as follows:
[0115] Client The locally collected raw trajectory data is cleaned, denoised, interpolated, and standardized to obtain formatted trajectory data.
[0116] ;
[0117] in, For the first The latitude and longitude coordinates of each trajectory point For the corresponding timestamp, For the client The total number of trajectory points.
[0118] To reduce noise in trajectory data and enhance semantic consistency, the client can perform a stop point extraction operation: when the trajectory point is within the spatial radius... Duration exceeding the threshold At that time, the trajectory segment is aggregated into a set of stopping points. The client uses a preset trajectory similarity measurement function. Cluster analysis is performed on the trajectories or stop points to aggregate similar trajectories into multiple trajectory clusters:
[0119] ;
[0120] in, The number of trajectory clusters, each trajectory cluster This represents a set of trajectories whose spatial distribution and temporal patterns are similar, typically corresponding to "family location clusters", "work location clusters", "shopping clusters", "commuting route clusters", etc.
[0121] S2: Privacy sensitivity assessment of trajectory clusters; for each trajectory cluster constructed by the client. The client calculates privacy sensitivity from semantic feature dimensions, spatiotemporal feature dimensions, and individual identifier correlation dimensions, and generates a comprehensive privacy sensitivity score for trajectory clusters. The steps include:
[0122] S2.1: Extract trajectory clusters Calculate the semantic feature dimension sensitivity of the trajectory points or dwell points;
[0123] Extract trajectory clusters The trajectory points or dwell points obtained in the data are mapped to obtain a set of POI types, denoted as:
[0124] ;
[0125] In the above formula, P represents the trajectory cluster. The set of interest point types obtained by mapping the trajectory points or dwell points. Let P be the set of POI types, j=1,2,…,|P|; representing the j-th POI type in set P, such as category labels like hospital, residential area, school, shopping mall, etc.; |P| represents the number or cardinality of POI types in set P.
[0126] make Representing a cluster of trajectories Among the POI types The total number of POI occurrences within a trajectory cluster is defined as follows:
[0127] ;
[0128] In the above formula, This represents the j-th type of POI in set P; To count the number of times this type appears in this cluster; This is the sum of the occurrences of all POI types within the cluster;
[0129] Normalizing by the sum of the occurrence counts of all POI types within the cluster, we get:
[0130] ;
[0131] in, According to the preset sensitive POI weight table;
[0132] Assign sensitive weights to each POI type Combined with the corresponding sensitivity weights in the preset sensitive POI weight table The semantic sensitivity value is obtained by summing the "weight × proportion" of each type:
[0133] ;
[0134] in, The semantic sensitivity value for this cluster; weight. The larger the value, the more sensitive the POI type (e.g., hospital, home address, etc.).
[0135] If the trajectory cluster contains a set of highly sensitive semantic tags When high-sensitivity semantic tag set There exists a certain type And the number of times it appears within the cluster When, then the indicator function The semantic sensitivity value increases as follows:
[0136] ;
[0137] in, The penalty coefficient is... For indicator functions, For POI types in the high-sensitivity set, Given the number of occurrences, the final output is the value after adding the penalty term. .
[0138] S2.2: Calculate the temporal concentration and output the trajectory cluster. The spatiotemporal sensitivity value;
[0139] To characterize whether trajectory clusters possess stable temporal patterns and spatial concentration, temporal concentration and spatial concentration are calculated separately; the access time of trajectory clusters is divided into several time interval sets. For each time slice Statistical Trajectory Clusters Number of trajectory points appearing within this time slice Let the number of trajectory points Divide by the total number of times the cluster appears in all time slices. The time distribution probability is obtained as follows:
[0140] ;
[0141] Then, summing the squared probabilities of each time slice yields the time concentration:
[0142] ;
[0143] The final output is the temporal concentration of the cluster. ; The larger the value, the more concentrated the trajectory cluster is in a specific time period (e.g., nighttime or weekdays).
[0144] Spatial concentration is calculated using the spatial radius of the trajectory cluster. Let the coordinates of the center of the trajectory cluster be... Then the average radius of space is:
[0145] ;
[0146] In the above formula, For trajectory clusters The average spatial radius; Indicates the number of samples or points within a cluster; (x,y) represents the coordinates of a point within the cluster; (x,y) represents the coordinates of the cluster center; the radius is calculated as the average distance from the point to the center.
