A content recommendation system cold start security risk detection method and device

Abstract

The application discloses a content recommendation system cold start security risk detection method and device, comprising: generating a user portrait according to the registration rules of the measured recommendation system, and constructing a corresponding user in the local system according to the generated user portrait; for any user, executing interaction in the local system based on the configured interaction strategy; constructing training data according to the interaction result; inputting the training data into a potential feature learning model to perform training; inputting the content data to be detected into the trained potential feature learning model; counting and inversely sorting the reconstruction error of the output result of the potential feature learning model; taking the specified number of output data after sorting as abnormal data and performing clustering; and judging whether the content recommendation system cold start exists a security risk according to the clustering result. The method can be used to judge whether the measured recommendation system is maliciously manipulated in the cold start stage.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computers, and in particular to a content recommendation system cold start security risk detection method and device. BACKGROUND

[0002] In a recommendation system, the addition of a new user usually has a cold start problem, which is specifically manifested in that the recommendation system has insufficient recommended content, the recommended content cannot meet the user's demand, and it is difficult to make recommendations due to insufficient user interaction, which is a common phenomenon in recommendation systems. However, the recommended content may be maliciously manipulated in this process, such as recommending topics related to the interests of the platform but not interesting to the user, or excessively pushing content related to the interests of the platform, or even pushing illegal content, which prevents users from obtaining information and legal content in a fair manner and leads to an overabundance of commercial information, thereby seriously infringing on the rights and interests of users and the user experience, and even causing security problems. Therefore, it is necessary to detect the cold start security problem of the recommendation system to protect the rights and interests of users.

[0003] Current research has focused on solving the cold start problem or improving the cold start speed, such as the MWUF framework, which is used to solve the cold start problem of new item embedding in deep models; the MvDGAE model, which proposes a denoise graph auto-encoder framework that allows more similar user or item features on the meta-path to be closer; and the ICAN framework, which improves the cold start recall effect by strengthening the feature domain interaction between multiple queues and environmental information. These cutting-edge researches can to some extent solve the cold start problem well, making the recommendation system more efficient and the recommended content more targeted, however, their current research mostly stays at the cold start problem itself, without focusing on the security problem behind the cold start technology. Although the cold start problem is being optimized, it has not been eliminated, and the cold start problem still exists in the recommendation system, so a malicious recommendation system may manipulate the recommended content during the cold start process, causing harm to the rights and interests of users. SUMMARY

[0004] The embodiments of the present application provide a content recommendation system cold start security risk detection method and device to determine whether a measured recommendation system is maliciously manipulated during the cold start phase.

[0005] The embodiments of the present application provide a content recommendation system cold start security risk detection method, comprising:

[0006] The network model is trained in advance by the following method:

[0007] According to the registration rule of the measured recommendation system, a user portrait is generated, and a corresponding user is constructed in the local system according to the generated user portrait;

[0008] For any user, an interaction strategy suitable for the user is configured, and interaction is performed in the local system based on the configured interaction strategy;

[0009] The interaction result is vectorized to construct training data;

[0010] The training data is input into a latent feature learning model, and training is performed based on a given reconstruction error threshold;

[0011] For the content data to be detected, the following steps are performed to complete the detection:

[0012] The trained latent feature learning model is input into the content data to be detected;

[0013] The reconstruction error of the output result of the latent feature learning model is counted and inversely sorted;

[0014] The first designated number of output data after sorting is taken as abnormal data and is clustered;

[0015] According to the clustering result, it is judged whether the content recommendation system cold start exists a security risk.

[0016] Optionally, constructing a corresponding user in the local system according to the generated user portrait comprises:

[0017] According to the generated user portrait, a completely identical user is constructed in the local system.

[0018] Optionally, the configured interaction strategy includes random interaction and / or fixed interaction, wherein:

[0019] The random interaction is that the user does not directly interact with the recommendation system;

[0020] The fixed interaction is that a set of interaction tracks matching the user portrait is set for any user.

[0021] Optionally, performing interaction in the local system based on the configured interaction strategy comprises:

[0022] One or more users created in the local system are selected, and one user corresponds to one latent feature learning model;

[0023] Any user is controlled to perform interaction with the local recommendation system according to the interaction strategy, so as to record the recommended content returned by the local system.

[0024] Optionally, based on the configured interaction strategy, the performing interaction in the local system further includes:

[0025] For any returned recommended content, the text content is not recorded, and the recommended content is vectorized to construct training data of any user.

