Information content security filtering system based on deep learning and multi-modal fusion

Abstract

The invention discloses an information content security filtering system based on deep learning and multi-modal fusion, and relates to the technical field of information content security filtering. The working process of the system comprises the following steps: acquiring historical filtering data of a first information type, determining a periodic weight and comprehensively considering a plurality of factors to calculate a first filtering recommendation degree; for the information source subjected to filtering, the recommendation degree is calculated according to the method; otherwise, calculating a second filtering recommendation degree based on the number of information content occurrences and the number of second information type filtering times. The system also calculates an applicable filtering degree to determine the applicability of the third information type to the target content; and finally, the system pushes a corresponding filtering instruction to a filtering execution module according to the recommendation degree and the applicable filtering degree. According to the method, multiple factors such as historical data, time periods and filtering modules are comprehensively considered, the filtering accuracy is improved, and flexible processing is performed aiming at different information source conditions; the applicable filtering degree is calculated, pertinence and effectiveness of a filtering instruction are ensured, and information content safety is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of information content security filtering, and specifically relates to an information content security filtering system based on deep learning and multimodal fusion. Background Art

[0002] With the rapid development of the Internet and the explosive growth of information, cyberspace is flooded with a large number of different types of information content. This information includes not only legal and beneficial content, but also some information that is not suitable for public dissemination, such as advertisements, violence and pornography, politically sensitive information, rumors, false advertisements, etc. How to efficiently and accurately filter these bad information and ensure the safety and health of the network environment has become an urgent problem to be solved.

[0003] Traditional information filtering methods mainly rely on technical means such as keyword matching and rule engines. Although these methods can identify and filter some bad information to a certain extent, they have the following limitations: insufficient accuracy, poor adaptability, and single modality limitation.

[0004] In recent years, deep learning technology, especially convolutional neural networks, recurrent neural networks and their variants, has made significant progress in image recognition, natural language processing and other fields. At the same time, multimodal data fusion technology can more comprehensively describe information features and improve the accuracy of classification and recognition by integrating data from different sources (such as text, images, audio, etc.). However, the existing information filtering system based on deep learning and multimodal fusion still faces some challenges: Lack of dynamic adjustment mechanism: Most systems use static models for prediction and fail to fully consider the changing patterns in the time dimension.

[0005] Ignoring the value of historical data: Many systems only focus on immediate data and fail to fully utilize historical filtered data to optimize the decision-making process.

[0006] Low efficiency of collaboration between modules: There is a lack of effective coordination mechanism between different types of filtering modules, which affects the overall performance.

[0007] In order to overcome the above problems, researchers began to explore a new generation of information content security filtering systems that combine deep learning and multimodal fusion technology. This type of system aims to achieve more intelligent and accurate identification and filtering of bad information by analyzing historical filtering data and dynamically adjusting weight parameters. Specifically, it requires the system to be able to process not only text information, but also multiple media types such as images and videos, and to flexibly adjust strategies according to different application scenarios.

[0008] Therefore, this embodiment proposes an information content security filtering system based on deep learning and multimodal fusion, which calculates the first filtering recommendation degree, the second filtering recommendation degree and the applicable filtering degree by comprehensively considering factors such as statistical cycle, time period, filtering log data and filtering module, and pushes the corresponding filtering instructions to the filtering execution module. This innovation not only improves the accuracy and efficiency of information filtering, but also provides strong support for building a more secure and reliable network environment. Summary of the invention

[0009] In order to overcome the shortcomings and deficiencies of the above-mentioned prior art, the present invention adopts the following technical solutions: The information content security filtering system based on deep learning and multimodal fusion has the following workflow: Obtaining historical filtering data of the first information type: extracting the first filtering times of the first information type in each statistical period of N statistical periods included in the historical period from the log database, where the historical period is the period before the filtering decision moment, and each statistical period includes M time periods arranged in sequence; Determine the cycle weight: Determine the cycle weight of each statistical cycle according to the cycle interval between the statistical cycle and the statistical cycle at the filtering decision moment, and the cycle weight is negatively correlated with the cycle interval; Calculating a first filtering recommendation degree: determining a first filtering recommendation degree for a first information type; Processing according to different information sources: if the first information source has been filtered before the filtering decision moment, it is processed according to the method of calculating the first filtering recommendation degree; if the first information source has not been filtered before the filtering decision moment, relevant data is obtained and the second filtering recommendation degree of the second information type is calculated; Calculating the applicable filtering degree: if the total number of first information contents subjected to filtering operations in the historical period is Z, obtaining first filtering execution data, generating second filtering execution data, and calculating the applicable filtering degree of the third information type for the target information content in the first information source; Push filtering instructions.