[0147] The spatial radius is normalized into a concentration index, and a preset maximum radius is introduced. As a normalization metric, the following is calculated:
[0148] ;
[0149] The final output is spatial concentration; where The larger the value, the more concentrated the trajectory cluster is in a fixed location (such as a residential area or workplace), and the higher the privacy risk.
[0150] The spatiotemporal sensitivity value is obtained by weighted fusion of temporal and spatial concentration.
[0151] ;
[0152] in, It is time concentration. It is spatial concentration. and Weight and satisfy .
[0153] S2.3: A function that measures the degree of correlation between trajectory clusters and individual identifier information. Calculate the sensitivity of the individual identifier association dimension;
[0154] To measure whether a trajectory cluster can be used to identify individuals, the individual identifier association sensitivity value for the trajectory cluster is defined as:
[0155] ;
[0156] in, It is the final output of the individual identifier associated with the cluster's sensitivity value; It is the k-th trajectory cluster of the i-th client; This is a correlation evaluation function used to measure the degree of correlation between trajectory clusters and individual identification information. It can be calculated based on the uniqueness of the trajectory cluster, the probability of re-identification, or the degree of matching with external identity information, and satisfies the following: A higher value indicates that it is easier to identify individual users and that the sensitivity is higher.
[0157] constraint ∈[0,1] indicates that the value of the individual identifier associated with the sensitivity value is between 0 and 1, and the final result is the range constraint that "the value is limited to the interval between 0 and 1".
[0158] S2.4: Map the comprehensive score to discrete sensitivity levels according to the threshold function, and classify the comprehensive privacy sensitivity score and stratification; the comprehensive privacy sensitivity score mentioned in step S2.4 is obtained by weighting three parts, and is defined as:
[0159] ;
[0160] ;
[0161] in, For semantically sensitive values, It is a time-space sensitive value. To identify associated sensitive values, , , There are three weights;
[0162] The comprehensive score is mapped to discrete sensitivity levels based on a threshold function:
[0163] , ;
[0164] Where 1, 2, and 3 represent low-sensitivity, medium-sensitivity, and high-sensitivity trajectory clusters, respectively, to achieve subsequent hierarchical forgetting strategy control.
[0165] Sensitivity level middle, It is a threshold mapping function. It is a comprehensive score. This indicates low / medium / high sensitivity levels, and the final output is a level. .
[0166] S3: Federated learning model training and trajectory cluster contribution calculation;
[0167] In the federated learning framework, the server downloads the global model parameters to the client during the t-th training round. The client trains local model parameters based on local trajectory cluster data; local training updates:
[0168] ;
[0169] in, These are the global model parameters for round t. These are the local model parameters trained on the client side. It is the local learning rate. It is the client-side loss function. For parameters The gradient is ultimately output as the local model parameters on the client side. .
[0170] Client-side calculation of trajectory cluster contribution:
[0171] ;
[0172] The resulting training batch set; This represents one of the training batches; Indicates in global parameters Next, batch The gradient of the loss function obtained above; Represents the gradient norm; Contribution to the final output trajectory cluster.
[0173] The contribution of trajectory clusters can also be calculated using an influence function:
[0174] ;
[0175] In the form of the influence function Represents the trajectory cluster of client i The contribution; among which Represents the loss function. For the loss function enclosed in parentheses, the values are taken within the parentheses. Indicates the current model parameters. This represents the local dataset of client i, therefore This indicates that when the model parameters are Time-use dataset The calculated loss value; where " "Used to indicate that the dataset is included in the corresponding loss calculation" It belongs to the trajectory cluster The loss term corresponding to the loss calculation performed after removing data samples; symbol This indicates taking the absolute value of the difference between two items.
[0176] S4: Generation of differentiated forgetting strategies based on privacy sensitivity and contribution; the generation of differentiated forgetting strategies includes the calculation of forgetting trigger conditions and forgetting intensity coefficients, specifically:
[0177] S4.1: The indicator function for the forgetting trigger condition is:
[0178] ( ) ( > ) ;
[0179] In the above formula, ( ) represents a trajectory cluster Whether to trigger the forgetting indicator result, 0 or 1; ( ) is an indicator function. For data retention duration, The retention threshold, To forget the number of requests, The threshold for the number of requests, symbol This represents a logical "OR" operation; based on the above formula, the final output is an indication of whether forgetting has been triggered. ( );
[0180] The monotonic relationship between threshold and sensitivity indicates that when the overall sensitivity score... When increased, retention time threshold The threshold for the number of requests should be reduced. It should also be reduced, ultimately resulting in the constraint that "the threshold decreases as sensitivity increases";
[0181] Forgetting triggers include data retention time and satisfy the following negative correlation: The negative correlation between data retention threshold and sensitivity is expressed as follows: ;in, Retention duration threshold, Based on the maximum threshold, For coefficients, For comprehensive sensitivity scoring, according to The final output is the retention threshold related to sensitivity. .
[0182] S4.2: The Forgetting Intensity Coefficient (DUR) is based on the comprehensive privacy sensitivity score. Contribution of trajectory clusters Generate a differential forgetting intensity coefficient for each trajectory cluster:
[0183] ;
[0184] in: Representing a cluster of trajectories The forgetting intensity coefficient; Sensitivity level (1 / 2 / 3); The sensitivity level mapping function satisfies a monotonically increasing relationship: The monotonicity constraint of the mapping function is: Sensitivity rating Changes satisfy In other words, the higher the sensitivity score, The larger the output, the more you will get in the end. The constraint that "increases with increasing sensitivity"; Output of the sensitivity level mapping function; This is the contribution suppression coefficient, used to avoid excessive forgetting of high-contribution trajectory clusters, which could lead to a significant decrease in model performance. As a contribution level.
[0185] Higher sensitivity levels correspond to larger trajectory clusters. This allows for stronger reverse updates, prioritizing the forgetting of highly privacy-sensitive trajectories; while the forgetting intensity of trajectory clusters with lower sensitivity levels is weaker, thus avoiding excessive degradation of model utility and achieving a balance between privacy protection and model performance.
[0186] S5: Differentiated forgetting execution in federated learning models; when the trajectory cluster meets the forgetting trigger condition. ( ) At time 1, the client performs a differential forgetting update on the local model.
[0187] Define the trajectory cluster loss function as follows:
[0188] ;
[0189] in, These are model parameters; For trajectory cluster data, The loss function is used to determine the final result, which is "only the clusters..." Defined cluster-level loss function .
[0190] The gradient for forgetting the trajectory cluster is the reverse gradient:
[0191] ;
[0192] Representing a cluster of trajectories The forgetting gradient is updated so that the final output is the forgetting gradient of the cluster; This is the forgetting intensity coefficient; The gradient of the cluster-level loss with respect to the parameters; the negative sign indicates reverse update.
[0193] The client forgetting update amount is obtained by summing the forgetting update gradients for all trajectory clusters on the client side.
[0194] ;
[0195] in, This represents the total number of forgotten updates for client i, and the final output is the number of forgotten updates uploaded by the client. The number of trajectory clusters for the client; The forgetting gradients for each cluster are summed, and the final output is the amount of forgetting updates uploaded by the client. .
[0196] The client will forget the update amount. Uploaded to the server, the server then processes the set of clients that initiated the forget request. Perform an aggregate update to obtain the forgotten global model:
[0197] ;
[0198] in, This represents the global model after forgetting, and the final output is the parameters of the global model after forgetting. Represents the global model before forgetting; This represents the set of clients that initiated the forget request; This indicates the number of forgotten updates uploaded by client i; The client aggregates the weights, and the final output is the global model after forgetting. .
[0199] FedAvg weights: ;
[0200] in, This represents the amount of local data on client i. It requests the total amount of data within the set, and the final output is the aggregate weight. .
[0201] The server uses FedAvg to aggregate forgotten updates, and the operation method is as follows: [Instructions for operation are missing from the original text]. Aggregated updates;
[0202] in, Represents the global model after forgetting. Represents the global model before forgetting. For the set of clients that initiated the forget request, This represents the amount of local data on client i. The client forgets the update amount, and the final output is the global model after forgetting. .
[0203] S6: Distribution of the global model after forgetting and subsequent training: The server distributes the forgotten global model parameters... Send to the client, the client uses As an initial model for subsequent federated training or continuous learning, it enters the next round of training, thereby effectively removing the impact of privacy-sensitive trajectory clusters without performing full retraining.
