CN-121996847-A - Sensitive attribute forgetting method and device for recommendation system

CN121996847ACN 121996847 ACN121996847 ACN 121996847ACN-121996847-A

Abstract

The application provides a sensitive attribute forgetting method and device for a recommendation system. The recommendation system comprises an embedding layer of a user embedding matrix and an article embedding matrix, wherein the method comprises the steps of carrying out matrix decomposition on the article embedding matrix to obtain an irrelevant space orthogonal basis, carrying out characteristic decoupling on a user representation vector to obtain a relevant vector and an irrelevant vector, representing the irrelevant vector as a linear combination of parameters to be optimized and the irrelevant space orthogonal basis, constructing user representation distribution and sensitive attribute distribution, selecting an optimal bandwidth of an HSIC core, constructing a forgetting loss function for measuring the dependence between the user representation distribution and the sensitive attribute distribution, replacing the user representation vector in the forgetting loss function with the combination of the relevant vector and the irrelevant vector to obtain an objective function, and carrying out iterative optimization with the minimum objective function as a target under the condition of keeping the relevant vector unchanged to obtain a final user representation vector which is used as a result after forgetting the sensitive attribute.

Inventors

Yi Qingxiong
LI YUYUAN
Wu Machao
TENG XUYANG
LI XINYUE
ZHAO LIN
CHEN CHAOCHAO

Assignees

杭州电子科技大学

Dates

Publication Date: 20260508
Application Date: 20260410

Claims (10)