[0010] Preferably, calculating the first filtering recommendation degree includes: Obtaining cycle weights and filtering log data: obtaining the cycle weight of each statistical cycle in the historical period, and obtaining first filtering log data including a source identifier of a first information source, a type identifier of a first information type, a filtering time of the first information type, a statistical cycle in which the filtering time is located, and a module identifier of a filtering module that performs filtering; Calculating the first filtering times in each statistical period: determining the total number of first filtering log data in each statistical period at the filtering time as the first filtering times of the first information type in the statistical period; Generate extended filtering log data: if the first filtering times of the first information type in a certain statistical period is greater than or equal to 1, then generate second filtering log data corresponding to M-1 other time periods one by one; if multiple first filtering log data include R different module identifiers, R>1, then generate R-1 groups of third filtering log data corresponding to R-1 other module identifiers one by one; Determine the time period weight: According to the interval between the time period of the filtering decision moment and the time period within the statistical period, use the exponential decay function to calculate the time period weight of each time period; Determine the module weight of the filtering module: set the module weight of the filtering module indicated by the module identifier in the first filtering log data and the second filtering log data to a first value, and the module weight of the filtering module indicated by the module identifier in the third filtering log data to a second value, and the first value is greater than the second value; Calculate the first filtering recommendation degree: determine the first filtering recommendation degree of the first information type according to the period weights of N statistical periods, the time period weights of the time periods corresponding to each filtering log data, and the module weights of the filtering modules indicated by the module identifiers in each filtering log data.

[0011] Preferably, in processing according to different information source situations, if the first information source has not been subjected to filtering operation before the filtering decision moment, the system is further configured to: Obtaining the number of occurrences of each type of first information content in a historical period, the third number of filtering times of the second information type in the historical period, and the fourth number of filtering times of each type of second information content in a plurality of types of second information content corresponding to the filtered second information type in the historical period; Based on the intersection of the types of the multiple types of first information content and the multiple types of second information content, at least one type of reference information content is obtained; Based on the number of occurrences of at least one type of reference information content, the fourth filtering number and the third filtering number, the second filtering recommendation degree of the second information type is calculated, and the second filtering recommendation degree is positively correlated with the product of the number of occurrences of each type of reference information content and the fourth filtering number, and negatively correlated with the third filtering number.

[0012] Preferably, in calculating the applicable filtering degree, the system is further configured to: Acquire multiple pieces of first filtering execution data of the first information source within a historical period, each piece of the first filtering execution data including a type identifier of an information type, a module identifier of an auxiliary filtering module that performs filtering of the information type, and the number of times the auxiliary filtering module performs filtering of the information type within the historical period; For each piece of first filtering execution data, determine Z-1 pieces of second filtering execution data based on the data, each piece of second filtering execution data including a type identifier of the information type in the first filtering execution data, a module identifier of another first filtering module, and an execution count in the first filtering execution data; The applicable filtering degree of the third information type to the target information content in the first information source is determined based on the product of the number of executions in each target filtering execution data and the module weight of the filtering module indicated by the module identifier, the module weight of the filtering module indicated by the module identifier in the first filtering execution data is a first value, the module weight of the filtering module indicated by the module identifier in each second filtering execution data is a second value, and the first value is greater than the second value.

