CN107133527B - A kind of personalized recommendation method based on location privacy protection - Google Patents

Abstract

The invention discloses a personalized recommendation method based on location privacy protection. The user's real location P0 is used as the center of the circle, and dmax is used as the radius to generate a hidden area Z0. is the center of the circle, dmax is the radius, regenerate the hidden area Z′ ₀ , and the application server converts the radius of the hidden area Z′ ₀ to Dmax to generate a recommended area Z ₁ . _1. Sort the merchants within 1 to get a personalized recommendation list. The present invention ensures that the generated position (false trajectory) information is structurally consistent with the real position (trajectory) on the whole, thereby effectively resisting background knowledge attacks. At the same time, since the hidden area and the recommended area are in the same circle center, it can provide users with high-quality recommendation services while effectively resisting privacy attacks.

Description

A personalized recommendation method based on location privacy protection

technical field

The invention belongs to the technical field of data mining and privacy protection, and more specifically relates to a personalized recommendation method based on location privacy protection.

Background technique

Since 2003, some researchers have begun to work on the privacy protection of mobile user locations, and proposed some classic algorithms. These algorithms are classified, mainly including false trajectory data method, suppression method and data generalization method.

Usually the false trajectory data method is relatively simple to implement, with a large amount of data storage and relatively poor data availability. The suppression method protects trajectory privacy by restricting the release of sensitive information. This method is simple and requires little computation, but the data is easily distorted. The data generalization method is a trajectory privacy protection algorithm based on generalization, which ensures that the data will not be distorted, but the amount of calculation is relatively large.

At present, location privacy protection technology usually adopts literature [Gedik, Bu&#, Liu L.Protecting LocationPrivacy with Personalized k-Anonymity: Architecture and Algorithms[J].IEEETransactions on Mobile Computing,2008,7(1):1-18.]k -anonymity is the location K anonymity algorithm, which is the current mainstream method for location privacy protection. The location K anonymous algorithm is a method commonly used for location privacy protection. This method is to send the query user to the location server together with other k-1 users within a certain area, so that it is difficult to determine the real query user , location The K-anonymity algorithm has a good effect on the privacy protection of the query user's single location, but it is not suitable for continuous query. The attacker can calculate the real location information of the query user by querying the intersection of the user's location at continuous time.

The document [Theodoridis, State-of-the-art in privacy preserving data mining[C]//ACM SIGMOD Record.2004.] divides mainstream privacy-preserving data mining methods into five categories: ① Some ways of data distribution; ② 1. Based on the way of hiding data or rules, it is divided into data distortion, data anonymity, data encryption, etc.; ③. At the level of data mining technology, there are cluster mining, association rule mining, classification mining, etc.; ④. Said that it is divided into original data hiding, rule or pattern hiding, etc.; ⑤, in terms of privacy protection technology, it is divided into heuristic-based, cryptography-based and reconstruction-based methods.

Privacy protection and data mining are a pair of contradictions. The research and application of knowledge mining, machine learning, artificial intelligence and other technologies have made the power of big data analysis more and more powerful, and at the same time brought more severe challenges to the protection of personal privacy.

Contents of the invention

The purpose of the present invention is to overcome the deficiencies in the prior art and propose a personalized recommendation method based on location privacy protection to effectively resist privacy attacks such as background knowledge attacks and user behavior pattern attacks, and to provide users with Provide high-quality referral services.

In order to achieve the purpose of the above invention, the personalized recommendation method based on location privacy protection of the present invention is characterized in that it includes the following steps:

(1) Generate a hidden area according to the location of the query user

1.1), the location server receives the location service request Q={P ₀ (x,y),c,query} sent by the querying user, where P ₀ (x,y) is the real location of the querying user, and (x,y) is its Coordinates, c is the degree of privacy protection set by the user, c>1, query is the service request information sent by the user;

1.2), take the query user's real position P ₀ (x, y) as the center, and generate a hidden area Z ₀ with a radius of dmax, where the radius dmax=R×c, and R is the protection coefficient;

1.3) Determine whether the number n of the neighbors of the user’s real position P ₀ (x, y) in the hidden area Z ₀ satisfies n>k, and if not, kn position points need to be randomly inserted, where k is the hidden area The required location points are determined according to the specific implementation situation;

(2) Recalculate the hidden area according to the query user's real position P ₀ (x, y) neighboring position points

