Social recommendation method based on differential privacy

Abstract

The invention discloses a social recommendation method based on differential privacy. A target function is disturbed based on Laplacian differential privacy so as to ensure the safety of user data. Firstly, data collection and initialization of a user description matrix and an article description matrix are carried out; and secondly, traversing is carried out for each article and each user in each round, the user carries out data interaction with a recommendation server and other users, a predicted value of the score of each article by the user is finally calculated after multi-round iteration, and personalized recommendation is carried out on the user based on the predicted value. On one hand, the method can protect privacy data of a user, so that intermediate data and results can resist attacks of stealing, reasoning and the like carried out by hostile attackers. On the other hand, the influence of the social relation between the users and the similar preferences between the users on the preferences of the users is considered at the same time, and a better recommendation effect can be provided.

Description

Technical Field

[0001] The present invention relates to the fields of information security, privacy protection and recommendation systems, and specifically to a social recommendation method based on differential privacy. Background Art

[0004] Commonly used privacy protection models such as k-anonymity and homomorphic encryption have been applied to recommendation systems to control the leakage of user privacy. However, k-anonymity limits the assumption of the attacker's background knowledge and the sensitive attributes of a group of users may be the same after generalization, so it is vulnerable to background attacks and similar attacks, which is not safe enough for privacy protection. Although cryptographic schemes such as homomorphic encryption can provide a sufficient level of security, they are difficult to apply in practical scenarios due to excessive computational overhead. In comparison, differential privacy has a lower computational overhead and provides a strict and quantitative concept of privacy, which limits the size of the difference in the final output of the algorithm caused by a record in the data set. It can ensure that even if the attacker knows all records except one in all data sets, the attacker cannot infer information about the unknown record. Therefore, social recommendation methods based on differential privacy help to provide good recommendation effects while protecting user privacy. Summary of the invention

[0005] The present invention provides a social recommendation method based on differential privacy. The main purpose of the present invention is to protect user privacy while providing good recommendation effects, ensuring that attackers cannot infer the user's specific privacy data from related information such as user historical scores and social relationships. At the same time, it provides a high-quality and low-overhead recommendation method for service providers such as government enterprises, increases user satisfaction and user stickiness in various application scenarios, and helps service providers better understand user needs. See the following description for details:

[0006] A social recommendation method based on differential privacy, the method comprising:

[0007] Step 1: The user collects his / her historical rating data on the local client and initializes his / her latent vector The recommendation server collects and generates the current item set and initializes the item description matrix where u i represents the hidden vector of the i-th user, which is a d-dimensional vector in the real field. V represents the item description matrix, which consists of the hidden vectors of each item, that is, v j Composition, v j is a d-dimensional vector in the real number field, V is a d-row and m-column matrix in the real number field, where m represents the total number of items;

[0008] Step 2: For each item in the dataset, the recommendation server calculates the total number of users n who have rated this item. j Sent to users who have rated it, where n j Represents the total number of users who rated the jth item n j ;

[0009] Step 3: After the user receives the number, noise is generated based on the infinite divisibility of the Laplace distribution The symbol represents the noise generated by user i on item j, which is in the form of a d-dimensional vector in the real number field. Afterwards, the user sends the intermediate result To the recommended server;

[0010] Step 4: The recommendation server aggregates all the collected intermediate results and updates the item latent vector based on them

[0011] Step 5: While the recommendation server is updating the item latent vector, the user updates his own user latent vector based on the social regularization term on the local client. First, the user sends the total number of his friends |F i |To your friends;

[0012] Step 6: The user's friends, that is, other users in the current social media that the user trusts through following, tagging, etc., receive the total number of |F i |After that, based on the infinite divisibility of Laplace, noise is generated The symbol represents the noise generated by user i's friend f when optimizing item j, which is in the form of a d-dimensional vector in the real number field. Sent to this user, where represents the similarity between user i and his friend user f for item j, u f represents the latent vector of user f;

[0013] Step 7: Other users who have rated the item generate noise The symbol represents the noise generated by other users c who have rated item j to user i, which is in the form of a d-dimensional vector in the real number field. After the noise is generated, other users will Sent to the current user, where represents the similarity between user i and user c for item j, u c represents the latent vector of user c;

[0014] Step 8: The user aggregates all intermediate results collected from friends and other scoring users and updates its own latent vector

[0015] Step 9: The above steps 2-8 are iterated continuously until convergence. The recommendation server calculates the final item hidden matrix and sends it to the user. The user side calculates The predicted values ​​of the user's ratings for each item are obtained, and the recommendation server recommends the items with higher scores to the current user.