[0013] Preferably, the push filtering instruction includes: If the first filtering recommendation degree of the first information type is greater than the first filtering recommendation degree threshold, pushing an instruction to filter the first information type to the filtering execution module after the filtering decision moment; If the second filtering recommendation degree is greater than the second filtering recommendation degree threshold, then after the filtering decision moment, an instruction to filter the second information type is pushed to the filtering execution module; If the applicable filtering type request for the target information content is received from the filtering request module, and it is determined that the applicable filtering degree of the third information type for the target information content is greater than the applicable filtering degree threshold, an instruction to filter the third information type is pushed to the filtering execution module.

[0014] Preferably, an exponential decay function is used to calculate the time period weight of each time period, specifically: let the current time period be t, and the time period of the filtering decision moment be , then the time period weight ;in, is the attenuation coefficient that can be adjusted according to actual conditions.

[0015] Preferably, the calculation of the first filtering recommendation degree satisfies the formula: ; Wherein, Wn is the period weight of the nth statistical period in N statistical periods, and n is an integer greater than or equal to 1 and less than or equal to N; The time period weight corresponding to the t-th filtered log data in the n-th statistical period; The module weight of the filtering module indicated by the module identifier in the t-th filtering log data in the n-th statistical period, t is greater than or equal to 1 and less than or equal to An integer.

[0016] Preferably, the second filtering recommendation degree of the second information type is calculated to satisfy the formula: ;in, is the number of occurrences of the i-th type of reference information content; is the fourth filtering order of the i-th category of reference information content; is the third filtering number of the second information type, is the number of categories of reference information content.

[0017] Preferably, the calculation of the applicable filtering degree satisfies the formula: ,in, is the number of executions in the jth target filtering execution data, is the corresponding module weight, and m is the number of target filter execution data In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: 1. The present invention calculates the first filter recommendation by using historical data and multiple factors (such as cycle weight, time period weight and module weight), and generates extended filter log data (second and third filter log data) according to different information types. The system can more accurately predict the potential threats of specific information types. This multi-level data analysis method not only considers the time and frequency of information occurrence, but also considers the distribution of information in different time periods and filter modules, thereby improving the accuracy and reliability of filtering decisions.

[0018] 2. The present invention provides corresponding processing strategies for different information sources (first information sources that have or have not performed filtering operations). In particular, for information sources that have not performed filtering operations, the reference information content is determined by analyzing the intersection of its content and other known information types, and the second filtering recommendation degree is calculated accordingly. This method makes the system highly flexible and adaptable, and can effectively deal with various types of information sources and information content, thereby enhancing the scope of application and processing capabilities of the system.

[0019] 3. The present invention sets a reasonable filter recommendation threshold and adjusts it according to actual needs. The system can automatically send filter instructions to the execution module after the filter decision moment, ensuring timely and effective interception of bad information. In addition, according to different information types, targeted filter instruction push strategies are set (such as instruction push corresponding to the first, second and third information types), which further improves the efficiency and pertinence of the filtering work. This method can not only respond to new threats quickly, but also ensure that users have a cleaner and safer information environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0021] Figure 1 The working flow chart of the information content security filtering system based on deep learning and multi-modal fusion of the present invention is shown; Figure 2 A flowchart showing the first filtering recommendation degree calculation of the present invention is shown; Figure 3 A flow chart of filtering instruction push according to the present invention is shown. DETAILED DESCRIPTION

[0022] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0023] In addition, the described features, structures or characteristics may be combined in one or more example embodiments in any suitable manner. In the following description, many specific details are provided to provide a full understanding of the example embodiments of the present disclosure. However, those skilled in the art will appreciate that the technical solutions of the present disclosure may be practiced while omitting one or more of the specific details, or other methods, components, steps, etc. may be adopted. In other cases, well-known structures, methods, implementations or operations are not shown or described in detail to avoid obscuring various aspects of the present disclosure.

[0024] Embodiment 1: See also Figure 1 As shown, the information content security filtering system based on deep learning and multimodal fusion in this embodiment has the following workflow: Step 1: Obtain basic data and determine periodic weights.

[0025] S11. Obtain historical filtering data of the first information type.

[0026] Extract the first filtering times of the first information type (such as advertisement, violent pornography, politically sensitive, etc.) in each statistical period in the historical period (before the filtering decision moment) in N statistical periods (each statistical period is, for example, one week, N is the past 12 weeks, N>1) from the system's log database. The log database records detailed information of each information filtering operation, including information type, filtering time, etc.

[0027] Each statistical period includes M time periods arranged in sequence (for example, each time period is one day, M = 7, M> 1).

[0028] S12. Determine the cycle weight.

[0029] The cycle weight of each statistical cycle is determined based on the cycle interval between the statistical cycle and the statistical cycle at the filtering decision moment, and is negatively correlated with the cycle interval. It can be calculated using a linear attenuation function. Suppose the current statistical cycle is n, and the statistical cycle at the filtering decision moment is , then the period weight .

[0030] Step 2: Calculate the first filter recommendation. (Comprehensively consider factors such as statistical cycle, time period, filter log data, and filter module) See also Figure 2 As shown, the process of calculating the first filtering recommendation degree is as follows: S21. Obtain period weight and filter log data.

[0031] Get the cycle weight of each statistical cycle in the historical period , obtain filtering log data: obtain multiple first filtering log data received in the historical period, each of which contains the source identifier of the first information source (such as website domain name, social media platform ID, etc.), the type identifier of the first information type (such as 1 for advertising, 2 for violence and pornography, etc.), the filtering time of the first information type, the statistical period of the filtering time, and the module identifier of the filtering module that performs filtering (such as A for text filtering module, B for image filtering module, etc.). These log data are generated when the filtering execution module filters the first information type content in the first information source.

[0032] S22. Calculate the first filtering times in each statistical period.

[0033] For each statistical period within the N statistical periods, the total number of first filtered log data in the statistical period at the filtering time is determined as the first filtering number of the first information type in the statistical period. .

[0034] S23: Generate extended filtering log data.

[0035] S231, generating second filtering log data: for each statistical period within N statistical periods, if the first filtering times of the first information type within the statistical period If it is greater than or equal to 1, then based on each piece of first filtering log data received in the statistical period, M-1 pieces of second filtering log data corresponding to M-1 other time periods are obtained. These second filtering log data contain the source identifier of the first information source, the type identifier of the first information type, the statistical period in the first filtering data, and the same module identifier as the corresponding first filtering log data, which is used to consider the potential occurrence of information in different time periods.

[0036] S232, generating third filtering log data: if the plurality of first filtering log data include R different module identifiers (R>1), based on each first filtering log data received within the statistical period, obtain R-1 groups of third filtering log data corresponding to R-1 other module identifiers. Each group includes M third filtering log data corresponding to M time periods, each including the source identifier of the first information source, the type identifier of the first information type, the statistical period of the first filtering log data, and the module identifier of the corresponding filtering module.

[0037] S24. Determine the time period weight.

[0038] The time period weight of each time period is negatively correlated with the time period interval. The time period interval is the interval between the time period at the time of the filtering decision and the time period within the statistical period. Using the exponential decay function to calculate, let the current time period be t, and the time period at the time of the filtering decision be , then the time period weight ;in, is the attenuation coefficient that can be adjusted according to actual conditions.

[0039] S25. Determine the module weight of the filtering module.

[0040] The module weight of the filter module indicated by the module identifier in the first filter log data and the second filter log data is a first value V1 (such as 0.8), the module weight of the filter module indicated by the module identifier in the third filter log data is a second value V2 (such as 0.2), and .

[0041] S26: Calculate the first filtering recommendation degree.

[0042] According to the cycle weight of N statistical cycles , obtained in each statistical period The time period weight of each filtered log data in the filtered log data , and the module weight of the filtering module indicated by the module identifier in each filtered log data , determine the first filtering recommendation degree of the first information type .

[0043] ; Wherein, Wn is the period weight of the nth statistical period in N statistical periods, and n is an integer greater than or equal to 1 and less than or equal to N; The time period weight corresponding to the t-th filtered log data in the n-th statistical period; The module weight of the filtering module indicated by the module identifier in the t-th filtering log data in the n-th statistical period, t is greater than or equal to 1 and less than or equal to An integer.