1. The sensitive attribute forgetting method for the recommendation system is characterized in that an embedding layer of the recommendation system comprises a user embedding matrix and an article embedding matrix, each row vector of the user embedding matrix corresponds to a different user representation vector, each column vector of the article embedding matrix corresponds to an article representation vector of a different article, and the method comprises the following steps: the method comprises the steps of S1, carrying out matrix decomposition on an object embedding matrix to obtain a relevant space orthogonal base and an irrelevant space orthogonal base, carrying out characteristic decoupling on a user representation vector to obtain a relevant vector and an irrelevant vector which are orthogonal to each other, wherein the irrelevant vector is expressed as a linear combination of parameters to be optimized and the irrelevant space orthogonal base, the difference between a dot product result of the relevant vector and the object embedding matrix and a dot product result of an original user representation vector and the object embedding matrix is smaller than a first preset threshold value, and the dot product result of the irrelevant vector and the object embedding matrix is smaller than a second preset threshold value; step S2, respectively constructing user representation distribution about the user representation vector and sensitive attribute distribution about sensitive attributes, and obtaining optimal bandwidth by optimizing the bandwidth of Hilbert-Schmitt independence criterion (HSIC core) to maximize test energy efficiency; S3, constructing a forgetting loss function for measuring the dependence between the user representation distribution and the sensitive attribute distribution by using the HSIC core under the optimal bandwidth, and replacing a user representation vector in the forgetting loss function with a combination of the related vector and the unrelated vector to obtain an objective function related to the parameter to be optimized; Step S4, under the condition that the relevant vector is kept unchanged, the parameters to be optimized are optimized to update the irrelevant vector by taking the minimization of the objective function as a target, and the parameters to be optimized are combined with the relevant vector to obtain a new user representation vector; and S5, reconstructing user representation distribution based on the new user representation vector, and repeatedly executing the steps S3 to S4 until a preset iteration stop condition is met, wherein the final user representation vector is used as a result after forgetting the sensitive attribute.
2. The method according to claim 1, wherein the step S1 comprises: Singular value decomposition is carried out on the article embedded matrix to obtain a left singular value matrix, a right singular value matrix and a singular value diagonal matrix; Based on the singular value diagonal matrix, before calculation Cumulative energy duty cycle of the individual singular values; Decomposing the right singular value matrix into a correlation space orthogonal base and an irrelevant space orthogonal base which are mutually orthogonal based on a minimum rank corresponding to the energy threshold super-parameter value which enables the accumulated energy duty ratio not to be smaller than a preset energy threshold super-parameter value; The user representation vectors are decoupled into the correlation vectors and the independent vectors that are mutually orthogonal based on the correlation spatial orthogonal basis and the independent spatial orthogonal basis.
3. The method according to claim 1, wherein the linear combination of the step S1 with respect to the extraneous vector is: ; Wherein, the The vector is represented for the user and, As the vector of the correlation, In order for the parameters to be optimized, Is an independent spatially orthogonal basis.
4. The method according to claim 1, wherein the step S2 comprises: Performing distribution construction processing on all user representation vectors embedded in the matrix based on Gaussian verification to obtain user representation distribution; And carrying out distribution construction processing on the attribute set of the recommendation system based on label verification to obtain sensitive attribute distribution.
5. The method according to claim 1, wherein the forgetting loss function in step S3 is: ; Wherein, the The distribution is represented for the user and, For sensitive attribute distribution, the user represents the distribution Distribution of sensitive properties The optimal bandwidth corresponding to the HSIC cores is that User representation distribution The optimal bandwidth corresponding to the HSIC core between itself is , As a function of the forgetfulness loss, Is a balance coefficient.
6. The method according to claim 5, wherein said step S2 comprises calculating an optimal bandwidth by : User representation distribution for vectors related to original user representations Distribution of sensitive properties Calculating an unbiased estimate of HSIC The said For bandwidth A function; calculating the said Standard deviation estimation of (2) The said For bandwidth Is a function of (2); by maximising Obtaining the optimal bandwidth 。
7. The method according to any one of claims 1 to 6, further comprising the step of: And S6, updating the user representation vector in the recommendation system into the user representation vector with the forgetting sensitive attribute.
8. The sensitive attribute forgetting device for the recommendation system is characterized in that an embedding layer of the recommendation system comprises a user embedding matrix and an article embedding matrix, wherein each row vector of the user embedding matrix corresponds to a different user representation vector, each column vector of the article embedding matrix corresponds to an article representation vector of a different article, and the device comprises: The device comprises a first processing unit, a user representation vector, a first processing unit, a second processing unit and a third processing unit, wherein the first processing unit is used for carrying out matrix decomposition on the object embedding matrix to obtain a relevant space orthogonal base and an irrelevant space orthogonal base, carrying out characteristic decoupling on the user representation vector to obtain a relevant vector and an irrelevant vector which are orthogonal to each other, wherein the irrelevant vector is expressed as a linear combination of parameters to be optimized and the irrelevant space orthogonal base, the difference between a dot product result of the relevant vector and the object embedding matrix and a dot product result of an original user representation vector and the object embedding matrix is smaller than a first preset threshold, and the dot product result of the irrelevant vector and the object embedding matrix is smaller than a second preset threshold; A second processing unit, configured to construct a user representation distribution about the user representation vector and a sensitivity attribute distribution about the sensitivity attribute, respectively, and obtain an optimal bandwidth by optimizing a bandwidth of a hilbert-schmitt independence criterion (HSIC core) to maximize a test energy efficiency; The third processing unit is used for constructing a forgetting loss function for measuring the dependence between the user representation distribution and the sensitive attribute distribution by utilizing the HSIC core under the optimal bandwidth, and replacing a user representation vector in the forgetting loss function with the combination of the related vector and the irrelevant vector to obtain an objective function related to the parameter to be optimized; the target optimization unit is used for optimizing the parameters to be optimized to update the irrelevant vector with the aim of minimizing the objective function under the condition of keeping the relevant vector unchanged, and combining the parameters to be optimized with the relevant vector to obtain a new user representation vector; And the iteration processing unit is used for reconstructing user representation distribution based on the new user representation vector, repeatedly executing the contents of the third processing unit and the target optimization unit until a preset iteration stop condition is met, and taking the final user representation vector as a result after forgetting the sensitive attribute.
9. An electronic device, comprising: A processor; A computer readable storage medium having stored therein computer program instructions which, when executed by the processor, cause the processor to perform the method of any of claims 1 to 7.
10. A computer readable storage medium, characterized in that the computer readable storage medium has stored thereon a computer program which is executed by a processor by the method according to any of claims 1 to 7.