[0026] Optionally, taking the top specified number of output data after sorting as abnormal data, and performing clustering includes:

[0027] A bag-of-words model is constructed in advance, and the top specified number of output data after sorting is taken as abnormal data input into the bag-of-words model to vectorize the abnormal data.

[0028] The vectorized abnormal data is clustered by K-means to obtain a clustering result.

[0029] Optionally, according to the clustering result, determining whether the content recommendation system cold start exists a security risk includes:

[0030] If the recommended content in a certain cluster of the clustering belongs to a certain type or publisher, it is determined that the tested recommendation system exists malicious manipulation.

[0031] The embodiment of the application also proposes a content recommendation system cold start security risk detection device, comprising a processor and a memory, the memory stores a computer program, and the computer program is executed by the processor to realize the steps of the content recommendation system cold start security risk detection method as described above.

[0032] The embodiment of the application also proposes a computer readable storage medium, the computer readable storage medium stores a computer program, and the computer program is executed by the processor to realize the steps of the content recommendation system cold start security risk detection method as described above.

[0033] The embodiment of the application learns the potential features of the recommended content by using the neural network model, so that the differences of specific content can be ignored, and the potential features of the content are compared and analyzed to determine whether the tested recommendation system is maliciously manipulated in the cold start stage.

[0034] The above description is only a summary of the technical solutions of the application, in order to more clearly understand the technical means of the application, the specific embodiments of the application can be implemented according to the content of the specification, and in order to make the above and other purposes, characteristics and advantages of the application more obvious and easy to understand, the following specific embodiments of the application are described. BRIEF DESCRIPTION OF DRAWINGS

[0035] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of this application. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0036] Figure 1 This is a basic flowchart example of the cold start security risk detection method for the content recommendation system in this application embodiment;

[0037] Figure 2 This is an example of the overall framework of the potential feature learning model in the embodiments of this application. Detailed Implementation

[0038] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0039] The cold start problem in recommender systems is prevalent across various social media platforms. Current research can only optimize the problem as much as possible and minimize its impact, but the risk of manipulating recommended content for new users during the cold start process still exists. However, current research has not yet addressed this issue. For example, the MWUF framework optimizes the cold start problem by proposing a new method for item embedding and two cooperating neural networks. This method can better capture item features, but it does not supervise or discriminate the content recommended during the cold start. The framework's final AUC metric only reaches a maximum of 70%, indicating that the cold start problem still exists in recommender systems. Furthermore, the applicant's research has not discussed or addressed the security issues related to the cold start process. The MvDGAE framework simulates real-world scenarios by randomly discarding some user-item interactions, allowing the cold start problem to be initially optimized during model training. It then uses a decoder to make the features of more similar users or items on the meta-path more closely approximate, thus enabling the recommendation system to better capture their relationships. However, none of these processes address the monitoring and discrimination of cold start content. Final experiments show that this method cannot completely eliminate the cold start problem, and it also fails to address the security issues associated with cold starts.

[0040] Based on this, this application provides a method for detecting security risks during the cold start of a content recommendation system. This method monitors the recommended content during the cold start process for new users and determines whether the system has maliciously manipulated the recommended content during the cold start process. Assume that the recommended content provided by a benign local recommendation system is S = {S1, S2, ..., S...}. N The method in this application is used to train a network model capable of capturing potential user preferences based on a local benign recommendation sequence S. Assume the recommended content sequence provided by the tested recommendation system is C = {C1, C2, ..., C}. N The ultimate goal of this application is to use a trained latent feature learning network to discriminate the content in sequence C, determine whether the content deviates from user preferences based on the reconstruction error output by the decoder, and then determine whether the recommendation system has malicious manipulation issues based on the relevance of the deviating content. Specifically, such as... Figure 1 As shown, the method of this application includes the following steps:

[0041] The network model is trained in advance using the following method:

[0042] In step S101, for the recommendation system under test, user profiles are generated according to the registration rules of the recommendation system under test, and corresponding users are constructed in the local system based on the generated user profiles. In some specific examples, the construction of user profiles satisfies the following points:

[0043] 1. User characteristics should be clearly and explicitly defined to avoid vague or unexplainable features.

[0044] 2. The number of user characteristics should be sufficient to fully define and differentiate the user.

[0045] 3. User characteristics should be relatively closely related to each other, avoiding the appearance of completely unrelated characteristics.