[0204] At this point, all the specific implementation steps of this embodiment have been completed.
[0205] To verify the feasibility and performance impact of the proposed "differentiated forgetting method based on trajectory cluster privacy sensitivity and contribution constraints" under the federated learning framework, a server-side-multi-client federated training simulation environment was constructed.
[0206] like Figure 2 As shown, this experiment constructs a federated learning differential forgetting system based on trajectory cluster privacy sensitivity. Its specific architecture includes:
[0207] 1. Clients: Assume the total number of clients participating in federated learning is... The client set is represented as:
[0208] ;
[0209] Each client Holding local trajectory dataset The client is typically a user terminal device that holds sensitive trajectory data, such as a mobile phone, vehicle terminal, or smartwatch.
[0210] The client includes the following functional modules: (1) trajectory data preprocessing and trajectory cluster construction module; (2) trajectory cluster privacy sensitivity assessment module; (3) federated learning training and trajectory cluster contribution calculation module; (4) differentiated forgetting strategy generation module; (5) differentiated forgetting execution module.
[0211] 2. Server: The server is responsible for receiving training updates and forgetting updates from the client and performing aggregated updates to obtain global model parameters. The server is usually deployed by the model owner or service provider and includes the following functional modules: (1) Global model maintenance module; (2) Forgetting update aggregation module; (3) Forgetting request management module.
[0212] By constructing trajectory clusters on the client side and stratifying the privacy sensitivity of these clusters, and combining contribution estimation to generate differentiated forgetting intensity, we can achieve priority forgetting of highly privacy-sensitive trajectory clusters while maintaining the overall performance of the federated model. This solves the problem of "over-forgetting or under-forgetting caused by uniform forgetting intensity" in existing federated forgetting schemes.
[0213] The server is responsible for maintaining global model parameters and performing routine aggregation updates to the global model. When the preset forgetting round is reached and a forgetting request is triggered, the client generates a differentiated forgetting intensity coefficient based on the trajectory cluster privacy sensitivity level and trajectory cluster contribution, and performs cluster-level reverse gradient forgetting update to obtain the client forgetting update amount. The data is then uploaded to the server, where a weighted aggregation update is performed on the set of clients that initiated the forgetting request to obtain the forgotten global model. They were then distributed for use in subsequent training rounds.
[0214] In the simulation, the 20th round was set as the forgetting trigger point to observe the instantaneous changes in the model's performance before and after the forgetting trigger and the subsequent recovery process.
[0215] Comparison scheme settings:
[0216] This experiment set up four schemes for comparison:
[0217] (1) baseline: Do not perform forgetting;
[0218] (2) uniform: Apply a uniform forgetting intensity to the target cluster;
[0219] (3) sens_only: The intensity of forgetting is determined solely by the sensitivity level of the trajectory cluster;
[0220] (4) Proposed: The forgetting intensity is determined by a monotonically increasing mapping of sensitivity level and a contribution inhibition term, so as to constrain excessive forgetting of high-contribution information.
[0221] The comparative data of the four schemes are shown in Table 1:
[0222] Table 1 Summary of changes before and after the forgetting trigger point
[0223] setting round_pre round_post delta_global_acc delta_high_acc delta_non_acc delta_global_loss delta_high_loss delta_non_loss baseline 19 20 0.007 0.058 0.003 −0.009 −0.026 −0.007 proposed 19 20 −0.096 0.052 −0.109 0.211 −0.128 0.240 sens_only 19 20 −0.169 −0.152 −0.170 0.906 0.971 0.901 uniform 19 20 −0.064 0.094 −0.078 0.044 −0.174 0.063
[0224] Based on the difference data from rounds 19 to 20 (forgetting trigger point) in unlearning_impact_summary.csv, the following conclusions can be drawn: the baseline remains basically stable at the forgetting trigger point, global_acc increases slightly (+0.007), non_acc also increases slightly (+0.003), and global_loss decreases slightly (−0.009), indicating that there is no significant performance perturbation in this round during normal federated training without forgetting.
