Safety management framework based on cross-modal semantic understanding

Abstract

The invention discloses a security management framework based on cross-modal semantic understanding, a content input and output module carries out centralized management and analysis on cultural contents, an integrated security management and disposal system is utilized to convert processed information into visual output, an AI preprocessing module converts multimedia input into a standardized data format, and the AI preprocessing module carries out cross-modal semantic understanding. The AI identification module performs comprehensive scanning on the standardized multimedia content by using a pre-trained deep learning model, the large model reasoning module maps data of different modalities such as images, audios and texts to the same high-dimensional feature space through a vectorization technology, and the rule matching module performs rule matching on the basis of a knowledge base and a rule set. According to the method, content security assessment is automatically executed and judgment is carried out, the IaaS layer provides computing power, storage space, network configuration, security protection and container services, the level and capability of cultural content security management can be effectively improved, and the functions of media asset full content coverage, refined auditing, intelligent choice and cross-modal retrieval are achieved.

Description

Technical Field

[0001] The present application relates to the technical field of content security and management, and in particular, to a security management framework based on cross-modal semantic understanding. Background Art

[0002] In the media industry, the traditional content review method based on keyword matching can no longer meet the requirements of multimedia content security in existing media production. The following problems will arise when using the traditional content review method based on keyword matching: First, the limitations of keyword matching rely on a preset series of sensitive vocabulary libraries to screen inappropriate content, but the rapid development and ambiguous expressions of network language make it difficult for the keyword library to exhaust all potential illegal expressions. Therefore, it is easy to miss detection, that is, some malicious or illegal content escapes screening due to changes in expression. At the same time, overly broad or imprecise keyword settings may lead to false positives, affecting the release and dissemination of normal content. Second, the difficulty of identifying non-text content. Non-text media such as images, audio and video contain rich information and may contain potential illegal elements, such as violence, pornography, and terrorist symbols, which are difficult to capture directly through text analysis.

[0003] In summary, traditional methods often require manual review, which is inefficient and costly, and in a fast-spreading network environment, the timeliness is greatly reduced.

[0004] Therefore, how to improve the content review method in the existing technology to make it meet the requirements of existing media production for multimedia content security has become a research direction for technical personnel in this field. Summary of the invention

[0005] The purpose of this application is to provide a security management framework based on cross-modal semantic understanding in order to overcome the existing technical defects. Through AI and cross-modal semantic understanding technology, it solves the problems existing in content security management in the traditional media industry, improves the level and ability of cultural content security management, and provides solid technical guarantee for multimedia content security in the media industry.

[0006] The purpose of this application is achieved through the following technical solutions:

[0007] In the first aspect, the present application proposes a security management framework based on cross-modal semantic understanding, the framework comprising a content input and output module, a security management and disposal system, and an IaaS layer;

[0008] Content input and output module, which is used for centralized management and analysis of cultural content, including but not limited to video and audio file analysis, image content recognition, electronic newspaper and periodical material review, and comprehensive multimedia manuscript evaluation. It uses an integrated security management and disposal system to convert processed information into a form of output, including classification, annotation and interpretation of identified risk content, and intelligent report generation;

[0009] The security management and disposal system includes AI preprocessing module, AI recognition module, large model reasoning module and rule matching module. The AI ​​preprocessing module converts multimedia input into a standardized data format. The AI ​​recognition module uses a pre-trained deep learning model to comprehensively scan the standardized multimedia content and intelligently identify and classify risk scenarios. The large model reasoning module uses vectorization technology to map data of different modalities to the same high-dimensional feature space. The rule matching module automatically performs content security assessment and makes judgments based on the knowledge base and rule set.

[0010] The IaaS layer includes computing resource pool, storage resource pool, network resource pool, security resource pool and container resource pool, which are used to provide computing power, storage space, network configuration, security protection and container services.

[0011] In one possible implementation, the AI ​​preprocessing module converts multimedia input into a single image, a set of sequenced images, a set of unrelated images, audio content, and text content through full-media multi-dimensional conversion technology.

[0012] In a possible implementation, the AI ​​recognition module performs logo recognition, face recognition, scene recognition, map classification, OCR recognition on a single image, a set of images with a sequence, and a set of unrelated images based on the AI ​​preprocessed data, performs ASR recognition on audio content, and performs segmentation / sentence processing on text content;

[0013] Conduct security review of all media content based on risk scenarios, with review granularity including pornography, violence and terrorism, politics, sensitive flags, bad artists, custom faces, bad behavior, and geography. Figure 2 Classification, disgusting pictures, advertisements, logos and black image databases; risk scenarios include pornography, politics, violence and terrorism, contraband, advertisements and low quality.