Description

Sensitive attribute forgetting method and device for recommendation system Technical Field The invention relates to the technical field of data security, in particular to a sensitive attribute forgetting method and device for a recommendation system. Background Along with the rapid development of internet and mobile internet technologies, the personalized recommendation system is widely applied to various online service platforms, such as an e-commerce platform, a short video platform and the like. The recommendation system constructs user portraits by collecting historical behavior data of the user, including clicking, browsing, collecting, praying, purchasing, scoring and other behaviors, and further models potential interests of the user so as to realize personalized content pushing. However, as the deployment scale of the recommender system continues to expand, its potential privacy risks develop. Because the recommendation system has strong user behavior modeling capability, sensitive attribute information related to gender, age, race and the like is often implicitly encoded while user preference is learned. Once the attribute information is obtained illegally or misused by a third party, the disclosure of personal identity information of a user may be caused, so that serious privacy security problems are caused. Currently, aiming at the requirement that sensitive attributes in a recommendation system are forgotten, related technologies are mainly divided into two types: The first is a forgetting method in training, that is, an antagonistic constraint or privacy protection mechanism is introduced in the model training process to reduce the inferability of the model to sensitive attributes. Representative methods include countermeasure training, privacy aware regularization, and coding of representations of the training data with perturbations to attenuate sensitive properties. The method can carry out explicit constraint on dependence of the model on sensitive attributes in the learning process, but usually needs to redesign and execute a complete training process, and can cause the problems of unstable training process, difficult convergence, obviously increased computing overhead and the like, so that the application is limited in a recommendation system which is deployed on a large scale or needs to be updated frequently. The second type is a training attribute forgetting method, namely under the premise of avoiding retraining, the inference capability of the model on specific attributes is reduced to an unavailable level through means such as a post-processing mechanism or parameter fine adjustment and the like. Typical schemes include introducing a dual objective loss function, designing pluggable fine tuning modules (e.g., sub-network structures based on information theory constraints), etc., which have better compatibility with deployed models, requiring only limited intervention based on existing parameters. However, the post-training forgetting method still faces two key bottlenecks, namely, the first key bottleneck is that an effective performance maintenance mechanism is lacked, effective decoupling is not realized between forgetting intervention and recommended precision, excessive weakening of precision in practical application directly influences system effectiveness, the second key bottleneck is that the supporting capability of bottom-layer representation is insufficient, the related method deeply reveals entanglement mechanisms between user representation vectors and sensitive attributes, so that the generality and adaptability of forgetting strategies are weak, and diversified attribute forgetting requests in a real scene are difficult to flexibly cope with. In summary, the related schemes have certain limitations in dealing with the problem of forgetting the attribute in the recommendation system. Therefore, there is a need for a sensitive attribute forgetting method with high efficiency, low cost and performance maintenance, which can not only realize stable protection of the performance of the recommendation system, but also flexibly adapt to the forgetting requirements of various sensitive attributes. Disclosure of Invention In view of the above, the present application provides a method and apparatus for forgetting sensitive attribute of recommendation system. Specifically, the application is realized by the following technical scheme: According to a first aspect of embodiments of the present specification, there is provided a sensitive attribute forgetting method for a recommendation system, the recommendation system including an embedding layer, the embedding layer including a user embedding matrix and an item embedding matrix, each row vector of the user embedding matrix corresponding to a different user representation vector, each column vector of the item embedding matrix corresponding to an item representation vector of a different item, the method comprising the steps of: