Fake location privacy protection method and system based on mobile path semantic association

Abstract

The present invention belongs to the field of privacy protection technology, and discloses a pseudo-location privacy protection method and system based on the semantic relevance of mobile paths. The present invention uses a neural network to learn the spatiotemporal and semantic features between various locations in the user's real mobile path, and combines the transition probability between location sequences to quantify the impact of query time on the spatiotemporal and semantic features between locations, and proposes a pseudo-location privacy protection scheme based on the semantic relevance of mobile paths. Theoretical analysis and a large number of experiments show that the scheme can efficiently help users generate pseudo-location sequences that are indistinguishable from real location sequences in terms of spatiotemporal and semantic features, and effectively protect the location privacy of the requesting user.

Description

Technical Field

[0001] The present invention belongs to the technical field of privacy protection, and in particular relates to a pseudo-location privacy protection method and system based on mobile path semantic relevance. Background Art

[0002] With the development of mobile smart devices and networks, location-based services (LBS) have become an indispensable part of people's lives. However, while users enjoy the convenience brought by LBS, they also face the risk of personal privacy leakage. The reason is that when users send LBS, they need to provide location information containing personal privacy to the location service provider (LSP), which makes LSP collect and abuse the location information submitted by users, and then obtain users' personal privacy.

[0003] To solve the above problems, researchers at home and abroad have proposed location privacy protection methods such as those based on fake location, K-anonymity, differential privacy, cryptography, and coordinate transformation. Among the existing location privacy protection methods, the fake location-based method is widely used because it does not rely on a third party, does not require complex cryptographic technology, and allows users to obtain accurate query results.

[0004] However, existing location privacy protection schemes based on fake locations have not measured the semantic correlation between the real query location sequences when requesting users to enjoy continuously from a holistic perspective. But in real life, the next visited location is not only affected by the previous visited location, but also by other historical visited locations in this trip. This is because in daily life, the location sequences in the user's mobile path often have close semantic correlations. If users directly use existing fake location-based schemes to protect their location privacy, LSP can use the semantic correlation between the locations in the mobile path formed by the location set sequences submitted by the user to identify certain fake locations, thereby reducing the user's location privacy protection level, and can even directly identify the user's real location. For example, suppose a user is on a certain weekday. When LBS requests occur continuously at any time, the corresponding fake locations are generated using the existing fake location-based solution. and The fake locations of hospitals, shops, and bars are generated at all times. arrive and from arrive The movement paths "stadium-store" and "hospital-bar" formed at each moment are reasonable; however, from the perspective of location sequence, arrive The mobile path "hospital-store-bar" formed at all times is unreasonable in terms of customary association. At this time, LSP can infer with a high probability that the mobile path is a false path, and thus identify each location in the path as a false location. Therefore, the existing location privacy protection scheme based on false locations cannot effectively protect the user's location privacy.

[0005] The location privacy protection scheme based on fake location was first proposed by Kido et al. based on the idea of ​​confusion. The basic idea is that users send a fake location containing their real location and The location set of the false location is given to LSP, so that LSP has at most The probability of identifying the user's real location. Represents the number of privacy protection requirements of users. Lu et al. pointed out that when the fake locations generated by users are concentrated in a certain area, the location privacy of users cannot be effectively protected. To solve this problem, they introduced the concept of location dispersion, so that the distance between each fake location generated by the user and the real location is as far as possible. However, the above scheme does not take into account the background knowledge possessed by LSP, so that LSP can use background knowledge such as geographic information to remove some unreasonable fake locations such as those located in the middle of the intersection and the lake, thereby effectively reducing the privacy protection level of users. Therefore, Xia Xingyou et al. designed a privacy protection architecture based on a semi-trusted third-party server, using historical query probability and Stackelberg game model to generate fake locations. In addition, Zhang et al. pointed out that when using fake locations to protect the location privacy of users, the time factor should be introduced into the generation of fake locations. Liu et al. found that when users frequently send LBS in the same area, the query frequency of the area where the generated fake location is located should be the same as the query frequency of the area where the real location is located. Parmar and Rao introduced the query area into the generation of fake locations to avoid the user's real location and the generated fake location being located in the center of the corresponding query area.

[0006] The above schemes cannot effectively protect the location privacy of users when sending LBS. To solve this problem, Liu et al. proposed a pseudo-location privacy protection scheme based on spatiotemporal correlation. They measured the correlation between the real locations of users in continuous queries from both temporal and spatial aspects, so as to more effectively protect the location privacy of users when they continue to use LBS. Hara et al. proposed a pseudo-location generation scheme based on user preferences. This scheme ensures that there are points of interest that meet the user's preferences in the user's continuous query area. In addition, Tang et al. predicted the real location of the user when sending LBS queries in the future, and pre-generated pseudo locations for all future queries at the initial moment of continuous queries. This method ensures that the pseudo-movement path formed between the location sets submitted by the user is indistinguishable from the user's real movement path. However, once the user sends an LBS query outside the predicted location, this scheme cannot effectively protect the user's location privacy.

[0007] However, all the above schemes do not consider the semantic similarity between the real location and the fake location when generating fake locations. If the above schemes are directly used to protect the user's location privacy, LSP can use semantic similarity to identify some fake locations, or even directly identify the real location.

[0008] A fake location privacy protection scheme based on location semantics

[0009] Chen and Shen first introduced the concept of location semantics into pseudo location generation. They found that if the pseudo location generated has the same semantics as the user's real location, LSP can still infer the user's personal privacy. To solve this problem, they used the distance between nodes in the location semantic tree to measure the similarity of different location semantics. By maximizing the distance between different locations in the semantic tree, they ensured that the pseudo location generated was semantically diverse from the user's real location, thereby protecting the user's location privacy. However, this scheme would generate pseudo locations that are easy to identify, such as lakes or mountain tops. To solve this problem, Yang et al. combined query entropy with location semantics and used a multi-objective optimization algorithm to ensure the following two points: (1) The number of historical queries of the generated pseudo location and the user's real location is as equal as possible; (2) While ensuring that the number of queries for each location is as equal as possible, the generated pseudo location is semantically diverse from the real location. To prevent the generated false locations from being concentrated in a certain area with the real locations, Wang Jie et al. introduced the concept of location dispersion into the generation of false locations, so that the generated false locations and the real locations are semantically diverse and have the same number of queries, while making the distance between these locations (including false locations and real locations) as far as possible. However, when the generated false locations are too far away from the real locations geographically, LSP uses the city map to identify some false locations. To solve this problem, Tadakaluru et al. considered the time when the user sent the LBS and used the Voronoi diagram to split the map area, thereby avoiding the generated false locations being too far away from the real locations geographically.

[0010] The fake location privacy protection scheme based on location semantics only considers the semantic similarity between the generated fake location and the user's real location, ignoring the sensitivity of different location semantics. If the above scheme is used directly, LSP can still identify some fake locations using location semantics. To solve this problem, Li and Wu used knowledge graphs to measure the sensitivity of location semantics; MIN et al. used differential privacy mechanisms to protect the semantic sensitivity of generated fake locations and users' real locations.