[0016] Furthermore, the noise generation method in the above step 3 specifically includes:

[0017] (1) User i generates intermediate results based on the Gamma distribution Among them G 1i , G 2i are d-dimensional vectors in the real number field, and each element in the vector is an independent and random variable that conforms to the Gamma distribution. The shape parameter is The inverse scale parameter is Gamma distribution, let Then its probability density function is: In the anti-scaling parameter, Δ is the difference between the maximum score and the minimum score in the current recommendation system, and ε is the privacy budget;

[0018] (2) User Calculation as the noise generated by user i for item j.

[0019] Furthermore, in step 4, the recommendation server aggregates the intermediate results and updates the item latent vector, which specifically includes:

[0020] (1) The recommendation server collects the intermediate results and uses the formula

[0021]

[0022] Calculate the gradient value of the hidden vector of the current round of optimization item j;

[0023] (2) Recommended servers based on the formula Update the latent vector of the current item.

[0024] Furthermore, in step 8, the user aggregates the intermediate results of friends and other users who rate an item and updates its own latent vector, which specifically includes:

[0025] (1) The user collects the intermediate results and calculates the gradient value of the current round of optimization of its own latent vector. The gradient formula is as follows:

[0026]

[0027] (2) The user follows the formula Update its own hidden vector.

[0028] Compared with the prior art, the technical solution of the present invention has the following beneficial effects:

[0029] 1. Based on differential privacy technology, the present invention provides users with better recommendation effects while protecting their historical rating records. According to the infinite divisibility of Laplace distribution, a new noise generation mechanism is designed and the size of the privacy budget is determined, so that intermediate data and results can resist attacks such as theft and inference by malicious users or recommendation services, while reducing the impact of noise on recommendation results. It can also be theoretically proved to comply with ε-differential privacy.

[0030] 2. The recommendation scheme proposed in the present invention considers the social regularization term and the user-based collaborative filtering idea at the same time. For each individual item, it considers the influence of the social relationship between users and the similar preferences between users on the user's own preferences in a fine-grained manner, making the rating prediction more accurate and providing better recommendation results. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 A system block diagram of a social recommendation method based on differential privacy provided by the present invention;

[0032] Figure 2 A flowchart of a social recommendation method based on differential privacy provided by the present invention. DETAILED DESCRIPTION

[0033] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific 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.

[0035] Social recommendation methods based on differential privacy can provide good recommendation effects while protecting user privacy. Among them, the solution combining differential privacy with matrix decomposition is one of the main methods of recommendation system. The core idea of ​​matrix decomposition is to decompose a high-dimensional sparse matrix into two low-dimensional dense matrices, namely the user description matrix and the item description matrix. The two matrices will be used to predict the user's rating of the unpurchased items in the item set, and then provide personalized recommendations for users based on the rating data.

[0036] At present, recommendation methods that combine differential privacy and matrix decomposition can be mainly divided into two categories: recommendation methods based on trusted servers and recommendation methods that consider untrusted servers. In the trusted server scenario, the recommendation server is considered to be completely trustworthy, and it usually uses centralized differential privacy technology to add noise to the user's historical ratings to protect the true value. However, for economic reasons, the recommendation server may misuse or sell the personal information contained in the user's historical ratings, so it is not completely trustworthy. In the untrusted server scenario, while guarding against malicious users or malicious attackers, it is also necessary to guard against information theft attacks such as reasoning attacks that may be carried out by the recommendation server. It is necessary to balance the needs between privacy protection and recommendation effect. It is necessary to provide sufficient privacy protection to prevent the user's personal information or privacy information that can be inferred from being stolen by malicious users, recommendation servers or malicious attackers, and to provide users with good personalized recommendations to increase user stickiness and improve system availability.