2.1), the location server randomly selects k adjacent location points of the hidden area _Z0 ;

2.2), obtain the coordinates of the k nearest neighbors, and calculate the mean value of the coordinates, through the formula

Get the mean coordinate position point Among them, x _i , y _i are the coordinates of the i-th point at position k;

2.3), the location server uses the mean coordinate location point is the center of the circle, dmax is the radius, regenerate the hidden area Z′ ₀ , and send the entire hidden area Z′ ₀ as the user’s current location to the application server, and at the same time, send the service request information to the application server;

(3) Recommend businesses near the user

3.1), the application server generates the recommended area Z ₁ from the hidden area Z′ ₀ radius to Dmax;

3.2), the application server sorts the merchants in the recommendation area Z1 according to the service request information query sent by the user, combined with the user's historical purchase merchant records, and then obtains _a personalized recommendation list and returns it to the query user.

The purpose of the present invention is achieved like this.

The present invention is based on a personalized recommendation method based on location privacy protection, which takes the user's real location coordinates, takes the user's real location P0 as the center, and dmax as the radius to generate a hidden area Z0, and calculates the mean value through the coordinates of the real location's neighbors, and then uses the mean coordinate position point is the center of the circle, dmax is the radius, regenerate the hidden area Z′ ₀ , and send the entire hidden area Z′ ₀ as the user’s current location to the application server, and at the same time, send the service request information to the application server; the application server will hide the area Z ′ ₀ radius to Dmax, generate a recommendation area Z ₁ , and then according to the service request information query sent by the user, combined with the user's historical purchase merchant records, sort the merchants in the recommendation area Z ₁ to obtain a personalized recommendation list. The present invention ensures that the generated position (false trajectory) information is structurally consistent with the real position (trajectory) on the whole, thereby effectively resisting background knowledge attacks. The hidden area is generated using the coordinates of the nearest neighbors. Since the calculation of the mean value of the position coordinates is used, the false location is closer to the place where the crowd is concentrated. This method can make the place corresponding to the real location consistent with the place corresponding to the hidden area, which can effectively resist targeting user behavior patterns. s attack. Since the hidden area and the recommended area have the same center, it can effectively resist privacy attacks and provide users with high-quality recommendation services.

Description of drawings

Fig. 1 is a schematic structural diagram of a personalized recommendation system applied in the present invention;

Fig. 2 is a flow chart of a specific embodiment of the personalized recommendation method based on location privacy protection in the present invention;

Fig. 3 is a schematic diagram of forming a hidden area;

Figure 4 is a diagram of the relationship between the hidden area and the recommended area;

Fig. 5 is mean value coordinate position point among the present invention track map;

Fig. 6 is a schematic diagram of location attack during continuous query;

Fig. 7 is a schematic diagram of maximum speed attack;

Fig. 8 is the track map of the query user's position that the third party can receive;

Fig. 9 is a comparison diagram of the protection degree of the position K anonymous algorithm without background knowledge;

Figure 10 is a comparison of random K-anonymous privacy protection degree with background knowledge;

Figure 11 is a graph showing the influence of location concealment on recommendation accuracy.

Detailed ways

Specific embodiments of the present invention will be described below in conjunction with the accompanying drawings, so that those skilled in the art can better understand the present invention. It should be noted that in the following description, when detailed descriptions of known functions and designs may dilute the main content of the present invention, these descriptions will be omitted here.

Fig. 1 is a schematic structural diagram of a personalized recommendation system applied in the present invention.

In this embodiment, the working process of the personalized recommendation system is as follows:

①. The querying user sends a location service request to the location server, and sends the location service request to the location server after encryption. Among them, the private key information of the inquiring user is known only to oneself, and the reliable communication between the inquiring user and the location server is encrypted.

②. The location server decrypts the received location information and service information, and generates a hidden area according to the real location information and the degree of privacy protection.

③. The location server packages and sends the obtained hidden area and service request information to the application server.

④. The application server combines the received service request information query with the user's historical purchase merchant records, sorts the merchants in the recommendation area Z1 to obtain _a personalized recommendation list, and sends it to the query user.

FIG. 2 is a flow chart of a specific embodiment of the location privacy protection-based personalized recommendation method of the present invention.