[0225] When forgetting is triggered, all three forgetting strategies introduce varying degrees of transient impact, with the sens_only strategy exhibiting the most significant impact. Global_acc drops sharply (−0.169), and non_acc also drops sharply (−0.170), while global_loss and non_loss increase significantly (+0.906 and +0.901, respectively). This indicates that setting the forgetting intensity solely based on the sensitivity level can easily cause significant collateral effects on non-target data and significantly weaken overall performance.
[0226] The impact of uniformity is relatively small, with global_acc decreasing (−0.064) and non_acc decreasing (−0.078), while the corresponding increases in global_loss and non_loss are also relatively mild (+0.044, +0.063), reflecting that the uniform forgetting intensity has a limited impact on the perturbation of the whole and non-target.
[0227] At the trigger point, the global_acc and non_acc of the proposed algorithm decreased by (−0.096, −0.109), while its global_loss and non_loss increased by (+0.211, +0.240). The overall perturbation amplitude was significantly smaller than that of sens_only, but larger than that of uniform. This indicates that under the current parameter settings, introducing a contribution suppression term can, to some extent, constrain the global and non-target performance impact caused by forgetting updates, making it more controllable than "sensitive forgetting only". Meanwhile, looking at the high-sensitivity cluster (target cluster) metrics, both high_acc and high_loss of sens_only showed significant changes (−0.152, +0.971), while high_acc of both proposed and uniform showed positive changes (+0.052, +0.094, respectively) and negative changes (−0.128, −0.174, respectively) at this trigger point. This suggests that the directionality of accuracy / loss alone cannot adequately characterize "whether forgetting has occurred," and a combination of... Figure 3 / Figure 4 The overall trend of the round curves and the loss curves of the target cluster / non-target clusters is used to comprehensively present the perturbation caused by forgetting and the subsequent recovery process. Based on this, the forgetting update step size and inhibition coefficient are further adjusted to enhance the differentiated effect of "more concentrated effect on the target cluster and less impact on non-target clusters".
[0228] Evaluation indicators and data division:
[0229] To characterize the impact of forgetting triggers on overall performance and non-target data, a global test set was constructed, and the test samples were divided into a high-sensitivity cluster (target cluster) and a non-target cluster (sensitivity ≠ 3). The results were recorded and plotted.
[0230] (1) Global Accuracy: Global test set accuracy (see...) Figure 3 );
[0231] (2) Non-target Accuracy: Accuracy on the test set for non-target clusters (see...) Figure 4 );
[0232] (3) At the same time, the changes in indicators before and after the forgetting trigger point (round 19 → round 20) are statistically summarized (see unlearning_impact_summary.csv) to quantify the instantaneous disturbance caused by the forgetting trigger.
[0233] Figure 3 The graphs show the global test set accuracy of four schemes during federated training as a function of training epochs (1-30). Before the forgetting trigger point (epochs 1-19), the accuracy of each scheme generally increases with each epoch, reflecting the convergence process of normal federated training; the baseline, serving as a control without forgetting, shows a relatively smooth curve. After the forgetting operation is triggered at epoch 20 (marked by the dashed line), the schemes that perform forgetting (uniform, sens_only, proposed) experience a one-time drop in accuracy, indicating that forgetting updates have an immediate perturbation on global performance; subsequently, as federated training continues, the accuracy of each scheme gradually recovers and tends to stabilize, demonstrating that the model can continue to iterate and recover performance after forgetting.
[0234] Figure 4 The diagram shows the accuracy curves of four schemes on non-target clusters (sensitivity ≠ 3, i.e., low / medium sensitivity data) as a function of training epochs, used to characterize the impact of forgetting operations on non-target data. Before the forgetting trigger point (epochs 1-19), the accuracy of non-target clusters generally increases with each epoch, consistent with the normal federated training process. After forgetting is triggered in epoch 20, different forgetting strategies show varying degrees of decrease in accuracy on non-target clusters, reflecting the differences in the impact of each method on the performance of non-target data. Subsequently, as training continues, the accuracy of non-target clusters gradually recovers and tends to stabilize, which can be used to illustrate that the model still has the ability to continue training and recover performance after performing forgetting updates.