[0037] The present invention provides a sarcasm detection method based on common sense enhanced sentiment dependency graph, aiming to achieve the following goals: 1. Seek a balance between recommendation effect and user data privacy protection. It can protect the user's historical rating data to protect the user's privacy information from external malicious attackers, malicious users and untrusted recommendation servers, and provide users with better personalized recommendation effects. 2. Improve the accuracy of recommendations. Make full use of social information and more thoroughly mine the remaining information contained in the historical rating data, comprehensively consider the social relationships between users and the impact of similar preferences between users on users' own preferences, and provide more accurate rating predictions to improve personalized recommendation effects.

[0038] The embodiment of the present invention provides a social recommendation method based on differential privacy, and its system block diagram is as follows: Figure 1 The embodiment of the present invention mainly involves two types of entities: users and recommendation servers.

[0039] 1. User: Rating the purchased items in the social media system to generate a historical rating data set, where the historical rating data set includes three parts: user ID, item ID, and rating value (numeric type, such as any value between 1-5 points). Users generate social trust relationships in social media through behaviors such as following. In this method, the trust relationship is directed binary data, that is, if a user trusts another user, the trust relationship is 1, otherwise the trust relationship is 0. The user's own trust relationship will generate a social relationship data set, which includes two parts: user ID and trusted user ID. There may be semi-honest users in the user set, that is, some users will follow the process to execute the method for reasons such as curiosity or economic interests, but will try to obtain the historical rating values ​​of other users in the process. For ease of understanding, in the following, we will describe the process of personalized recommendation method for a certain user, and use the user's friend set and other user sets that have rated a certain item to describe the process, but it should be noted that both the friend set and other user sets are users in this method, and will participate in the recommendation system as recommended users;

[0040] 2. Recommendation server: Maintain the current set of items, update the latent vector of each item in each round, and make the final item description matrix public to each user. In this method, the recommendation server is considered to be semi-honest, that is, although it will follow the process to execute the method, it may try to obtain the user's historical rating value due to curiosity or interests.

[0041] The threat model of this method is as follows:

[0042] 1. External malicious attackers, for reasons of interest, attempt to obtain the user's historical rating data and then infer the user's private information, possibly carrying out data theft attacks or reasoning attacks;

[0043] 2. The recommendation server may be semi-honest and can rely on methods to strictly implement the recommendation process, but may be curious about some user data and try to obtain it, or carry out reasoning attacks by analyzing the results;

[0044] 3. Some users, even those temporarily trusted by some people, may be semi-honest. Although they will execute the recommendation method, they are curious about user data and hope to obtain other users' personal data, and may carry out reasoning attacks.

[0045] The method flow provided by the example of the present invention is as follows Figure 2 The specific description is as follows:

[0046] This method is based on differential privacy and matrix decomposition technology to transform the original high-dimensional sparse rating matrix Decomposed into two low-dimensional dense matrices, namely the user description matrix and the item description matrix Where d represents the number of latent factors, that is, the number of hidden features of users and items. For example, if an item in the item description matrix is ​​a movie, its features may be different types of movie features such as action movies and comedies. The total number of types is the number of latent factors. The recommendation server is responsible for maintaining the item description matrix V throughout the process, that is, updating all item latent vectors in each iteration. For each item, the server will receive the intermediate results sent by all users who have scored it in each iteration, and is responsible for aggregating the intermediate results to update the item's latent vector. The user is responsible for maintaining and updating his or her own user latent vector throughout the process. At the same time, it is necessary to interact with the recommendation server and other users. Specifically, based on Laplace differential privacy, users generate noise for the items maintained by the recommendation server, friend users, and users who have rated the same item, and add it to the intermediate results, which are sent to the target server or user, so that the server can aggregate the results to update the item latent vector, and the user can aggregate the intermediate results to update the user latent vector. Finally, after several rounds of iterations and convergence, the recommendation server sends the final calculated item description matrix to each user, and the user calculates the product of its own latent vector and the item matrix to obtain the predicted value, and obtains the recommendation based on the predicted result.

[0047] The method perturbs the objective function based on Laplace differential privacy to ensure the security of user data. Multiple rounds of iterations are required, and in each round of iteration, the user needs to interact with the server and other users multiple times. The overall process of the method is shown in Algorithm 1:

[0048]

[0049] The algorithm input includes the following 7 parts:

[0050] (1) Objective function, the objective function of this method is as shown in formula (1), where the symbols appearing will be explained in detail in the subsequent description and process steps:

[0051]

[0052] (2) ε: Privacy budget. In privacy protection, the smaller the privacy budget, the higher the level of protection.

[0053] (3)γ: learning rate, which indicates the step size of each round in the iterative learning process.

[0054] (4) α: The influence weight of the social regularization term, that is, the influence of friends on one’s own interests.

[0055] (5) β: The influence weight based on the user's collaborative filtering idea, that is, the influence of other users who score the same item on their own interest.

[0056] (6)λ: matrix regularization coefficient, used to prevent overfitting.

[0057] (7) Raw data: including user historical rating data, user social relationship data, and item datasets maintained by the recommendation server.

[0058] The algorithm output is Contains the predicted ratings of each item by each user in the current system, based on which users will receive personalized recommendations.

[0059] This method consists of 9 steps, which can be divided into 3 stages, as detailed below:

[0060] 1. Initialization phase

[0061] The initialization phase includes step 1, completing data collection and initializing the user description matrix and item description matrix. The specific steps are as follows:

[0062] Step 1: The user collects his / her historical rating data and social relationship data on the local client and initializes his / her latent vector The recommendation server collects and generates the current item set and initializes the item description matrix where u i represents the hidden vector of the i-th user, which is a d-dimensional vector in the real field. V represents the item description matrix, which consists of the hidden vectors of each item, that is, v j Composition, v j is a d-dimensional vector in the real number field, V is a d-row and m-column matrix in the real number field, where m represents the total number of items.

[0063] 2. Iterative Calculation Phase

[0064] The iterative calculation phase includes steps 2-8, which can be represented as a three-layer loop in the algorithm. The first layer of loop is the training round, that is, the end condition of the iterative calculation phase is the convergence of the objective function (see formula (1)); the second layer of loop is the traversal operation of each item in each round; the third layer of loop is to consider each user who has rated it when traversing each item. Through the second and third layers of loops, this method can complete the traversal of each item and each user in each round, that is, traverse the existing historical rating records in the system. For ease of understanding, the following steps describe the update of the latent vector of an item and a user in a cycle. It should be noted that in the following, for the convenience of description, we use the user's friend set and the set of other users who have rated an item to describe the process, so as to characterize the specific process when a user updates his latent vector, but it should be noted that the users mentioned here, whether they are friend sets or other user sets, are users in this method, and will participate in the recommendation system as recommended users during the traversal process. The detailed description of each step is as follows:

[0065] Step 2: For each item in the dataset, the recommendation server calculates the total number of users n who have rated this item. j Sent to users who have rated it, where n j Represents the total number of users who rated the jth item.

[0066] Step 3: After the user receives the number, noise is generated based on the infinite divisibility of the Laplace distribution The symbol represents the noise generated by user i on item j, which is in the form of a d-dimensional vector in the real number field. Afterwards, the user sends the intermediate result To the recommendation server. Among them, Rij represents the score of the i-th user for the j-th item, and the noise generation method is as follows:

[0067] (1) User i generates intermediate results based on Gamma distribution Among them G 1i , G 2i are d-dimensional vectors in the real number field, and each element in the vector is an independent and random variable that conforms to the Gamma distribution. The shape parameter is The inverse scale parameter is Gamma distribution, let Then its probability density function is: In the anti-scaling parameter, Δ is the difference between the maximum score and the minimum score in the current recommendation system, and ε is the privacy budget.

[0068] (2) User Calculation as the noise generated by user i for item j.

[0069] Step 4: The recommendation server aggregates all the collected intermediate results and updates the item latent vector based on them in, Represents the gradient value of the item hidden vector, which is calculated according to the objective function (Formula (1)). In the objective function and the above gradient, I ij Indicates whether the i-th user has rated the j-th item. If yes, the value is 1, otherwise it is 0.