In this embodiment, as shown in Figure 1, the personalized recommendation method based on location privacy protection of the present invention is characterized in that it includes the following steps:

Step S1: Generate a hidden area according to the query user location

Step S1.1: The location server receives the location service request Q={P ₀ (x,y),c,query} sent by the querying user, where P ₀ (x,y) is the real location of the querying user, (x,y) Its coordinates, c is the degree of privacy protection set by the user, c>1, and query is the service request information sent by the user.

In this embodiment, x,y are latitude and longitude for the coordinates. The location of the query user is Niuwangmiao, Chenghua District, Chengdu, and its latitude and longitude are (104.099962, 30.651244). Send a location service request to the location server, and the location server generates a hidden area Z ₀ according to the location of the query user, specifically:

The location server (Location Based Service, referred to as LBS) receives the location service request sent by the query user, and the received information is Q={P ₀ (x,y),c,query}, where P ₀ (104.099962,30.651244) is the query user real location, c is the privacy protection level set by the user, c>1, query is the service request information sent by the user, and in this embodiment, query is the requested nearby restaurant information sent by the user.

Step S1.2: Take the real location P ₀ (x, y) of the query user as the center and generate a hidden area Z ₀ with a radius dmax, where the radius dmax=R×c, and R is the protection coefficient set by the location server.

After the location server receives the location service request, it starts to perform concealment calculations on the real location of the inquiring user, with P ₀ (104.099962, 30.651244) as the center and a radius of dmax to generate a hidden area Z ₀ , where dmax=R×c, in this embodiment , set the protection coefficient R=0.5km, the privacy protection degree c=2, and the location server will generate a hidden area Z ₀ within a radius of one kilometer.

Step S1.3: Determine whether the number n of the adjacent location points n of the user’s real location P ₀ (x, y) in the hidden area Z ₀ satisfies n>k, if not, it is necessary to randomly insert kn location points, where k is The location points required for the hidden area are determined according to the specific implementation. In this embodiment, k is determined to be 10.

Step S2: Recalculate the hidden area according to the user's real position P ₀ (x, y) neighboring position points

Step S2.1: The location server randomly selects k=10 neighboring location points in the hidden area _Z0 ;

Step S2.2: Obtain the coordinates of k=10 neighboring points, and calculate the mean value of the coordinates, through the formula

Get the mean coordinate position point Among them, x _i , y _i are the coordinates of the i-th point at position k. In this example, the mean coordinate position point The coordinates are (104.099692,30.650444).

Step S2.3: The location server coordinates the location point with the mean value is the center of the circle, and dmax is the radius, regenerate the hidden area Z' ₀ , and send the entire hidden area Z' ₀ as the user's current location to the application server, and at the same time, send the service request information to the application server.

Step S3: Recommend businesses near the user

Step S3.1: The application server sets the radius of the hidden area _Z'0 to Dmax to generate a recommended area Z1 _;

Step S3.2: _The application server sorts the merchants in the recommendation area Z1 according to the service request information query sent by the querying user, combined with the querying user's historical purchase merchant records, and obtains a personalized recommendation list.

In this embodiment, before the application server sorts the merchants, it needs to obtain the characteristics and weights of the merchants through training. Extract 1,000 merchants with a large number of purchasers in the application server database, divide them into positive and negative samples (purchased ones are positive samples, and those that browse without purchases are negative samples), and extract merchant characteristics, including whether parking, area, price , user ratings, etc., and then use the stochastic gradient descent method of the logistic regression algorithm to train the positive and negative samples to obtain the characteristics and weights of the merchants.

In this embodiment, as shown in FIG. 3, the circular area on the left is a hidden area Z ₀ within a radius of one kilometer with the real location P ₀ of the inquiring user as the center, and k=10 hidden areas Z ₀ are randomly selected. Neighboring position points (indicated by X), forming a circular area on the right is the mean coordinate position point is the hidden area Z′ ₀ within a radius of one kilometer from the center of the circle.

In this embodiment, the radius of the concealed area Z′ ₀ is set to Dmax=3km to generate a recommended area. The relationship between the recommended area and the hidden area is shown in Figure 4. In Figure 4, the inner circular area is the hidden area with a radius of dmax, and the entire circular area with a radius of Dmax including the hidden area is the recommended area. The recommended area can be described as the user mean coordinate position point is the center of the circle, a circular area with a radius of Dmax, and the hidden area is also a position point based on the user's mean value coordinates is the circular area with the center of the circle, using the user mean coordinate position point Represents the location of the entire hidden area. Although the user's location is hidden, the recommendation service for the querying user is a circular area with a large overlap with the hidden area. Therefore, while protecting the privacy of the querying user's location, it can also give the querying user Provide high-quality recommendation services and ensure data availability.

In this embodiment, the application server will put the user's mean value coordinate position point All restaurants within a three-kilometer radius are divided into positive and negative samples (purchased ones are positive samples, and those browsed but not purchased are negative samples), and the characteristics of the merchants are extracted, including whether parking, area, price, user rating... etc., and then use the stochastic gradient descent method of the logistic regression algorithm to train the positive and negative samples to obtain the characteristics and weights of the merchants, and finally obtain the historical records of the query user to obtain the historical purchase merchant records _of the query user. For the merchants in the recommended area Z1 Sort to get a personalized recommendation list.

Figure 5 shows that at three different moments, when the querying user sends a location service request, the hidden area Z ₀ and the location point based on the mean coordinates The formed trajectory, the position of the center of the circle (black origin) is the real position P ₀ of the query user, and the black triangle is the center of the generated hidden area Z′ ₀ The two are different, thereby hiding the location of the querying user.

The location K anonymous algorithm is a commonly used location privacy and query privacy protection method. This method is to send the query user to the location server together with other k-1 users within a certain area, so that it is difficult to judge the real query user. The location K anonymity algorithm has a good privacy protection effect on the location of a single query user, but it is not suitable for continuous query. Figure 6 is a continuous query attack: when the attacker intercepts the query user's location query information at different times, by observing the different locations included in the application query at different times, and intersecting the query user's location at consecutive time queries, the real location information of the query user is calculated.

As shown in Figure 6: the attacker initiates continuous query of the location information at four times, and the anonymous location information obtained at time t1 is (A, B, D, E, F), and at t2 (A, B, G, C, D), t3(A,B,C,E,G), t4(A,C,E,G,F). To find A in these location points, although each query can only get an anonymous set, it is impossible to tell who initiated the location request, but A can be identified by intersecting the location sets queried at these four times .

Figure 7 is a schematic diagram of the maximum speed attack, which is a kind of background knowledge attack. At T1, user C initiates a location query, and then generates two fake locations A and B, and generates an anonymous area S1; at T2, user C again A location query is initiated, and an anonymous area S2 is also generated. If the attacker obtains the user's transportation mode at this time, it can be roughly inferred that the user's speed is V, and according to the speed, the maximum range P that the user can reach between T1 and T2 can be obtained, so that the intersection of P and S2 can be inferred It is the real area that the user can go to, and the real location point C is further obtained. Through the hidden processing of the position coordinates in the present invention, the continuous query attack and the maximum speed attack can be well defended.

Figure 8 shows the location trajectory information of the querying user received by the third party after the processing of the present invention. Since the present invention performs generalized processing on the user's location, the third party can only receive the hidden location information of the querying user as a circle. area, whose center position is the mean coordinate position point However, the query user's real location _P0 is hidden, and the attacker cannot determine the real location.

By inventing and proposing an algorithm based on the mean value of location coordinates, the real location P ₀ of the query user is concealed, and Shannon entropy theory is used to measure the degree of anonymity of the algorithm. First, the definition of Shannon entropy is given:

Suppose a random variable x is a value in a finite set X, then the entropy of a random variable x is defined as:

p(x) is the probability when the value of the variable is x. The attacker launches an attack on the private information and successfully breaks the private information into an event set X, and the attacker successfully finds out that a user's private information is an event x in the event set, then the degree of privacy protection can pass the attacker's successful attack on the user's Measured by information entropy. The location K anonymous algorithm is the most widely used algorithm in the traditional location privacy protection, and the experiment measures the degree of privacy protection of the present invention and the location K anonymous algorithm.