[0046] 4. Different users' characteristics need to be differentiated to a certain extent. For the same category of characteristics, try to select different directions to avoid users with similar characteristics.

[0047] In this embodiment, user characteristics will affect subsequent model training and anomaly detection models; therefore, the above requirements should be fully considered when defining them. Regarding the number of users, some applications may construct 10 users with identical user profiles both locally and in the system under test.

[0048] In some embodiments, constructing a corresponding user on the local system based on the generated user profile includes: constructing an identical user on the local system based on the generated user profile.

[0049] In step S102, an interaction strategy adapted to each user is configured, and interaction is performed on the local system based on the configured interaction strategy.

[0050] In some embodiments, the configured interaction strategy includes random interaction and / or fixed interaction, wherein:

[0051] The random interaction refers to user actions that do not involve direct interaction with the recommendation system, such as clicking, liking, or saving. Instead, the user simply refreshes the recommended content, such as the pull-to-refresh operation on platforms like Weibo and Douyin. This interaction method allows the recommendation system to make recommendations based on the user's characteristic profile, allowing the test system to observe how well it captures and values ​​user profiles.

[0052] The fixed interaction refers to setting a set of interaction trajectories for each user that match their profile. In this embodiment, the fixed interaction is defined as setting a set of interaction trajectories for each user that matches their profile. For example, for a user whose profile is teenager, their interaction trajectory is defined as: upon first refresh, clicking the first game content that appears; upon second refresh, clicking the first entertainment news content that appears; upon third refresh, clicking the second educational content that appears, and so on. This interaction method allows the recommendation system to capture user preferences more quickly and recommend content that matches the user's interests more rapidly.

[0053] The present invention constructs the training set of the learning network by gradually generating the training set through the interaction between the local benign recommendation system and the user, which makes the judgment criteria more convincing and prevents the tested recommendation system from denying its rights by taking advantage of the cold start problem.

[0054] Since the local system's benign recommendation system lacks prior user data and suffers from cold start issues, the comparison with the tested recommendation system is fair. The specific construction method is as follows: select one or more users previously created in the local system; each user will correspond to a latent feature learning model. Manipulate this user with the local recommendation system using a fixed interaction strategy, and record the recommendations returned by the local system. The reason for using a fixed interaction strategy is that it can capture the latent features of user preferences more quickly, thus accelerating the training process.

[0055] In step S103, the interaction results are vectorized to construct training data.

[0056] In step S104, the training data is input into the latent feature learning model, and training is performed based on a given reconstruction error threshold.

[0057] This application proposes using neural networks to learn the latent features of a benign recommendation system's output. This allows for a deeper understanding beyond superficial differences in recommended content, enabling the determination of malicious intent based on the underlying differences in the recommended content between the two systems. This application addresses the cold-start manipulation problem by utilizing comparison results between the tested system and a benign system, and also resolves the issue of differing recommended content in cross-recommendation system comparisons.

[0058] In some specific examples, such as Figure 2 As shown, the latent feature learning model in this application embodiment has the following functions:

[0059] a. Latent Feature Extraction

[0060] Learning networks need to have the ability to extract latent features in order to learn the latent features of the recommended content. Ideally, the extraction of latent features should facilitate the reconstruction of the original data.

[0061] b. Latent Feature Learning

[0062] The learning network in this framework needs to be able to learn the commonalities of the latent features of the recommended content and record the latent features of the recommended content in a parameterized form.

[0063] c. Original data restoration

[0064] The learning network can reconstruct the original data based on latent features and output reconstruction error, thereby enabling a quantitative comparison of the differences in latent features between different recommended content. This application can use reconstruction error to quantitatively measure whether there is a gap between the recommended content of the tested system and the recommended content of good quality.

[0065] To complete the detection of the data, follow these steps:

[0066] In step S105, the trained latent feature learning model is input into the content data to be detected.

[0067] In step S106, the reconstruction error of the output results of the latent feature learning model is statistically analyzed and sorted in reverse order.

[0068] In step S107, a specified number of output data points after sorting are taken as outliers and clustered.

[0069] In step S108, based on the clustering results, it is determined whether there are any security risks associated with the cold start of the content recommendation system.

[0070] This application embodiment utilizes a neural network model to learn the latent features of recommended content, thereby ignoring the differences in specific content and comparing and analyzing the latent features of the content to determine whether the tested recommendation system has been maliciously manipulated during the cold start phase.

[0071] In some examples, performing interactions on the local system based on a configured interaction policy includes:

[0072] Select one or more users created in the local system, with each user corresponding to a latent feature learning model.

[0073] Control any user to interact with the local recommendation system according to the interaction strategy, and record the recommended content returned by the local system.

[0074] In some examples, performing interactions on the local system based on configured interaction policies also includes:

[0075] For any returned recommended content, the text content is not recorded. Instead, the recommended content is vectorized to construct training data for any user.

[0076] In some specific examples, for each recommended content, we don't record its actual text content, but instead vectorize its tags and features. Ultimately, for a user i, suppose the dataset D constructed based on i... i Contains n i Recommended content: D i ={c1,c2,c3,…,c ni}, further will D i Recommended content c j Represented as c j ={w1,w2,w3,…,w mj}, where w k This indicates the characteristics of the recommended content, such as category, subcategory, number of likes, number of comments, etc.

[0077] In some examples, clustering is performed by taking a specified number of outliers from the sorted output data, including:

[0078] A bag-of-words model is pre-built, and a specified number of output data after sorting are taken as abnormal data and input into the bag-of-words model to vectorize the abnormal data;

[0079] The vectorized outlier data is clustered using K-means to obtain the clustering results.

[0080] In a specific example, the analysis method can be K-means clustering, with the clustering object being the top-k, such as the tags of the top-30 content, like type, publisher, etc. Before clustering, a bag-of-words model needs to be constructed to count the tags appearing in the top-k content and store them in the bag. The tags are then represented as vectors using statistical measures such as the count of each word in the bag. After that, the tag vectors of the top-k content are clustered to obtain the results.

[0081] In some examples, determining whether a content recommendation system's cold start poses a security risk based on clustering results includes:

[0082] If all recommended content in a cluster belongs to a certain type or publisher, then the tested recommendation system is judged to be maliciously manipulated.

[0083] In some specific examples, the clustering results are analyzed. If the recommended content in a certain cluster all belong to a certain type or publisher, the tested system may have maliciously manipulated that content.

[0084] This application also proposes an application example of a cold start security risk detection method for a content recommendation system. This application example consists of four stages: training data generation, latent feature learning model training, test data generation, and abnormal content detection. The main application scenario for this example is a network social platform with content recommendation capabilities.

[0085] Training Data Generation: For an untested recommendation system platform, this example first generates a user profile based on its registration rules. Then, it constructs an identical user profile on the local system, using the same profile from the tested system. In this example, the user profile is: a 40-year-old unmarried male with interests in photography, military affairs, and current affairs; currently employed. Next, this example needs to define a set of fixed interaction strategies tailored to this user. Here, the defined interaction strategy is: the first 5 interactions are with photography-related content; the 6th, 8th, and 10th interactions are with military-related books; and the remaining interactions are with current affairs-related content. Afterward, the recommendation results returned by the local recommendation system in these 10 interactions are statistically analyzed and vectorized to generate training data.

[0086] Training the latent feature learning model: For the latent feature learning model in the framework, this example uses the VAE variational autoencoder. The training dataset is input into the VAE model in the form of vectors for training. The reconstruction error threshold is set according to the actual needs. When the model loss converges and reaches the error threshold, the training stops and the trained VAE model is output.

[0087] Test data generation: The test data generation process for the fixed interaction strategy is basically the same as the training data process, the only difference being that the data generation platform is the platform under test. For the test data of the random refresh strategy, this example uses a pull-down refresh directly on the recommendation platform under test, such as Weibo, to obtain the data. This example will refresh the platform 10 times and take the top-10 recommended content from each refresh as the test data.

[0088] Anomaly Content Detection: After the test data is generated, it is input into the trained VAE model. The reconstruction error of each data point is statistically analyzed and reverse-sorted. The top-30 anomaly data points with the highest reconstruction errors are then analyzed. This example uses the K-means algorithm to cluster the anomaly content. Other suitable clustering algorithms, such as dbScan and Mean Shift, can also be used. If the content in a cluster is mostly in the game category, and some of it belongs to the same game company, it is considered that the recommendation system tends to promote this game or game from this company to new customers. If the content in a cluster is mostly from a certain brand, it is considered that the recommendation system tends to promote this brand's products to new customers. If the content in a cluster is unrelated, such as all being games but from different categories and companies, or all being financial content but published by different authors and containing no advertising content, it indicates that the recommendation system may have a cold start over-push problem, but there is no manipulation of the recommended content.