[0011] The pseudo-location privacy protection scheme based on location semantics is only applicable to protecting the location privacy of users during a single discrete query. However, in real applications, users will send LBS continuously. Therefore, Bindschaedler and Shokri proposed that when generating pseudo-locations for users' continuous queries, the spatiotemporal similarity and semantic similarity between adjacent real locations should be considered. Xing et al. pointed out that when users send LBS continuously, the sensitivity of some of their real locations is low, so only the real locations with higher location sensitivity need to be protected. To achieve the above goals, they proposed a pseudo-location replacement algorithm, which uses the generated pseudo-locations to replace the real locations with high sensitivity during continuous queries.

[0012] Through the above analysis, the problems and defects of the prior art are as follows:

[0013] (1) Attackers and even location service providers themselves collect and abuse user location information, resulting in serious user data and privacy leakage problems;

[0014] (2) Existing location privacy protection schemes based on fake locations fail to fully consider the background knowledge possessed by LSPs. LSPs can use relevant prior knowledge to identify certain fake locations or even directly infer the real location.

[0015] (3) Existing location privacy protection schemes based on fake locations fail to fully consider the spatiotemporal correlation and semantic correlation between the generated fake locations and the user's real location sequence, and cannot ensure the indistinguishability of the generated fake locations;

[0016] (4) When users continuously enjoy LBS, the generation of fake locations in existing location privacy protection schemes based on fake locations fails to fully consider the characteristics of mobile trajectories from a holistic perspective, resulting in semantic conflicts between some fake locations and users' historical trajectories.

[0017] It can be seen that the existing location privacy protection scheme based on fake location cannot effectively protect the user's location privacy. Summary of the invention

[0018] In view of the problems existing in the prior art, the present invention provides a pseudo-location privacy protection method based on the semantic association of mobile paths.

[0019] The present invention is implemented as follows: a pseudo location privacy protection method based on mobile path semantic relevance includes:

[0020] Step 1: Use the long short-term memory network to extract the spatiotemporal and semantic features of each position sequence in the user's mobile path, and introduce the attention mechanism to assign different weights to the correlation between each position in the user's real position sequence, so as to generate candidate pseudo positions with the same spatiotemporal and semantic features as the user's real position sequence;

[0021] Step 2: Taking the user's historical movement path as a reference, the transition probability between position sequences is calculated, and the transition probability between positions is used to calculate the transition probability between positions at different time intervals. The similarity under quantifies the influence of the current query time on the generated candidate fake locations in terms of spatiotemporal features and semantic features. The higher the similarity, the greater the influence of the location on the historical moving path on the candidate fake location in terms of spatiotemporal semantic features. Conversely, the lower the similarity, the smaller the influence. At this time, when the generated candidate fake locations are screened, the weights of the influencing factors will be changed according to the different time intervals, so that the remaining fake locations after screening can effectively protect the user's location privacy.

[0022] Further, the candidate false positions having the same spatiotemporal features and semantic features as the user's real position sequence are generated:

[0023] (1) The generated fake location is consistent with the user's real location in terms of location sequence correlation. The neural network is used to learn and extract the location sequence features of the user's mobile path, and generate a fake location semantic category that conforms to the mobile features.

[0024] (2) According to the predicted pseudo-location semantic categories, candidate pseudo-locations are obtained, and pseudo-locations that meet the location transfer probability screening conditions are screened out, ensuring that the probability of each generated pseudo-location satisfying the user's historical location point on the historical movement trajectory transferring to the pseudo-location is greater than a preset threshold.

[0025] Furthermore, the steps of screening the generated candidate pseudo positions are as follows:

[0026] First, determine the dynamic position transfer probability screening threshold. This threshold is determined by the elements in the transfer probability set of the position when the different positions are themselves the semantic state before the transfer. The screening threshold is a certain percentile of the set, and is affected by the time factor. The longer the time interval between two positions, the smaller the threshold will be set. The probability of the point transfer candidate false position on the user's historical movement path must be greater than the dynamic threshold, and all candidate positions are scored according to the historical transfer probability, and the top position with the highest score is finally output. A fake location.

[0027] Another object of the present invention is to provide a pseudo location privacy protection system based on mobile path semantic association, comprising:

[0028] The generation module is used to extract the spatiotemporal features and semantic features of each position sequence in the user's mobile path using a long short-term memory network, and introduce an attention mechanism to assign different weights to the correlation between each position in the user's real position sequence, thereby generating candidate pseudo positions with the same spatiotemporal features and semantic features as the user's real position sequence;

[0029] The screening module is used to quantify the influence of the current query time on the generated candidate pseudo locations in terms of spatiotemporal features and semantic features by calculating the transition probability between location sequences based on the user's historical movement path. The screening module screens the generated candidate pseudo locations so that the remaining pseudo locations after screening can effectively protect the user's location privacy.

[0030] Another object of the present invention is to provide a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the pseudo-location privacy protection method based on the semantic association of the mobile path.

[0031] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, enables the processor to perform the steps of the pseudo-location privacy protection method based on the semantic association of the mobile path.

[0032] Another object of the present invention is to provide an information data processing terminal, which is used to implement the pseudo-location privacy protection system based on the semantic association of the mobile path.

[0033] In combination with the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solutions to be protected by the present invention are as follows:

[0034] First, the present invention uses a neural network to learn the spatiotemporal and semantic features between each location in the user's real mobile path, and combines the transition probability between location sequences to further quantify the impact of query time on the spatiotemporal and semantic features between locations. Based on this, a pseudo-location privacy protection scheme based on the semantic association of mobile paths is proposed. Theoretical analysis and a large number of experiments show that the scheme can efficiently help users generate pseudo-location sequences that are indistinguishable from real location sequences in terms of spatiotemporal and semantic features, thereby effectively protecting the location privacy of the requesting user.

[0035] Second, as auxiliary evidence of the inventiveness of the present invention, it is also reflected in the following important aspects:

[0036] This solution provides an innovative solution to the privacy leakage problem caused by users' continuous use of LBS. This solution uses a neural network to learn the spatiotemporal and semantic features between locations in the user's real mobile path, and combines the transition probability between location sequences to accurately quantify the impact of query time on the spatiotemporal and semantic features between locations, so that the generated fake location sequence is indistinguishable from the user's real location sequence in terms of spatiotemporal and semantic features, thereby effectively protecting the location privacy of the requesting user.

[0037] This solution optimizes the problem that the existing location privacy protection technology becomes ineffective when users send LBS multiple times in a period of time in daily life. In the scenario of continuous LBS service, the location sequences in the user's mobile path often have a close semantic correlation, that is, the next location visited by the user is not only affected by the previous location, but also by other historically visited locations in this trip. Therefore, the present invention selects a neural network to extract the spatiotemporal features and semantic features between the location sequences in the user's mobile path, and uses the transition probability between locations as a measurement standard to ensure the generation of a false location sequence that is indistinguishable from the real location sequence in terms of spatiotemporal features and semantic features.