[0235] Experimental Results and Analysis:
[0236] like Figure 3As shown, before the forgetting trigger point (rounds 1-19), the global accuracy of each scheme generally increases with each training round, reflecting the convergence trend of normal federated training. After the forgetting is triggered in round 20, all forgetting schemes except for the baseline scheme experience a one-time drop in accuracy, indicating that forgetting updates introduce immediate perturbations. However, the accuracy gradually recovers and stabilizes during subsequent training, demonstrating that iterative training can continue after forgetting without requiring a full retraining. To further quantify the impact of the trigger point, the changes from round 19 to 20 are statistically analyzed using unlearning_impact_summary.csv: baseline remains relatively stable (Δglobal_acc = +0.0071); uniform decreases (Δglobal_acc = −0.0642); sens_only shows the largest decrease (Δglobal_acc = −0.1688); and proposed decreases less than sens_only (Δglobal_acc = −0.0958), indicating that the global performance perturbation of the proposed scheme is relatively more controllable when forgetting is triggered.
[0237] like Figure 4 As shown, the accuracy of non-target clusters (sensitivity ≠ 3) reflects the degree of influence of forgetting on non-target data. When forgetting was triggered in the 20th round, each forgetting scheme showed varying degrees of decrease in accuracy for non-target clusters. The difference table statistics show that the baseline remained relatively stable (Δnon_acc = +0.0027), uniform (Δnon_acc = −0.0779), sens_only (Δnon_acc = −0.1702), and proposed (Δnon_acc = −0.1086). This result indicates that sens_only, which sets the forgetting intensity solely based on sensitivity, has a more significant impact on non-target data. In contrast, the proposed scheme of this invention, by introducing a contribution suppression term, can constrain the performance impact on non-target data to a certain extent while performing forgetting updates, reflecting the balance between privacy protection and model utility preservation in differentiated forgetting.
[0238] Experiment Summary:
[0239] comprehensive Figure 3 , Figure 4The statistical results of the trigger point difference show that forgetting triggers introduce short-term performance disturbances, but the model can continue to train and gradually recover in subsequent rounds. In the comparison of trigger point quantification, the proposed solution of this invention shows a smaller performance impact on the global and non-target clusters compared with the sensitivity-only forgetting strategy (sens_only). This indicates that the differentiated forgetting intensity design with the introduction of the "contribution suppression term" helps to reduce the impact on non-target data and the overall model utility when performing forgetting updates.
Claims
1. A federated learning differential forgetting method based on trajectory cluster privacy sensitivity, characterized in that: Including the following steps: S1: Preprocess the client trajectory data and construct trajectory clusters. ; S2: Privacy sensitivity assessment of trajectory clusters; for each trajectory cluster constructed by the client. The client calculates privacy sensitivity from semantic feature dimensions, spatiotemporal feature dimensions, and individual identifier correlation dimensions, and generates a comprehensive privacy sensitivity score for trajectory clusters. ; Including the following steps: S2.1: Extract trajectory clusters Calculate the semantic feature dimension sensitivity of the trajectory points or dwell points; S2.2: Calculate the temporal concentration and output the trajectory cluster. The spatiotemporal sensitivity value; S2.3: A function that measures the degree of correlation between trajectory clusters and individual identifier information. Calculate the sensitivity of the individual identifier association dimension; S2.4: Map the comprehensive score to discrete sensitivity levels based on the threshold function, and classify the comprehensive privacy sensitivity score and stratification into levels; S3: Federated learning model training and trajectory cluster contribution calculation; S4: Generation of differentiated forgetting strategies based on privacy sensitivity and contribution; S5: Federated learning model differential forgetting execution.
2. The differential forgetting method for federated learning based on trajectory cluster privacy sensitivity according to claim 1, characterized in that: The specific operation method of step S1 is as follows: Client The locally collected raw trajectory data is cleaned, denoised, interpolated, and standardized to obtain formatted trajectory data. ； in, For the first The latitude and longitude coordinates of each trajectory point For the corresponding timestamp, For the client The total number of trajectory points.
3. The differential forgetting method for federated learning based on trajectory cluster privacy sensitivity according to claim 2, characterized in that: The client can perform a stop point extraction operation: when the trajectory point is within the spatial radius Duration exceeding the threshold At that time, the trajectory segment is aggregated into a set of stopping points. ; The client uses a preset trajectory similarity measurement function. Cluster analysis is performed on the trajectories or stop points to aggregate similar trajectories into multiple trajectory clusters: ； in, The number of trajectory clusters, each trajectory cluster This represents a set of trajectories whose spatial distribution and temporal patterns are similar.