[0070] Step 5: While the recommendation server is updating the item latent vector, the user updates his own user latent vector based on the social regularization term on the local client. First, the user sends the total number of his friends |F i |To your friends.

[0071] Step 6: The user's friends, that is, other users in the current social media that the user trusts through following, tagging, etc., receive the total number of |F i |After that, based on the infinite divisibility of Laplace, noise is generated The symbol represents the noise generated by user i's friend f when optimizing item j. Its form is a d-dimensional vector in the real number field. The noise generation method is similar to that in step 3. The only difference is that That is, the parameters of the Gamma distribution are different. Afterwards, the user's friend will Sent to this user, where Indicates the similarity between user i and his friend user f for item j, using u f represents the latent vector of user f.

[0072] Step 7: Other users who have rated the item generate noise The symbol represents the noise generated by other users c who have rated item j to user i. It is in the form of a d-dimensional vector in the real number field. The noise generation method is similar to that in step 3. The only difference is that After the noise is generated, other users will Sent to the current user, where represents the similarity between user i and user c for item j, u c represents the latent vector of user c;

[0073] Step 8: The user aggregates all intermediate results collected from friends and other scoring users and updates its own latent vector in, As shown below, it represents the gradient value of the user latent vector, which is also calculated according to the objective function (Formula (1)).

[0074]

[0075] Iterate the above steps for each item and each user in each round until convergence, and the final item description matrix and user description matrix can be obtained.

[0076] 3. Result acquisition stage

[0077] The result acquisition includes step 9, which ultimately generates the predicted value of the user's rating for each item and makes recommendations based on it. The specific description is as follows:

[0078] Step 9: The recommendation server sends the final calculated item description matrix to the user. The user calculates The predicted values ​​of the user's ratings for each item are obtained, and the recommendation server recommends the items with higher scores to the current user.

[0079] The security analysis of the method provided by the embodiment of the present invention is as follows:

[0080] 1. This method satisfies ε-differential privacy

[0081] First, according to Laplace's infinity, we have in Laplace's infinite divisibility is described as follows:

[0082] Laplace's infinite divisibility: Let represents a Laplace distribution with location parameter θ and scale parameter s, which satisfies infinite divisibility. Specifically, for n ≥ 1, Each X i It is consistent with θ / n+Y 1n -Y 2n The distribution of Y 1n and Y 2n is gamma independent and identically distributed.

[0083] Assume that D1 and D2 are two datasets with only one record that is different. The records with differences are R ij and Two different data sets will produce the same output, that is, the same V. Since and V are the final results obtained after multiple rounds of iterations converge, we have Therefore, there is

[0084]

[0085] Since there is only one record in the two data sets, namely R ij and Different, so there are because And ||ui||<=1, so we have

[0086] For each item hidden vector v in V j ,have

[0087]

[0088] Therefore, the final item description matrix V satisfies ε-differential privacy.

[0089] The infinite divisibility of the Laplace distribution is stable, and we have Among them B n-1 is a random variable that follows a Beta distribution with parameters 1 and n-1. In the process of users updating their own latent vectors, Furthermore, Similar to the proof process of whether the item description matrix satisfies ε-differential privacy, the latent vector of each user also satisfies ε-differential privacy.

[0090] 2. Attack resistance analysis

[0091] This method uses differential privacy to perturb the objective function and add noise to the gradient to prevent the user's original score or user hidden vector from being stolen. For possible stealing attacks by external malicious attackers, since only the intermediate results with added noise are transmitted during the transmission process without any plaintext, the attacker cannot get the original score data or the real user hidden vector, and can only get the meaningless results after the perturbation. Therefore, it can effectively resist stealing attacks. For possible reasoning attacks by external malicious attackers, semi-honest recommendation servers and semi-honest users, since the noise is added to the gradient and the different noises generated by each user are updated once in each round, it is impossible to analyze any valid information from it, which can effectively resist.