[0038] Among the existing location privacy protection methods, the location privacy protection methods based on fake locations have a major technical difficulty to be overcome, that is, how to make the generated fake location more consistent with the user's real trajectory, making it difficult for attackers to identify, thereby protecting the user's real location. Although the existing research schemes based on fake locations have paid attention to the impact of the correlation between locations on the credibility of fake locations, most of them are limited to the correlation between adjacent locations, and rarely pay attention to the correlation between non-adjacent locations, especially the semantic correlation between non-adjacent locations. The present invention uses the user's historical movement path as a reference, uses a neural network to extract the user's movement path features and uses the attention mechanism to adjust the feature weights. Finally, based on the location transfer probability, the fake location that meets the rationality is selected to ensure that the user's privacy can be effectively protected when using LBS.

[0039] Third, the existing technical problems solved by the technical solution of the present invention in industrial applications are mainly concentrated in the protection of mobile user privacy, especially in the protection of location privacy. In the prior art, commonly used privacy protection methods, such as fake location generation, obfuscation strategies and anonymity technologies, often fail to effectively balance the spatiotemporal correlation and semantic consistency of locations. Specifically, the generated fake location sequence usually lacks correlation with the user's actual movement trajectory, resulting in limited effect on user location privacy protection, and even being cracked by attackers through spatiotemporal feature analysis or semantic matching. In addition, the existing methods lack in-depth analysis of historical mobile data and user behavior patterns when generating fake locations, resulting in unreasonable fake locations or not conforming to the user's actual behavior habits, further reducing the effectiveness of user location privacy protection.

[0040] The significant technical advances of the present invention are mainly reflected in the following aspects:

[0041] 1. Introducing the long short-term memory network (LSTM) and attention mechanism: Compared with traditional methods, this invention uses LSTM to extract spatiotemporal features in the user location sequence, which can effectively capture the dynamic changes in the user's mobile path. Through the attention mechanism, different weights are assigned to different positions in the location sequence, which enhances the correlation between positions and makes the generated fake location closer to the real path in terms of spatiotemporal features and semantic features. The introduction of this deep learning model significantly improves the rationality of fake location generation and the difficulty of privacy protection.

[0042] 2. Screening in combination with historical mobility data: After generating a fake location, the present invention refers to the user's historical mobility data, calculates the transition probability between location sequences, and screens out fake locations that match the user's behavior pattern. This makes the fake location similar to the real location sequence and can maintain consistency with the spatiotemporal characteristics of the user's behavior, thereby effectively resisting location privacy attacks. Compared with the prior art method of generating fake locations in a purely statistical manner, the present invention significantly improves the authenticity of the fake location by introducing neural networks and analyzing historical data.

[0043] 3. Dynamic optimization of spatiotemporal and semantic features: In real-time query scenarios, the present invention dynamically adjusts spatiotemporal and semantic features by analyzing the impact of the current time point on candidate pseudo locations, so that the generated pseudo location sequence is more consistent with the user's current mobility status. This technological advancement makes the pseudo location sequence more flexible, significantly enhances the ability to protect privacy in dynamic query environments, and further improves the protection strength of location privacy.

[0044] 4. Enhance user privacy protection and improve user experience: Traditional privacy protection methods often affect user experience when pursuing high privacy protection effects. For example, the deviation between the false location and the actual location features is too large, resulting in reduced quality of service to users. The present invention can improve the high privacy protection effect without significantly affecting the user's normal application experience through reasonable simulation of spatiotemporal features and semantic features, thereby achieving a balance between privacy protection and system functions.

[0045] In summary, the present invention has made significant technological progress in the field of location privacy protection by innovatively combining deep learning technology, historical data analysis and semantic optimization, solving the problems of poor correlation and limited privacy protection effect in the prior art of false location generation, and can effectively enhance the privacy security of users in industrial applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 It is a flow chart of a pseudo-location privacy protection method based on mobile path semantic association provided by an embodiment of the present invention.

[0047] Figure 2It is a structural block diagram of a pseudo-location privacy protection system based on mobile path semantic relevance provided by an embodiment of the present invention.

[0048] Figure 3 This is a user path diagram provided by an embodiment of the present invention.

[0049] Figure 4 It is a system structure diagram provided by an embodiment of the present invention.

[0050] Figure 5 This is a POI classification diagram provided by an embodiment of the present invention.

[0051] Figure 6 It is a structural diagram of the solution provided by an embodiment of the present invention.

[0052] Figure 7 This is a flow chart for generating a pseudo location semantic type provided by an embodiment of the present invention.

[0053] Figure 8 It is a structural diagram of the input layer provided by an embodiment of the present invention.

[0054] Fig. 9 It is a neural network structure diagram provided by an embodiment of the present invention.

[0055] Fig.10 This is a path location time interval distribution diagram provided by an embodiment of the present invention.

[0056] Fig.11 It is a similarity graph between the original transition probability set and different conditional probability sets provided by the embodiment of the present invention.

[0057] Fig.12 2 is a graph of training set and validation set loss (a) training set (b) validation set provided by an embodiment of the present invention.

[0058] Fig.13 1 is a diagram of the accuracy of the training set and the validation set provided by an embodiment of the present invention (a) training set (b) validation set.

[0059] Fig.14 It is a relationship diagram between the privacy protection requirement and the anonymity success rate provided by the embodiment of the present invention.

[0060] Fig.15 This is a relationship diagram between the number of generated semantic types and the anonymity success rate provided by an embodiment of the present invention.

[0061] Fig.16 4 is a relationship diagram between the screening threshold and the anonymity success rate provided by an embodiment of the present invention.

[0062] Fig.17 It is a graph of the number of paths that meet the position transfer probability condition provided by an embodiment of the present invention.

[0063] Fig.18 It is a privacy protection strength diagram of different solutions provided by the embodiments of the present invention when LBS is used continuously.

[0064] Fig.19 This is a comparison diagram of average computational delays for generating anonymity sets provided by an embodiment of the present invention.

[0065] Fig. 20 It is an average calculation delay diagram of generating false positions under different conditions provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0066] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0067] like Figure 1 As shown, a pseudo location privacy protection method based on mobile path semantic association provided by an embodiment of the present invention includes the following steps:

[0068] S101, using the long short-term memory network to extract the spatiotemporal features and semantic features of each position sequence in the user's mobile path, and introducing the attention mechanism to assign different weights to the correlation between each position in the user's real position sequence, and generate candidate pseudo positions with the same spatiotemporal features and semantic features as the user's real position sequence;

[0069] S102, taking the user's historical movement path as a reference, by calculating the transition probability between position sequences, and using the similarity between the transition probabilities of each position at different time intervals, quantifying the influence of the current query time on the generated candidate pseudo positions in terms of spatiotemporal features and semantic features. The higher the similarity, the greater the influence of the position on the historical movement path on the candidate pseudo positions in terms of spatiotemporal semantic features; conversely, the lower the similarity, the smaller the influence. At this time, when screening the generated candidate pseudo positions, the weights of the influencing factors will be changed according to different time intervals, so that the remaining pseudo positions after screening can effectively protect the user's location privacy.

[0070] The working principle of the pseudo-location privacy protection method based on mobile path semantic relevance provided by the embodiment of the present invention is mainly divided into two steps: extracting user location features through a deep learning model, and filtering pseudo-locations through probability calculation.