[0044] Step 3: Handling different information sources.

[0045] a. The first information source has performed a filtering operation.

[0046] If the first information source has been filtered before the filtering decision moment, the first filtering times of the first information type in each of the N statistical periods included in the historical period are obtained and calculated according to the above method for calculating the first filtering recommendation degree.

[0047] b. The first information source does not perform filtering operations.

[0048] Data acquisition: Get the number of occurrences of each type of first information content (such as text, image, video, etc.) in a historical period.

[0049] Get the third filtering times of the second information type (such as rumors, false advertisements, etc.) in the historical period.

[0050] Obtain the fourth filtering times of each type of second information content being filtered in a historical period among multiple types of second information content corresponding to the filtered second information type (such as text rumors, image rumors, etc. corresponding to the rumor type).

[0051] Determine the content of the reference information: Based on the intersection of the types of the multiple types of first information content and the multiple types of second information content, at least one type of reference information content is obtained. For example, if the first information source contains text and image content, and the filtered content in the second information type also contains text and image, then the text and image are the reference information content.

[0052] Calculate the second filter recommendation: Based on the number of occurrences of at least one type of reference information content and the fourth filter times , and the third filter times , determine the second filtering recommendation degree of the second information type , D2 is positively correlated with the product of the number of occurrences of each type of reference information content and the fourth filtering number, and negatively correlated with the third filtering number. The calculation formula is ;in, is the number of occurrences of the i-th type of reference information content; is the fourth filtering order of the i-th category of reference information content; is the third filtering number of the second information type, is the number of categories of reference information content.

[0053] Step 4: Apply filtration degree calculation.

[0054] If the total number of first information contents that have been filtered in the historical period is Z (Z>1): Get the first filter execution data A plurality of first filtering execution data of the first information source in a historical period are obtained, each of which includes a type identifier of an information type, a module identifier of an auxiliary filtering module that performs filtering of the information type, and the number of times the auxiliary filtering module performs filtering of the information type in the historical period.

[0055] Generate the second filtering execution data: For each piece of first filtering execution data, Z - 1 pieces of second filtering execution data are determined based on the data, each piece of which includes a type identifier of the information type in the first filtering execution data, a module identifier of another first filtering module, and the number of executions in the first filtering execution data. Another first filtering module is a module other than the auxiliary filtering module indicated by the module identifier in the first filtering execution data among the Z first filtering modules that perform filtering operations in the historical period.

[0056] Calculate the applicable filtration degree: Determine the applicable filtering degree S of the third information type for the target information content in the first information source. The applicable filtering degree is positively correlated with the product of the number of executions Cj in each target filtering execution data and the module weight Vj of the filtering module indicated by the module identifier. The calculation formula is: ,in, is the number of executions in the jth target filtering execution data, is the corresponding module weight, and m is the number of target filtering execution data.

[0057] The module weight of the filter module indicated by the module identifier in the first filter execution data is a first value (such as 0.8), and the module weight of the filter module indicated by the module identifier in each second filter execution data is a second value (such as 0.2), and the first value is greater than the second value.

[0058] Step 5: Filter instruction push.

[0059] See also Figure 3 As shown in the figure, the process of pushing filtering instructions is as follows: S51. Pushing a first information type filtering instruction.

[0060] If the first filtering recommendation degree of the first information type is greater than the first filtering recommendation degree threshold (which can be adjusted according to the actual needs of the system and historical data), an instruction to filter the first information type is pushed to the filtering execution module (such as an information firewall, content review engine, etc.) after the filtering decision moment, so as to filter the first information type content in the first information source (such as a website, a social media platform, a news client, etc.).

[0061] S52: Pushing a second information type filtering instruction.

[0062] If the second filtering recommendation degree is greater than the second filtering recommendation degree threshold, an instruction to filter the second information type is pushed to the filtering execution module after the filtering decision moment.

[0063] S53: Pushing a third information type filtering instruction.

[0064] If a request for an applicable filtering type for target information content is received from a filtering request module (such as a user terminal, a management system, etc.), and it is determined that the applicable filtering degree of the third information type for the target information content is greater than the applicable filtering degree threshold, an instruction to filter the third information type is pushed to the filtering execution module.

[0065] The beneficial effects of this embodiment include: comprehensively considering multiple factors such as historical data, time period, and filtering modules to improve filtering accuracy; flexibly handling different information sources; calculating applicable filtering degrees to ensure the pertinence and effectiveness of filtering instructions and improve information content security.

[0066] The weights of the present invention are used to measure the degree of influence of different factors or variables on a certain result or decision. The definition of weight refers to the value assigned to each factor when comparing and evaluating multiple factors to reflect its importance or priority. These weights can be determined according to specific circumstances and needs, and are usually jointly formulated and confirmed by professionals or relevant stakeholders. By setting weights reasonably, it can help programs or systems make decisions or predictions more accurately.

[0067] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

[0068] The preferred embodiments of the present invention disclosed above are only used to help explain the present invention. The preferred embodiments do not describe all the details in detail, nor do they limit the invention to only specific implementation methods. Obviously, many modifications and changes can be made according to the content of this specification. This specification selects and specifically describes these embodiments in order to better explain the principles and practical applications of the present invention, so that those skilled in the art can understand and use the present invention well. The present invention is limited only by the claims and their full scope and equivalents.

Claims

1. The information content security filtering system based on deep learning and multimodal fusion is characterized by: The system workflow is as follows: Obtaining historical filtering data of the first information type: extracting from the log database the first filtering times of the first information type in each statistical period of N statistical periods included in the historical period, wherein the historical period is the period before the filtering decision moment, and each statistical period includes M time periods arranged in sequence; Determine the cycle weight: Determine the cycle weight of each statistical cycle according to the cycle interval between the statistical cycle and the statistical cycle at the filtering decision moment, and the cycle weight is negatively correlated with the cycle interval; Calculating a first filtering recommendation degree: determining a first filtering recommendation degree for a first information type; Processing according to different information sources: if the first information source has been filtered before the filtering decision moment, it is processed according to the method of calculating the first filtering recommendation degree; if the first information source has not been filtered before the filtering decision moment, relevant data is obtained and the second filtering recommendation degree of the second information type is calculated; Calculating the applicable filtering degree: if the total number of first information contents subjected to filtering operations in the historical period is Z, obtaining first filtering execution data, generating second filtering execution data, and calculating the applicable filtering degree of the third information type for the target information content in the first information source; Push filtering instructions.

2. According to claim 1, the information content security filtering system based on deep learning and multimodal fusion is characterized in that: The calculating of the first filtering recommendation degree comprises: Obtaining cycle weights and filtering log data: obtaining the cycle weight of each statistical cycle in the historical period, and obtaining first filtering log data including a source identifier of a first information source, a type identifier of a first information type, a filtering time of the first information type, a statistical cycle in which the filtering time is located, and a module identifier of a filtering module that performs filtering; Calculating the first filtering times in each statistical period: determining the total number of first filtering log data in each statistical period at the filtering time as the first filtering times of the first information type in the statistical period; Generate extended filtering log data: if the first filtering times of the first information type in a certain statistical period is greater than or equal to 1, then generate second filtering log data corresponding to M-1 other time periods one by one; if multiple first filtering log data include R different module identifiers, R>1, then generate R-1 groups of third filtering log data corresponding to R-1 other module identifiers one by one; Determine the time period weight: According to the interval between the time period of the filtering decision moment and the time period within the statistical period, use the exponential decay function to calculate the time period weight of each time period; Determine the module weight of the filtering module: set the module weight of the filtering module indicated by the module identifier in the first filtering log data and the second filtering log data to a first value, and the module weight of the filtering module indicated by the module identifier in the third filtering log data to a second value, and the first value is greater than the second value; Calculate the first filtering recommendation degree: determine the first filtering recommendation degree of the first information type according to the period weights of N statistical periods, the time period weights of the time periods corresponding to each filtering log data, and the module weights of the filtering modules indicated by the module identifiers in each filtering log data.