When there is no background knowledge, the probability that an attacker can successfully obtain user location information:

If Q is set as the weight value of the background knowledge of a single trajectory owned by the attacker, 1≤Q≤n. The attacker may have the background knowledge of certain locations in the area, so the probability that the location privacy of the node with the background knowledge is successfully obtained by the attacker is Q/n+1 or Q/k+1, let m be the anonymous The number of nodes with background knowledge mastered by the attacker in the area, the probability that the attacker can obtain user location information without background knowledge is:

In the formula, Q _i means that the attacker has mastered it according to the i-th user. When the weight Q _i =1, it means that the attacker has not mastered any useful background knowledge. Qi ₌ n means that the attacker has mastered enough background knowledge to be able to Determine the user's location information. By calculating the probability, the degree of privacy protection is measured according to the information entropy. The larger the entropy value, the better the degree of privacy protection.

Figure 9 is a comparison of the privacy protection degree of the present invention and the position K anonymous algorithm without the attacker having background knowledge. The privacy protection degree here is represented by the information entropy mentioned above, and the abscissa k represents position K anonymity In the algorithm, there are k neighbor nodes, and the vertical coordinate H/bit represents the value of information entropy. When the attacker has certain background knowledge, the privacy protection degree comparison between the present invention and the position K anonymous algorithm is shown in Fig. 10 . It can be seen from Fig. 10 that the privacy protection degree of the present invention is stronger than that of the position K anonymous algorithm when the attacker has no background knowledge.

Recommendation accuracy can be defined as: the number of correct information extracted divided by the number of information extracted, since the present invention is a recommendation system implemented by a classification algorithm, accuracy can be defined as: the number of positive samples extracted divided by The total number of extractions. In order to measure the impact on the recommendation system after location concealment processing, the recommendation accuracy without concealment processing is compared with that after concealment processing, as shown in Figure 11: Figure 11 is five recommendation requests initiated, and there will be location concealment Comparing the recommended accuracy rate with that without concealment, it can be seen from the figure that the hidden data has little effect on the recommendation accuracy rate, and the recommendation accuracy rate can always be maintained above 90%.

Although the illustrative specific embodiments of the present invention have been described above, so that those skilled in the art can understand the present invention, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those of ordinary skill in the art, As long as various changes are within the spirit and scope of the present invention defined and determined by the appended claims, these changes are obvious, and all inventions and creations using the concept of the present invention are included in the protection list.

Claims (2)

1. a kind of personalized recommendation method based on location privacy protection, which comprises the following steps:

(1), secret area is generated according to inquiry user location

1.1), location server receives inquiry user and sends location service request Q={ P₀(x, y), c, query }, wherein P₀(x, It y) is inquiry user's actual position, (x, y) is its coordinate, and c is the secret protection degree of user setting, and c > 1, query are user The service request information of transmission；

1.2), to inquire user's actual position P₀(x, y) is the center of circle, and radius is that dmax generates secret area Z₀, wherein radius Dmax=R × c, R are protection factor；

1.3), determine secret area Z₀Intra domain user actual position P₀Whether the neighbor positions point number n of (x, y) meets n > k, if It is unsatisfactory for, needs k-n location point of radom insertion, wherein k is that position needed for secret area is counted, according to specific implementation situation It determines；

(2), according to inquiry user's actual position P₀(x, y) neighbor positions point recalculates secret area

2.1), location server selectes secret area Z at random₀K neighbor positions point；

2.2) coordinate of k neighbor positions point, and coordinates computed mean value, are obtained, formula is passed through

Obtain HCCI combustion location pointWherein, x_i,y_iFor i-th of coordinate of k location point；

2.3), location server is with HCCI combustion location pointFor the center of circle, dmax is radius, regenerates secret area Z '₀, and Entire secret area Z '₀It is sent to application server as user current location, meanwhile, service request information is also sent to Application server；

(3), businessman near recommended user

3.1), application server is by secret area Z '₀Radius generates to Dmax and recommends region Z₁；

3.2), the service request information query that application server is sent according to user buys merchant record in conjunction with user's history, To recommendation region Z₁Interior businessman's sequence obtains personalized recommendation list and returns to inquiry user.

2. the personalized recommendation method according to claim 1 based on location privacy protection, which is characterized in that application service Device needs to be trained the feature and weight of businessman, obtains the feature and weight of businessman before being ranked up to businessman:

The businessman that purchase number is more in application server database is extracted, is divided into positive and negative example sample, wherein purchase is positive example The sample that is negative that sample, browsing are not bought extracts businessman feature, and feature includes whether parking, area, price, user's scoring, Then positive and negative example sample is trained using the stochastic gradient descent method of logistic regression algorithm, obtains the feature and power of businessman Weight.