[0092] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above embodiments can be implemented by software, or by means of software plus necessary general hardware platforms. Based on such understanding, the technical solutions of the above embodiments can be embodied in the form of software products, which can be stored in a non-volatile storage medium (which can be a CD-ROM, a USB flash drive, a mobile hard disk, etc.), including several instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0093] Those skilled in the art will appreciate that the accompanying drawing is only a schematic diagram of a preferred embodiment, and the serial numbers of the embodiments of the present invention are only for description and do not represent the advantages or disadvantages of the embodiments.

[0094] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A social recommendation method based on differential privacy, characterized in that: The following steps are involved: (101) Step 1: The user collects his / her historical rating data on the local client and initializes his / her latent vector The recommendation server collects and generates the current item set and initializes the item description matrix where u i represents the hidden vector of the i-th user, which is a d-dimensional vector in the real number field. V represents the item description matrix, which consists of the hidden vectors of each item, that is, v j Composition, v j is a d-dimensional vector in the real number field, V is a d-row and m-column matrix in the real number field, where m represents the total number of items; (102) Step 2: For each item in the dataset, the recommendation server calculates the total number of users n who have rated this item. j Sent to users who have rated it, where n j Represents the total number of users who rated the jth item n j ; (103) Step 3: After receiving the number, the user generates noise based on the infinite divisibility of the Laplace distribution The symbol represents the noise generated by user i for item j, which is in the form of a d-dimensional vector in the real number field. After that, the user sends the intermediate result To the recommended server; (104) The recommendation server aggregates all the collected intermediate results and updates the item latent vector based on them. (105) Step 5: While the recommendation server is updating the item latent vector, the user updates his own user latent vector based on the social regularization term on the local client. The user sends the total number of his friends |F i |To your friends; (106) Step 6: The user's friends, i.e. other users trusted by the user in the current social media through following, tagging, etc., receive the total number of |F i |After that, based on the infinite divisibility of Laplace, noise is generated The symbol represents the noise generated by user i's friend f when optimizing item j, which is in the form of a d-dimensional vector in the real number field. Afterwards, the user's friend will Sent to this user, where represents the similarity between user i and his friend user f for item j, u f represents the latent vector of user f; (107) Other users who have rated the item generate noise respectively The symbol represents the noise generated by other users c who have rated item j to user i, which is in the form of a d-dimensional vector in the real number field. After the noise is generated, other users will Sent to the current user, where represents the similarity between user i and user c for item j, u c represents the latent vector of user c; (108) Step 8: The user aggregates all intermediate results collected from friends and other scoring users and updates its own latent vector (109) The above steps 2-8 are iterated continuously until convergence. The recommendation server calculates the final item hidden matrix and sends it to the user. The user side calculates The predicted values ​​of the user's ratings for each item are obtained, and the recommendation server recommends the items with higher scores to the current user.

2. A social recommendation method based on differential privacy according to claim 1, characterized in that: The noise generation method is specifically as follows: (201) User i generates an intermediate result G based on the Gamma distribution 1i , Among them G 1i , G 2i are d-dimensional vectors in the real number field, and each element in the vector is an independent and random variable that conforms to the Gamma distribution. The shape parameter is The inverse scale parameter is Gamma distribution, let Then its probability density function is: x ≥ 0, where Δ in the inverse scale parameter is the difference between the maximum and minimum scores in the current recommendation system, and ε is the privacy budget; (202) User Calculation as the noise generated by user i for item j.

3. According to claim 1, a social recommendation method based on differential privacy is characterized in that: The method for the recommendation server to aggregate the intermediate results and update the item latent vector is specifically as follows: (301) The recommendation server collects the intermediate results and uses the formula Calculate the gradient value of the hidden vector of the current round of optimization item j; (302) Recommended server according to the formula Update the latent vector of the current item.

4. According to claim 1, a social recommendation method based on differential privacy is characterized in that: The user aggregates the intermediate results of friends and other users who score an item and updates its own latent vector specifically including: (401) The user collects the intermediate results and calculates the gradient value of the current round of optimization of its own latent vector. The gradient formula is as follows: (402) User according to the formula Update its own hidden vector.

Images