3. The information content security filtering system based on deep learning and multimodal fusion according to claim 1 is characterized in that: In the processing according to different information source situations, if the first information source has not been subjected to filtering operation before the filtering decision moment, the system is further used for: Obtaining the number of occurrences of each type of first information content in the historical period, the third number of filtering times of the second information type in the historical period, and the fourth number of filtering times of each type of second information content in the plurality of types of second information content corresponding to the filtered second information type being filtered in the historical period; Based on the intersection of the types of the multiple types of first information content and the multiple types of second information content, at least one type of reference information content is obtained; Based on the number of occurrences of at least one type of reference information content, the fourth filtering number and the third filtering number, the second filtering recommendation degree of the second information type is calculated, and the second filtering recommendation degree is positively correlated with the product of the number of occurrences of each type of reference information content and the fourth filtering number, and negatively correlated with the third filtering number.

4. The information content security filtering system based on deep learning and multimodal fusion according to claim 1 is characterized in that: In the calculation of applicable filtering degree, the system is further used for: Acquire multiple pieces of first filtering execution data of a first information source within a historical period, each piece of the first filtering execution data including a type identifier of an information type, a module identifier of an auxiliary filtering module that performs filtering of the information type, and the number of times the auxiliary filtering module performs filtering of the information type within the historical period; For each piece of first filtering execution data, determine Z-1 pieces of second filtering execution data based on the data, each piece of second filtering execution data including a type identifier of the information type in the first filtering execution data, a module identifier of another first filtering module, and an execution count in the first filtering execution data; The applicable filtering degree of the third information type to the target information content in the first information source is determined based on the product of the number of executions in each target filtering execution data and the module weight of the filtering module indicated by the module identifier, the module weight of the filtering module indicated by the module identifier in the first filtering execution data is a first value, the module weight of the filtering module indicated by the module identifier in each second filtering execution data is a second value, and the first value is greater than the second value.

5. The information content security filtering system based on deep learning and multimodal fusion according to claim 1 is characterized in that: The push filtering instruction includes: If the first filtering recommendation degree of the first information type is greater than the first filtering recommendation degree threshold, pushing an instruction to filter the first information type to the filtering execution module after the filtering decision moment; If the second filtering recommendation degree is greater than the second filtering recommendation degree threshold, then after the filtering decision moment, an instruction to filter the second information type is pushed to the filtering execution module; If the applicable filtering type request for the target information content is received from the filtering request module, and it is determined that the applicable filtering degree of the third information type for the target information content is greater than the applicable filtering degree threshold, an instruction to filter the third information type is pushed to the filtering execution module.

6. The information content security filtering system based on deep learning and multimodal fusion according to claim 2 is characterized in that: The exponential decay function is used to calculate the time period weight of each time period, specifically: let the current time period be t, and the time period of the filtering decision moment be , then the time period weight ;in, is the attenuation coefficient that can be adjusted according to actual conditions.

7. The information content security filtering system based on deep learning and multimodal fusion according to claim 1 is characterized in that: The calculation of the first filtering recommendation degree satisfies the formula: ; Wherein, Wn is the period weight of the nth statistical period in N statistical periods, and n is an integer greater than or equal to 1 and less than or equal to N; The time period weight corresponding to the t-th filtered log data in the n-th statistical period; The module weight of the filtering module indicated by the module identifier in the t-th filtering log data in the n-th statistical period, t is greater than or equal to 1 and less than or equal to An integer.

8. The information content security filtering system based on deep learning and multimodal fusion according to claim 1 is characterized in that: The calculation of the second filtering recommendation degree of the second information type satisfies the formula: ;in, is the number of occurrences of the i-th type of reference information content; is the fourth filtering order of the i-th category of reference information content; is the third filtering number of the second information type, is the number of categories of reference information content.

9. The information content security filtering system based on deep learning and multimodal fusion according to claim 1 is characterized in that: The calculation is applicable to the filtering degree that satisfies the formula: ,in, is the number of executions in the jth target filtering execution data, is the corresponding module weight, and m is the number of target filtering execution data.

Images