4. The differential forgetting method for federated learning based on trajectory cluster privacy sensitivity according to claim 1, characterized in that: The specific operation steps of step S2.1 include: Extract trajectory clusters Mapping the trajectory points or dwell points obtained from the data yields a set of POI types, denoted as: ； In the above formula, P represents the trajectory cluster. The set of interest point types obtained by mapping the trajectory points or dwell points. Let P be the set of POI types, j = 1, 2, ..., |P|; representing the j-th POI type in set P, such as category labels like hospital, residential area, school, shopping mall, etc.; |P| represents the number or cardinality of POI types in set P. make Representing a cluster of trajectories Among the POI types The total number of POI occurrences within a trajectory cluster is defined as follows: ； In the above formula, This represents the j-th type of POI in set P; To count the number of times this type appears in this cluster; This is the sum of the occurrences of all POI types within the cluster; Normalizing by the sum of the occurrence counts of all POI types within the cluster, we get: ； in, According to the preset sensitive POI weight table; Assign sensitive weights to each POI type Combined with the corresponding sensitivity weights in the preset sensitive POI weight table The semantic sensitivity value is obtained by summing the "weight × proportion" of each type: ； in, The semantic sensitivity value for this cluster; weight. The larger the value, the more sensitive the POI type.
5. The differential forgetting method for federated learning based on trajectory cluster privacy sensitivity according to claim 1, characterized in that: The specific operation steps of step S2.2 include: The access time of the trajectory cluster is divided into several time period sets. For each time slice Statistical Trajectory Clusters Number of trajectory points appearing within this time slice Let the number of trajectory points Divide by the total number of times the cluster appears in all time slices. The time distribution probability is obtained as follows: ； Then, summing the squared probabilities of each time slice yields the time concentration: ； The final output is the temporal concentration of the cluster. ; The larger the value, the more concentrated the trajectory cluster is within a specific time period; Spatial concentration is calculated using the spatial radius of the trajectory cluster. Let the coordinates of the center of the trajectory cluster be... Then the average radius of space is: ； In the above formula, For trajectory clusters The average spatial radius; Indicates the number of samples or points within a cluster; (x, y) represents the coordinates of a point within the cluster; (x, y) represents the coordinates of the cluster center; the radius is calculated as the average distance from the point to the center. The spatial radius is normalized into a concentration index, and a preset maximum radius is introduced. As a normalization metric, the following is calculated: ； The final output is spatial concentration; where The larger the value, the more concentrated the trajectory cluster is in a fixed location (such as a residential area or workplace), and the higher the privacy risk. The spatiotemporal sensitivity value is obtained by weighted fusion of temporal and spatial concentration. ； in, It is the spatiotemporal sensitivity value of the trajectory cluster. It is time concentration. It is spatial concentration. and Weights and satisfying .
6. The differential forgetting method for federated learning based on trajectory cluster privacy sensitivity according to claim 1, characterized in that: The specific operation steps of step S2.3 include: Define the sensitivity value for individual identifier association of trajectory clusters as: ； in, It is the final output of the individual identifier associated with the cluster's sensitivity value; It is the k-th trajectory cluster of the i-th client; This is a correlation evaluation function used to measure the degree of correlation between trajectory clusters and individual identification information. It can be calculated based on the uniqueness of the trajectory cluster, the probability of re-identification, or the degree of matching with external identity information, and satisfies the following: A higher value indicates that it is easier to identify individual users and that the sensitivity is higher.
7. The differential forgetting method for federated learning based on trajectory cluster privacy sensitivity according to claim 1, characterized in that: The comprehensive privacy sensitivity score mentioned in step S2.4 is obtained by weighting three parts, and is defined as follows: ； ； in, For semantically sensitive values, It is a time-space sensitive value. To identify associated sensitive values, , , There are three weights; The comprehensive score is mapped to discrete sensitivity levels based on a threshold function: ， ； Where 1, 2, and 3 represent low-sensitivity, medium-sensitivity, and high-sensitivity trajectory clusters, respectively, to achieve subsequent hierarchical forgetting strategy control.
8. The differential forgetting method for federated learning based on trajectory cluster privacy sensitivity according to claim 1, characterized in that: The specific steps of step S3 include: In the federated learning framework, the client downloads the global model parameters during the t-th training round. The local model parameters are obtained by training based on local trajectory cluster data. ； in, These are the global model parameters for round t. These are the local model parameters trained on the client side. It is the local learning rate. It is the client-side loss function. For parameters The gradient is ultimately output as the local model parameters on the client side. ; Client-side calculation of trajectory cluster contribution: ； in, This represents the set of training batches generated by the trajectory cluster; This represents one of the training batches; Indicates in global parameters Next, batch The gradient of the loss function obtained above; Represents the gradient norm; Contribution to the final output trajectory cluster.
9. The differential forgetting method for federated learning based on trajectory cluster privacy sensitivity according to claim 8, characterized in that: The generation of the differentiated forgetting strategy in step S4 includes the calculation of forgetting trigger conditions and forgetting intensity coefficients, specifically: S4.1: The indicator function for the forgetting trigger condition is: ； In the above formula, ( ) represents a trajectory cluster Whether to trigger the forgetting indicator result, 0 or 1; ( ) is an indicator function. For data retention duration, The retention threshold, To forget the number of requests, The threshold for the number of requests, symbol This represents a logical "OR" operation; based on the above formula, the final output is an indication of whether forgetting has been triggered. ( ); The monotonic relationship between threshold and sensitivity indicates that when the overall sensitivity score... When increased, retention time threshold The threshold for the number of requests should be reduced. It should also be reduced, ultimately resulting in the constraint that "the threshold decreases as sensitivity increases"; S4.2: The Forgetting Intensity Coefficient (DUR) is based on the comprehensive privacy sensitivity score. Contribution of trajectory clusters Generate a differential forgetting intensity coefficient for each trajectory cluster: ； in: Representing a cluster of trajectories The forgetting intensity coefficient; Sensitivity level (1 / 2 / 3); The sensitivity level mapping function satisfies a monotonically increasing relationship: The monotonicity constraint of the mapping function is: Sensitivity rating Changes satisfy In other words, the higher the sensitivity score, The larger the output, the more you will get in the end. The constraint that "increases with increasing sensitivity"; Output of the sensitivity level mapping function; This is the contribution suppression coefficient, used to avoid excessive forgetting of high-contribution trajectory clusters, which could lead to a significant decrease in model performance. As a contribution level.
10. The differential forgetting method for federated learning based on trajectory cluster privacy sensitivity according to claim 1, characterized in that: The operation method for performing differential forgetting in step S5 is as follows: When the trajectory cluster is satisfied When the forgetting condition is triggered, the client performs a cluster-level forgetting update on the local model; Define the trajectory cluster loss function as follows: ； in, These are model parameters; For trajectory cluster data, The loss function is used to determine the final result, which is "only the clusters..." Defined cluster-level loss function ; The gradient for forgetting the trajectory cluster is the inverse gradient: ； Representing a cluster of trajectories The forgetting gradient is updated so that the final output is the forgetting gradient of the cluster; This is the forgetting intensity coefficient; This represents the gradient of the cluster-level loss with respect to the parameters; the negative sign indicates a reverse update. The client forgetting update amount is obtained by summing the forgetting update gradients for all trajectory clusters on the client side. ; This represents the total number of forgotten updates for client i, and the final output is the number of forgotten updates uploaded by the client. The number of trajectory clusters for the client; The forgetting gradients for each cluster are summed, and the final output is the amount of forgetting updates uploaded by the client. ; The client uploads the forgotten updates to the server, and the server processes the set of clients that initiated the forgotten requests. Perform an aggregate update to obtain the forgotten global model: ； in, This represents the global model after forgetting, and the final output is the parameters of the global model after forgetting. Represents the global model before forgetting; This represents the set of clients that initiated the forget request; This indicates the number of forgotten updates uploaded by client i; The client aggregates the weights, and the final output is the global model after forgetting. ; Higher sensitivity levels correspond to larger trajectory clusters. This allows for stronger reverse updates, prioritizing the forgetting of highly privacy-sensitive trajectories; while the forgetting intensity of trajectory clusters with lower sensitivity levels is weaker, thus avoiding excessive degradation of model utility and achieving a balance between privacy protection and model performance.