[0071] First, in step S101, LSTM is used to extract the spatiotemporal features and semantic features of each location sequence in the user's mobile path. LSTM has the ability to process time series data and can capture the law of the user's mobile trajectory changing over time. It combines the attention mechanism to assign different weights to different locations to ensure that the generated pseudo locations are similar to the real locations in spatiotemporal features and semantic features. Then, the attention mechanism is introduced to further enhance the relevance of the location sequence. The attention mechanism assigns different weights to each location based on the contextual information of each location to highlight the more important locations in the user's actual mobile path. In this way, the generated pseudo location sequence is not only similar to the user's actual path, but also has higher fidelity in the relevance of specific locations. This makes the movement features of the pseudo location sequence more consistent with the real location, thereby increasing the difficulty of privacy protection. In step S102, the system uses the user's historical mobile path as a reference, and evaluates the spatiotemporal features and semantic features of each candidate pseudo location at the current time point by calculating the transition probability between location sequences. The transition probability represents the possibility of a user moving from one location to another, and these candidate locations are evaluated in combination with the time factor. In this way, the system can dynamically select the fake location that is closest to the actual path characteristics based on the current query time and the user's mobile behavior history path to ensure the rationality and authenticity of the fake location. Finally, the candidate fake locations after the transition probability and spatiotemporal feature analysis will be further screened, leaving behind fake locations that can effectively confuse trackers. The screened fake locations can maximize the privacy of the user's actual location without affecting the user experience, while ensuring that the fake location is highly similar to the real path in spatiotemporal features and semantic features, ultimately achieving the effect of improving privacy protection.

[0072] The embodiment of the present invention provides a method for generating candidate false locations having the same spatiotemporal features and semantic features as the user's real location sequence:

[0073] (1) Using neural networks to learn and extract the location sequence features of the user's mobile path, and generate pseudo location semantic categories that match the mobile features;

[0074] (2) Obtain candidate pseudo locations based on the predicted pseudo location semantic categories, and select pseudo locations that meet the location transfer probability screening conditions.

[0075] The screening method provided by the embodiment of the present invention is as follows:

[0076] First, determine the dynamic position transfer probability screening threshold. This threshold is determined by the elements in the transfer probability set of different positions when they are the semantic states before the transfer. The screening threshold is a percentile of the set, and is affected by the time factor. The longer the time interval between two positions, the smaller the threshold will be set. The probability of the point transfer candidate false position on the user's historical movement path must be greater than the dynamic threshold, and all candidate positions are scored according to the historical transfer probability, and the highest score is finally output. A fake location.

[0077] like Figure 2 As shown, a pseudo location privacy protection system based on mobile path semantic association provided by an embodiment of the present invention includes:

[0078] The generation module is used to extract the spatiotemporal features and semantic features of each position sequence in the user's mobile path using the long short-term memory network (LSTM), and introduce an attention mechanism to assign different weights to the correlation between each position in the user's real position sequence, thereby generating candidate pseudo positions with the same spatiotemporal features and semantic features as the user's real position sequence;

[0079] The screening module is used to take the user's historical movement path as a reference, calculate the transition probability between location sequences, quantify the impact of the current query time on the generated candidate pseudo locations in terms of spatiotemporal features and semantic features, and screen the generated candidate pseudo locations so that the remaining pseudo locations after screening can effectively protect the user's location privacy.

[0080] Another object of the present invention is to provide a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the pseudo-location privacy protection method based on the semantic association of the mobile path.

[0081] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, enables the processor to perform the steps of the pseudo-location privacy protection method based on the semantic association of the mobile path.

[0082] Another object of the present invention is to provide an information data processing terminal, which is used to implement the pseudo-location privacy protection system based on the semantic association of the mobile path.

[0083] The present invention is specifically implemented:

[0084] The existing privacy protection schemes based on fake locations do not measure the semantic relevance between the real query location sequences when users continuously enjoy location services from a holistic perspective. If these schemes are directly used to protect the location privacy of users during continuous queries, LSP can still use the semantic relevance between the location sequences in the mobile path to identify some fake locations, or even directly identify the user's real location, such as Figure 3 shown.

[0085] 1.1 System Structure

[0086] like Figure 4 As shown in Figure 1, the system structure adopted in this paper mainly consists of two parts: users and LSP. Among them, LSP is a semi-trusted entity, and there is a reliable transmission channel between the user and LSP.

[0087] When a user wants to send an LBS query request, the map information of the current location is first obtained, and the semantic location database is accessed. At the same time, the historical query probability of the map location and the transfer probability between locations are calculated based on the user's historical movement path; then, according to the user's own privacy protection requirements, the number of , generate a containing its real position and The generated location is a collection of fake locations; finally, the generated location is sent to LSP together with its own query content. After receiving the query request sent by the user, LSP searches in its database according to the query content and location collection, and returns the search results to the user. After receiving the query results returned by LSP, the user filters the received query results according to his real location to obtain accurate query results.

[0088] 1.2 Definitions

[0089] Some basic concepts used in this article are given below, including the definition of user location and user historical semantic path.

[0090] Definition 1 User location , the user's location is a set consisting of location coordinates and access time, that is .in, represents the coordinates of the location; and Respectively represent the latitude and longitude of the location; Indicates the time the user visited the location.

[0091] This paper classifies the semantic type of any location according to the map classification document There are three levels:

[0092] (1) Major categories ;

[0093] (2) Middle category ;

[0094] (3) Subcategories .

[0095] in, Indicates the detailed semantic position in the category to which it belongs. Figure 5 As shown, the middle , belonging to the category , which includes subcategories .

[0096] Definition 2 User historical semantic path , the user The movement path formed by the real position sequence when sending LBS queries at a certain time in history is called a historical semantic path of the user: .in, Indicates user Send the first The location at the time of the LBS query.

[0097] In addition, if the map is divided into grids, then the number of times the user has sent LBS queries in each grid in history can be counted. At this point, the historical query probability of the user in any grid can be calculated. ,Right now

[0098] (1),

[0099] in, Indicates the user's history in the grid unit Obviously, the sum of the historical query probabilities of users in all grids is .

[0100] Similarly, the transition probability between positions The following formula can be used for calculation:

[0101] (2);

[0102] in, Indicates from position Move to location The number of times; Indicates from position The total number of times the When (i.e. location does not appear in the path as a position before the transfer), .

[0103] In LBS scenarios, when using location privacy protection technology based on fake locations, it should be ensured that the generated fake location sequences are semantically indistinguishable from the real location sequences. Otherwise, LSP can use the semantic correlation between the location sequences in the formed mobile path to identify the fake location, or even directly identify the user's real location. However, the existing location privacy protection schemes based on fake locations do not consider the semantic correlation between location sequences from a holistic perspective when generating fake locations. This paper uses neural networks to learn the spatiotemporal and semantic features between each location in the user's real mobile path, and combines the transition probability between location sequences to quantify the impact of query time on the spatiotemporal and semantic features between locations, and proposes a fake location privacy protection scheme based on the semantic correlation of mobile paths.