[0089] This application presents a locally benign recommendation system as ground truth. By comparing the differences in recommended content between the tested system and the benign system during the cold start phase, it assesses whether the tested system exhibits manipulation issues. It also addresses two technical problems: first, how to conduct cross-system comparisons and assess the differences in recommended content between different systems when each recommendation system acquires different content resources, leading to potentially inconsistent final recommendation results; and second, how to compare these differences and assess whether the tested system exhibits malicious manipulation issues. To address the first problem, this application proposes using a neural network model to learn the latent features of the recommended content, thus ignoring the differences in specific content and comparing and analyzing the latent features. To address the second problem, this application uses the reconstruction error in the latent feature learning model to quantify the inconsistency between the tested system and the benign system's recommended content, and uses a clustering algorithm to analyze the content connections of recommendations with large reconstruction errors to determine whether the system exhibits malicious manipulation issues.

[0090] This application also proposes a cold start security risk detection device for a content recommendation system, including a processor and a memory. The memory stores a computer program, and when the computer program is executed by the processor, it implements the steps of the aforementioned cold start security risk detection method for a content recommendation system.

[0091] This application also proposes a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the aforementioned content recommendation system cold start security risk detection method.

[0092] It should be noted that, in the embodiments of this application, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0093] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0094] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0095] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims. All of these forms are within the protection scope of this application.

Claims

1. A method for detecting cold start security risks in a content recommendation system, characterized in that, include: The network model is trained in advance using the following method: For the recommendation system under test, a user profile is generated according to the registration rules of the recommendation system under test, and the corresponding users are constructed in the local system based on the generated user profile; Configure an appropriate interaction strategy for any user, and perform the interaction on the local system based on the configured interaction strategy; Vectorize the interaction results to construct training data; The training data is input into the latent feature learning model, and training is performed based on a given reconstruction error threshold; To complete the detection of the data, follow these steps: The trained latent feature learning model is input into the content data to be detected; The reconstruction error of the output results of the latent feature learning model is statistically analyzed and reverse-sorted. Take a specified number of output data points after sorting as outliers and perform clustering. Based on the clustering results, determine whether there are security risks associated with the cold start of the content recommendation system.

2. The cold start security risk detection method for content recommendation systems as described in claim 1, characterized in that, Based on the generated user profile, the corresponding user is built on the local system, including: Based on the generated user profile, create an identical user on the local system.

3. The cold start security risk detection method for content recommendation systems as described in claim 1, characterized in that, The configured interaction strategies include random interaction and / or fixed interaction, wherein: The random interaction refers to user interactions that do not involve direct interaction with the recommendation system. The fixed interaction refers to setting a set of interaction trajectories for any user that match their profile.

4. The method for detecting cold start security risks in a content recommendation system as described in claim 1, characterized in that, Based on the configured interaction policy, performing interactions on the local system includes: Select one or more users created in the local system, with each user corresponding to a latent feature learning model; Control any user to interact with the local recommendation system according to the interaction strategy, and record the recommended content returned by the local system.

5. The cold start security risk detection method for content recommendation systems as described in claim 4, characterized in that, Based on the configured interaction strategy, performing interactions on the local system further includes: For any returned recommended content, the text content is not recorded. Instead, the recommended content is vectorized to construct training data for any user.

6. The method for detecting cold start security risks in a content recommendation system as described in claim 1, characterized in that, Selecting a specified number of output data points after sorting as outliers and performing clustering includes: A bag-of-words model is pre-built, and a specified number of output data after sorting are taken as abnormal data and input into the bag-of-words model to vectorize the abnormal data; The vectorized outlier data is clustered using K-means to obtain the clustering results.

7. The cold start security risk detection method for content recommendation systems as described in claim 6, characterized in that, Based on the clustering results, determining whether a content recommendation system has security risks during cold start includes: If all recommended content in a cluster belongs to a certain type or publisher, then the tested recommendation system is judged to be maliciously manipulated.

8. A cold start security risk detection device for a content recommendation system, characterized in that, It includes a processor and a memory, wherein the memory stores a computer program, and when the computer program is executed by the processor, it implements the steps of the content recommendation system cold start security risk detection method as described in any one of claims 1 to 7.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the content recommendation system cold start security risk detection method as described in any one of claims 1 to 7.

Images