[0104] like Figure 6 As shown in Figure 1, the proposed scheme consists of two parts: pseudo location semantic type generation and pseudo location screening. When a user wants to send an LBS request, the semantic type of the pseudo location that meets the characteristics of the user's mobile path is first generated using a neural network based on the historical mobile path. ; False position filtering is to filter all the semantic types of the middle class The basic idea is to first remove the false locations with low transfer probability from the historical location to the location, and use the remaining false locations as candidate false locations; then combine the historical movement path of the user and position transfer probability Score candidate fake locations and use historical movement paths Calculate the position transfer probability , the probability that the historical trajectory position is transferred to the candidate false position Multiply by the corresponding time weight is a transition score of the candidate false position. Calculate the transition score from each historical trajectory position to the candidate false position and add them up to get the total score of the candidate false position , and finally select the one with the highest total score The candidate false positions are used as the false positions submitted to the LSP.

[0105] 2.1 Generation of pseudo-position semantic types

[0106] The generation of pseudo-location types mainly consists of three modules, namely, the input module, the semantic association learning module and the pseudo-location semantic type output module, e.g. Figure 7 The input module mainly encodes and embeds the user path and each location type, thereby extracting the time, space and semantic information in the user's historical movement path and generating an input sequence;

[0107] The semantic association learning module learns the spatiotemporal and semantic associations between different locations in the user's historical movement path; the pseudo-location semantic type output module generates the semantic type of the pseudo-location that conforms to the characteristics of the user's real movement path by extracting the spatiotemporal and semantic association features.

[0108] 2.1.1 Input Module

[0109] The input module preprocesses the user's original input path and connects the latitude and longitude, access time and semantics of each location in the normalized path to form an input sequence. Specifically, the longitude and latitude of each location in the path are first normalized and connected to form a spatial position tensor. in, Represents the real number space.

[0110] Since user movement paths often have certain periodicity and regularity in time, this article will use a certain day of the week and a certain time of day Mapped to Peacekeeping dimensional vector, i.e. , This helps to learn the weekly and daily cyclical patterns of users, thereby effectively describing the correlation between locations in the mobile path. Select the middle category in the location semantic type As the learning object of the semantic relevance of each location in the mobile path, its encoding is encoded with integers based on the classification of Amap. Since this encoding form is discrete, word embedding technology is used to embed them into a continuous numerical form, that is, the embedding of location semantics is , Represents the embedding dimension. Word embedding technology can make semantically similar words closer in the vector space and better capture the semantic relevance between words. After normalization and embedding operations, the input sequence is generated through concatenation operations. , so as to be input into the subsequent neural network for feature extraction and learning, that is, ;in, Represents the dimension of the embedding space, and the input layer structure is as follows Figure 8 shown.

[0111] 2.1.2 Semantic Association Learning Module

[0112] The semantic association learning module is used to extract the features of the user's real moving path. This module consists of a long short-term memory LSTM network layer and an attention layer, such as Fig. 9 shown.

[0113] In the network layer of LSTM, the correlation between the positions in the input position sequence is captured to generate a hidden state sequence which is input into the attention layer. In the attention layer, the hidden state sequence output by LSTM is weighted through the attention mechanism to generate a weighted output sequence, which assigns different weights to the correlation between different positions.

[0114] 1) LSTM network layer

[0115] This paper uses stacked LSTM to better learn the correlation between the locations in the user's mobile path. LSTM is a special recurrent neural network that is suitable for processing and predicting spatiotemporal sequence data. The network consists of a series of LSTM units, including a forget gate, an input gate, a candidate cell state, a cell state, and an output gate, which are , specifically,

[0116] (3);

[0117] (4);

[0118] (5);

[0119] (6);

[0120] (7);

[0121] (8);

[0122] in, is the time step The hidden state of the moment; is the activation function; is the time step Input of time; , , , They are the weights of the forget gate, input gate, candidate cell state, and output gate; , , , are the biases corresponding to each gate; Represents element-wise multiplication.

[0123] 2) Attention Layer

[0124] The attention layer can learn weights on the output sequence of the LSTM to distribute the attention to the time step Always hide status The hidden state corresponding to the output of LSTM at each moment The normalized historical query probability of the current location Concatenate to calculate the attention score. The attention score is composed of the historical query probability of the position Decision. Historical query probability The larger the value, the higher the attention score, which indicates that the position is in a key position for predicting user behavior, and vice versa. Attention score at a time step It can be calculated by the following formula:

[0125] (9);

[0126] in is a linear function defined as:

[0127] (10);

[0128] and denote the weight matrix and bias vector respectively.

[0129] According to the importance of different positions, the attention layer assigns different attention weights to each moment in the LSTM output sequence, and obtains the weighted output of the hidden state. This output is used as the input of the pseudo position semantic type output module. The specific calculation method is as follows:

[0130] (11);

[0131] (12);

[0132] in, is the time step The attention weight of is the weighted sequence output; is the sequence length.

[0133] 2.1.3 Output Module

[0134] 1) Model optimization

[0135] The output module is a fully connected neural network consisting of two linear layers, which can map the learned feature vector to a smaller dimension and finally transform it to the output space that matches the task. The specific calculation formula is as follows:

[0136] (13);

[0137] (14);

[0138] in, is the hidden state of the output module, is the final output of the model; and The linear layers are and The weight matrix of and The linear layers are and The bias vector of .

[0139] In order to optimize the prediction results, this paper uses mean squared error (MSE) as the loss function ,in For real data, is the number of samples. The stochastic gradient descent (SGD) algorithm is combined to optimize the parameters in the model and minimize the loss function. The SGD algorithm is suitable for large-scale data sets and can significantly reduce the amount of calculation and speed up training.

[0140] 2) Model output

[0141] This module is output by calculating the model With each position semantics The similarity of the embeddings is calculated and the location semantic types are sorted based on the similarity; then the closest one to the predicted vector is selected. Position semantic embedding vectors to select corresponding semantic labels , Represents a set of location semantic types, Indicates the number of prediction types.

[0142] According to the above description, the pseudo position semantic type is given below Generate an algorithm.

[0143] Algorithm 1: Semantic type generation algorithm for predicting false locations

[0144] Input: Model output , location semantic type set And the number of prediction types

[0145] Output: Predict the semantic type of the false position

[0146] 1. Get the position semantic type from the embedding layer The set of embedding vectors

[0147] 2. Calculate model output Embedding collections with positional semantic types The similarity of each embedding in ;

[0148] 3. Semantic labels in the location semantic type set based on similarity Sort by

[0149] 4. Select the first one from the sorted index semantic tags;

[0150] 5. Return the predicted location semantic type label as ;

[0151] 2.2 Fake location screening

[0152] Based on the generated pseudo location semantic type, a pseudo location generation scheme based on transition probability is designed. Since the influence of transition probability will gradually decrease over time, that is, the historical location To current location The transfer probability is affected by the time interval between two locations, so this paper introduces the time weight To adjust the original transfer probability, so as to more accurately describe the user's mobility behavior.

[0153] Historical Location Move to current position The actual transition probability is:

[0154] (15);

[0155] in, Is a current location Visit time Go to historical location Visit time Time interval The time weight determined; is the original transition probability without distinguishing the time feature.

[0156] In addition, to ensure that the generated fake locations are semantically indistinguishable, this solution comprehensively considers the dynamic impact of the transition probability between locations and the time dimension. 2.2.1 has described in detail how to calculate the differences in the location transition probability sets at different time intervals based on the user's mobile path, and determine the location transition probability time weight based on this to quantify the impact of time on state transition. The specific solution is to compare the original transition probability that does not distinguish time intervals. , and filter the time intervals between adjacent positions The position transfer probability set under the condition , using the similarity between sets as the time weight The measurement basis of , and in 2.2.2, it is explained how to use the actual position transfer probability To generate pseudo locations that are semantically relevant to the user's movement path.

[0157] 2.2.1 Time Weight Calculation

[0158] Step 1. A set of historical moving paths formed based on the real location of the user when sending LBS in history , Indicates The trajectory of each user. Calculate the transition probability between adjacent positions , is the specific value of the original transition probability. This transition probability can effectively describe the movement pattern of users between different locations.

[0159] Step 2. Based on the time distribution between the location points in the statistical historical movement path set , filter out different historical movement path sets , so that the new historical moving path selected Each position in satisfies the time interval between adjacent positions .

[0160] Step 3. For the new historical movement path set after filtering , respectively calculate the transition probability between adjacent position points in different sets , put into the position transfer conditional probability set middle;

[0161] Step 4. By comparing the original position transfer probability set and different time intervals The conditional probability set of position transfer under the condition , the similarity between sets is used to quantify the impact of time on the transition probability, and the time weight of the position transition probability is obtained :

[0162] (1) Position transition probability set Contains multiple key-value pairs, each key-value pair element Indicates the previous transfer position status in the current collection Transfer to the next transfer position state The probability of position transfer .

[0163] (2) Find the original position transition probability set and the set of conditional probabilities of position transfer The intersection of two sets, that is, finding the transition states that are contained in both sets , and calculate its position transfer probability The Difference If the probability values ​​of the common elements in the two sets are closer, The Difference The smaller it is, the more similar the two sets are.

[0164] (3) Find the original position transition probability set and the set of conditional probabilities of position transfer A unique element in a set, that is, a transition state that only appears in one set , its position transfer conditional probability set Directly reflects the difference between the two sets .

[0165] (4) Based on the difference value Calculate the similarity between the two sets and convert it into a percentage form as the final position transfer probability time weight The specific calculation formula is as follows:

[0166] (16);

[0167] in, Represents the total number of all key-value pairs in the collection.

[0168] Similarity between two sets: ;

[0169] Based on the above description, a specific position transfer probability time weight algorithm can be given.

[0170] The details are as follows:

[0171] Algorithm2: Position transfer probability time weight determination algorithm;

[0172] Input: original position transfer probability set; position transfer conditional probability set;

[0173] Output: transition probability time weight;

[0174] 1. Obtain the common transition state contained in the original position transition probability set and the position transition conditional probability set;

[0175] 2. Calculate the absolute value of the difference between the probability values ​​of each common transition state and add it to the difference value;

[0176] 3. Obtain the unique transition state value of each set and add it to the difference value;

[0177] 4. Calculate the total number of all key-value pairs in the two sets;

[0178] 5. Calculate the similarity between two sets;

[0179] 6. Time weight is λ t =similar t ×100.

[0180] 2.2.2 Fake location generation

[0181] Based on the given position transfer probability time weight determination algorithm, the generation process of the false position is explained in detail below.

[0182] First, according to the prediction in Section 2.1.3 The semantic type of the pseudo position at the query time , from the set of positional semantic types The class to which the semantic type belongs is obtained All locations Select fake positions and evaluate each position The specific process of the algorithm is as follows:

[0183] Step 1. Filter all acquired locations , so that the candidate false position satisfies the historical position arrive The transition probability Both are greater than the transition probability screening threshold . It is not a fixed constant, but a position on different historical trajectories. The set of transition probabilities The calculation formula is as follows:

[0184] (17);

[0185] in, is a time interval The function that increases and decreases, With the time interval Dynamic filtering threshold that increases and decreases.

[0186] Step 2. The transfer score of the candidate false position is the historical position The probability of moving to this point Multiply by the corresponding time weight Calculate the transfer score from each historical position to this point and add them up to get the total score of this candidate false position . According to the scores, all candidate false positions are sorted Sort in descending order and output the one with the highest score The pseudo locations are used as anonymity sets.

[0187] The specific pseudo position generation algorithm is given below as follows: Algorithm 3 pseudo position generation algorithm

[0188] Input: The current time t at which the LBS query is sent current ; Set of pseudo-position semantic types Move_t ; Number of user privacy protection requirements ; User historical movement path Traj ; Location semantic type set Poi_set , dynamic screening threshold

[0189] Output: Fake location loc poi

[0190] 1. Calculate the time difference between the current query time and the access time of the historical moving path location point Δt= t current - t i ;

[0191] 2. Get the weight λ based on the time difference Δt t ;

[0192] 3. From the location semantic type set Poi_set The class for obtaining semantic types is Move_t , and the transition probability of the points on the path to this position is greater than all false positions of the threshold ε as a candidate false position;

[0193] 4. Calculate each candidate false position The score relative to each historical position point is accumulated ;

[0194] 5. If the number of candidate pseudo positions that meet the screening criteria is greater than Then sort the candidate false positions in descending order according to the scores;

[0195] 6. Otherwise, randomly generate Move_t Other semantic types in the same semantic category , and obtain the pseudo position of the random semantics in the middle class semantic type;

[0196] 7. Repeat steps 6 to 8;

[0197] 8. If all fake locations of random semantic types cannot meet the user's privacy protection needs;

[0198] 9. Anonymous failure;

[0199] 10. Otherwise, output the top score Candidate pseudo positions loc poi As fake locations form an anonymity set.

[0200] 3.1 Security Analysis

[0201] Assume that an attacker attempts to pass user path information The semantic information of each location point in the filter is used to filter the generated fake location. After using the fake location generation scheme proposed in this paper, if , after receiving the submitted location set, it is difficult for the attacker to infer the false location based on the transition probability between locations. Represents the attacker’s transition probability screening threshold.

[0202] Assuming that the attacker wants to filter out fake locations based on the rationality of the user's path, he needs to determine whether the location points on the user's historical movement path and the points in the submitted location set meet the following conditions:

[0203] ;

[0204] If the attacker successfully identifies an unreasonable fake location, then:

[0205] ;

[0206] From the algorithm, we can know that , so we have: It is established, which is obviously consistent with the known Therefore, if When , the attacker cannot filter out unreasonable false positions through the position transfer probability. The probability that an attacker can identify the user's real location is ,Therefore, this algorithm can effectively resist attackers from using background knowledge and ,collected user information to infer the user’s real location and protect the user’s ,location privacy.

[0207] 3.2 Time complexity analysis

[0208] The time complexity of the two models in the scheme will be analyzed.

[0209] For the LSTM network model, the time complexity of data loading and preprocessing depends on the amount of input data, while the time complexity of data preprocessing such as data normalization is usually , so let the number of input data points be , then the time complexity of this part is The time complexity of the LSTM network layer mainly depends on the sequence length , hidden layer size , input dimension size And the number of LSTM layers In each time step, LSTM needs to calculate the multiplication of the weight matrix and the input as well as the multiplication of the weight matrix of the previous hidden state and the current state. Therefore, the time complexity of each time step is , the time complexity of the entire sequence is Since a two-layer LSTM structure is used, each layer performs a calculation, and the time complexity of the two-layer network model is .

[0210] Based on the above analysis, the time complexity of the pseudo-location semantic type prediction model is mainly determined by the forward propagation of the LSTM model, that is, the time complexity of the pseudo-location semantic type prediction model is ;

[0211] For the pseudo position generation module, after receiving the predicted After the semantic type of the user's pseudo location is determined, the POI collection must be read first. Finding the corresponding pseudo position according to the semantic type of the generated pseudo position can usually be considered a constant time After selecting the fake location, it is necessary to traverse the user's historical mobile path data and check the location transition probability for each location point and fake location to calculate the score of each location. Suppose the size of the transition probability set is , then its time complexity is . Then, each position needs to be sorted in descending order according to the score, and the time complexity is .in, The number of privacy protection requirements for users.

[0212] In the best case, a single calculation will find valid false positions, so the best time complexity of the algorithm is ; In the worst case, the user needs to iterate multiple times to obtain enough valid false positions. Suppose the number of other semantic types belonging to the same category as the generated semantic type is In the worst case, it is necessary to Therefore, the worst time complexity is .

[0213] In order to verify the effectiveness of the pseudo-location privacy protection scheme based on the semantic correlation of mobile paths proposed in this paper, this paper selects the Geolife dataset [34-35] of Microsoft Research Asia as experimental data. This dataset records 17,621 mobile paths of 182 users in 5 years (from April 2007 to August 2012), with a total distance of approximately 1.2 million kilometers and a total duration of 50,176 hours. The experiment selects the most densely populated area in the central Beijing area of ​​the dataset, that is, the latitude range is [39.8, 40.0] and the longitude range is [116.2,116.5]. The area is divided into 100×100 rectangular grids, and each grid is represented by the coordinates of its center point. Its semantic information is obtained by calling the reverse geocoding API of Amap. For the convenience of the experiment, the mobile paths of users in the dataset are preprocessed with a time interval of 5 minutes as the standard. After the above processing, the time intervals between adjacent locations in the path are mostly concentrated between 5 and 25 minutes, such as Fig.10 The rest of the experimental settings are shown in Table 1. The experiment uses Python 3.8 as the programming language, the operating system is Windows 10-64bit, and the hardware environment is AMD Ryzen 3 2200U with Randeon Vega Mobile Gfx (4 CPUs), ~2.5GHz 8192MB RAM.

[0214] Table 1 Experimental default parameters

[0215] parameter default value 5 μ 3 ε 50 (median)

[0216] According to the algorithm in 2.2.1, the similarity values ​​of the position transfer probability set under different time interval conditions are calculated, such as Fig.11 As shown, when As the time interval increases, the similarity between the transfer probability set at different time intervals and the position transfer probability set without distinguishing time gradually decreases, which is consistent with the rule that the influence of position transfer probability weakens over time. For the convenience of calculation, this paper counts the similarity interval as the reference range of weight values, as shown in Table 2.

[0217] Table 2 Time weight

[0218] Time interval Δt <![CDATA[Weight λ t > Δt≤5 75 5<Δt≤10 21 10<Δt≤25 10 25<Δt≤50 5 50<Δt≤95 3 95<Δt≤180 2 180<Δt≤335 1

[0219] 4.1 Analysis of model training process

[0220] In the training process of the neural network model in this paper, the dimension is set , 3 and 6, and the semantics is embedded into a 10-dimensional continuous vector. 70% of the data set is divided into a training set and 20% is divided into a test set. The number of iterations of the training set is 20. The SGD optimization algorithm is used to optimize the model parameters. The learning rate is 0.01, the batch size is 128, and the loss of the training process is as follows Fig.12 As shown, the accuracy is Fig.13 As shown in the figure. The optimizer calculates the loss value and uses back propagation to update the model parameters until the model converges. Therefore, as the number of training times increases, the model parameters are gradually optimized, the loss value gradually decreases, the model performance gradually increases, and the accuracy rate increases accordingly. After the 10th iteration, the model began to converge, and the change range of the loss value of the training set and the validation set tended to be flat. Since the frequency of position semantics in the path varies greatly, the imbalance of samples causes the accuracy rate of the training set to fluctuate.

[0221] 4.2 Anonymity Success Rate

[0222] because and The unique parameters will be controlled by the variable method, and other parameters will be kept as the default values ​​in Table 1. The number of privacy protection requirements will be discussed separately. , the number of semantic types of false positions And the position transfer probability screening threshold Impact on the success rate of anonymity.

[0223] like Fig.14 As shown in the figure, as the number of user privacy protection requirements increases As the number of pseudo locations that satisfy semantic relevance increases, the anonymity success rate decreases. This is because the number of pseudo locations that satisfy semantic relevance gradually cannot meet the needs of users.

[0224] like Fig.15 As shown in Figure 2, the number of semantic types of generated pseudo positions The increase in the anonymity success rate will increase, because the number of semantic types of the generated fake locations The more there are, the more locations are obtained, and the more false locations that meet the conditions increase.

[0225] Screening threshold is the set of probabilities of moving from different locations to any location Determined by Fig.16 As shown in Figure 1, the screening threshold of each position point is the 25th percentile, median and 75th percentile of the elements in its transition probability set. As the screening threshold increases, the number of positions that meet the screening conditions gradually decreases, and the success rate of this scheme in generating reasonable false positions decreases.

[0226] In general, since the location semantics of the generated fake locations come from the predictions of the neural network, they are consistent with the semantics and spatiotemporal correlations, and the good performance of the LSTM neural network has been verified in the analysis of the model training process. Therefore, there are more locations that meet the requirements of the probability of transition between locations, which also makes the anonymity success rate of the schemes in different situations higher.

[0227] 4.3 Path anonymity success rate

[0228] In the scenario of sending LBS queries continuously, the attacker obtains anonymous location sets at different times. Therefore, this section mainly verifies whether the path composed of the fake location set generated by this scheme meets the user's privacy protection requirements. .

[0229] For n sets of positions submitted at a certain time, each set contains m positions, which can form paths. When the number of privacy protection requirements of users is At least Therefore, in order to verify the effectiveness of this scheme, different Experiments are conducted with the value of , and the number of indistinguishable paths that can be composed of the generated set of false positions is calculated.

[0230] like Fig.17 As shown in Figure 2, the number of privacy protection requirements for different users , the number of paths that meet the position transfer conditions in the path is greater than the number of user privacy protection requirements ,Therefore, this scheme can effectively resist attackers from using background knowledge and ,location transition probability to filter false locations and ,thus reducing the privacy protection effect.

[0231] In addition, this method is compared with the DLP algorithm and the MMDS algorithm. Fig.18 As shown in the figure, the set of fake locations generated by the DLP algorithm has the least number of indistinguishable paths and most of them do not meet the privacy protection requirements. This is because the DLP algorithm does not consider the influence of semantic factors when generating fake locations, but only considers side information attacks based on query probability. Although the MMDS algorithm considers both location semantics and query probability, the MMDS algorithm does not involve location transfer probability factors.

[0232] 4.4 Computational Overhead Analysis

[0233] The number of different user privacy protection requirements , select different algorithms to repeat the experiment and verify the generation efficiency of this scheme. Fig.19As shown in Figure 2, as the privacy protection requirements increase, the average computational delays of different algorithms are increasing. Since the DLP algorithm does not consider the semantic characteristics of the location, the average latency required to generate the location set is smaller. In contrast, the average computational latency of the MMDS algorithm and this solution is closer because both solutions require traversing and filtering the data set. As the number of user privacy protection requirements increases As the number of fake locations increases, users need to continuously expand the scope and further screen more fake locations that meet their needs, which leads to a higher average calculation delay of this solution. However, the overall calculation delay of this solution is within the acceptable range for users. This experiment will study the semantic types of fake locations. The size of affects the efficiency of generating false positions. Fig. 20 As shown, with the semantic type of the pseudo position As the value of the pseudo position increases, the average computational delay of the algorithm decreases. This is because the semantic type of the pseudo position The more the number of , the more false positions it contains, the more false positions that conform to the transfer rules, and the higher the efficiency of generating false positions.

[0234] Existing location privacy protection schemes based on fake locations do not consider the semantic relevance between the real query location sequences when requesting users to continuously enjoy services from a holistic perspective when generating fake locations. If these schemes are used directly to protect the user's location privacy, then after the LSP constructs a moving path based on the received location set, it can use the semantic relevance between the location sequences in the formed moving path to identify the fake location, or even directly identify the user's real location. Therefore, the existing privacy protection schemes based on fake locations cannot effectively protect the user's location privacy. To solve this problem, this paper proposes a fake location privacy protection scheme based on the semantic relevance of the moving path. The scheme first uses the LSTM network to extract the spatiotemporal features and semantic features between the user's real location sequences, and generates candidate fake locations based on this; then, based on the user's historical moving path, the generated candidate fake locations are screened by quantifying the impact of the current query time on the location's spatiotemporal features and semantic features, so that the remaining fake locations can effectively protect the user's location privacy.

[0235] The specific application fields or related products of the present invention.

[0236] With the popularization of smart devices and the development of mobile Internet, LBS, as a mobile Internet service, is very popular in people's daily lives. LBS has demonstrated significant creative and application value in many fields. Its typical applications mainly include information retrieval services, navigation services, advertising push services, motion detection services, etc. For example, in real-time traffic navigation, tourism and city guide scenarios, navigation systems using LBS technology can provide real-time traffic information, help users avoid congestion, choose the best route, and dynamically adjust navigation paths according to traffic conditions. For example, Amap and Google Maps can update traffic conditions in real time, provide route suggestions and estimated arrival times. Location-based applications can also provide personalized travel recommendations and navigation services. Tourists can obtain nearby attractions, restaurants and accommodation information through applications, and make travel plans according to their interests to achieve a more efficient travel experience. Therefore, when using LBS, submitting a large number of fake locations that are real enough to be mistaken for real to the LSP together with the user's real location can effectively protect the user's location privacy, and in addition, it can also enable users to obtain accurate query results.

[0237] It should be noted that the embodiments of the present invention can be implemented by hardware, software, or a combination of software and hardware. The hardware portion can be implemented using dedicated logic; the software portion can be stored in a memory and executed by an appropriate instruction execution system, such as a microprocessor or dedicated design hardware. A person of ordinary skill in the art will appreciate that the above-mentioned devices and methods can be implemented using computer executable instructions and / or contained in a processor control code, such as a carrier medium such as a disk, CD or DVD-ROM, a programmable memory such as a read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. Such code is provided on the carrier medium. The device and its modules of the present invention can be implemented by hardware circuits such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, etc., or programmable hardware devices such as field programmable gate arrays, programmable logic devices, etc., can also be implemented by software executed by various types of processors, and can also be implemented by a combination of the above-mentioned hardware circuits and software, such as firmware.

[0238] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any modification, equivalent substitution and improvement made by any technician familiar with the technical field within the technical scope disclosed by the present invention and within the spirit and principle of the present invention should be covered by the protection scope of the present invention.

Claims

1. A pseudo-location privacy protection method based on semantic association of mobile paths, characterized in that: The following steps are involved: Step 1: Use the long short-term memory network to extract the spatiotemporal features and semantic features of each position sequence in the user's mobile path, and introduce the attention mechanism to assign different weights according to the correlation between each position in the user's real position sequence, and generate candidate pseudo positions with the same spatiotemporal features and semantic features as the user's real position sequence; Step 2: Taking the user's historical movement path as a reference, the transition probability between position sequences is calculated, and the transition probability between positions is used to calculate the transition probability between positions at different time intervals. The similarity degree under the condition quantifies the influence of the current query time on the generated candidate fake locations in terms of spatiotemporal features and semantic features. The higher the similarity degree, the greater the influence of the location on the historical moving path on the candidate fake location in terms of spatiotemporal semantic features. On the contrary, the lower the similarity degree, the smaller the influence degree. At this time, when the generated candidate fake locations are screened, the weights of the influencing factors will be changed according to the different time intervals, so that the remaining fake locations after screening can effectively protect the user's location privacy. The generating of candidate false locations having the same spatiotemporal features and semantic features as the user's real location sequence comprises: (1) The generated fake location is consistent with the user’s real location in terms of the correlation of the location sequence; (2) Ensure that each generated fake location can satisfy the probability that the historical location points on the user's historical movement trajectory will be transferred to the fake location; The screening method is as follows: First, determine the dynamic position transfer probability screening threshold, which is determined by the elements in the transfer probability set of different positions when they are the semantic states before the transfer. The screening threshold is a certain percentile of the set, and is affected by the time factor. The longer the time interval between two positions, the smaller the threshold will be set. The probability of the point transfer candidate false position on the user's historical mobile path must be greater than the dynamic threshold, and all candidate positions are scored according to the historical transfer probability, and the top position with the highest score is finally output. A fake location.

2. A pseudo-location privacy protection system based on mobile path semantic association that implements the pseudo-location privacy protection method based on mobile path semantic association as claimed in claim 1, characterized in that: The pseudo location privacy protection system based on mobile path semantic association includes: The generation module is used to extract the spatiotemporal features and semantic features of each position sequence in the user's mobile path using the long short-term memory network, and introduce an attention mechanism to assign different weights to the correlation between each position in the user's real position sequence, so as to generate candidate pseudo positions with the same spatiotemporal features and semantic features as the user's real position sequence; The screening module is used to take the user's historical movement path as a reference, calculate the transition probability between location sequences, quantify the impact of the current query time on the generated candidate pseudo locations in terms of spatiotemporal features and semantic features, and screen the generated candidate pseudo locations so that the remaining pseudo locations after screening can effectively protect the user's location privacy.

3. A computer device, characterized in that: The computer device includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the pseudo-location privacy protection method based on the semantic association of the mobile path as claimed in claim 1.

4. A computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor executes the steps of the pseudo-location privacy protection method based on semantic association of mobile paths as claimed in claim 1.

5. An information data processing terminal, characterized in that: The information data processing terminal is used to implement the pseudo-location privacy protection system based on the semantic association of the mobile path as claimed in claim 2.

Images