[0014] In one possible implementation, the large model reasoning module establishes an information retrieval mechanism through vectorized representation of images, audio, and semantics, and uses a vector space model to associate cross-modal information.

[0015] In one possible implementation, the rule matching module builds a knowledge base, supports adding custom faces, custom text, custom behavioral actions, custom audio information, and blacklists according to tenants, adjusts audit policies as needed, selects audit categories based on audit scenarios, and supports adjusting audit item confirmation thresholds and suspected thresholds.

[0016] In one possible implementation, after obtaining the semantic vectorized data, the large model inference module determines whether it is necessary to detect the existence of illegal content in the semantic blacklist. If a comparison is required with the sentence blacklist, the semantic vectorized data is calculated for similarity with each sensitive word vector in the sentence blacklist. Based on the similarity threshold, it is determined whether the input semantic vectorized data is similar to the sensitive words in the blacklist, and cross-modal semantic retrieval of the text is performed to determine whether the semantic vectorized data contains sensitive word information.

[0017] In a possible implementation, the AI ​​preprocessing module is further used for:

[0018] When the multimedia input is a single type of business object, a single type of business object analysis service is executed. If the multimedia input is audio or image, it is input into the AI ​​recognition module. If the multimedia input is video, it is converted to low code, separated from video and audio, and input into the AI ​​recognition module.

[0019] When the multimedia input is a rich text service, rich text service processing is performed. If the rich text service is a multimedia manuscript, in-depth processing is performed using the HTML parsing service. If the rich text service is an e-newsletter, the e-newsletter is processed through the e-newsletter script rule parsing service, and text and images are extracted from the e-newsletter for preprocessing.

[0020] In a possible implementation, the AI ​​recognition module is further used to receive the preliminarily preprocessed video, text, picture, and audio for classification processing, transcode the video and extract frames into a picture set, convert the audio into text, and standardize the format of the picture and text;

[0021] Apply AI models to images for recognition, including logo recognition, face recognition, scene recognition, and map classification;

[0022] Convert audio content into text through speech recognition technology.

[0023] In a possible implementation, the large model reasoning module is further used to:

[0024] The images are converted into embedding vectors through the deep learning model, the texts are embedded into words through the deep learning model, the semantic features of each word or sentence are extracted, and the audio data is extracted and converted into vectors reflecting the audio characteristics;

[0025] Integrate the information obtained through recognition and reasoning into structured data, including but not limited to facial information, behavioral actions, and audio information;

[0026] Based on structured data, compare with the black database to identify potential illegal content or sensitive information;

[0027] Outputs the classification information, coordinates and timestamp of the offending content, and the sensitivity score.

[0028] The above-mentioned main scheme of the present application and its further options can be freely combined to form multiple schemes, all of which are schemes that can be adopted and claimed for protection in the present application; and in the present application, (non-conflicting options) options and other options can also be freely combined. After understanding the scheme of the present application, those skilled in the art can understand that there are multiple combinations based on the prior art and common knowledge, all of which are technical schemes to be protected by the present application, and they are not exhaustively listed here.

[0029] The present application discloses a security management framework based on cross-modal semantic understanding. The content input and output module centrally manages and analyzes cultural content, and uses an integrated security management and disposal system to convert the processed information into a formal output. The AI ​​preprocessing module converts multimedia input into a standardized data format. The AI ​​recognition module uses a pre-trained deep learning model to comprehensively scan the standardized multimedia content. The large model reasoning module maps data of different modalities such as images, audio and text to the same high-dimensional feature space through vectorization technology. The rule matching module automatically performs content security assessment and makes judgments based on the knowledge base and rule set. The IaaS layer provides computing power, storage space, network configuration, security protection and container services, which can effectively improve the level and ability of cultural content security management, and realize the functions of full content coverage, refined review, intelligent decision-making and cross-modal retrieval of media assets. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0031] Figure 1 A schematic diagram of a security management framework based on cross-modal semantic understanding proposed in an embodiment of the present application is shown.

[0032] Figure 2 A flow chart showing the management and handling methods of the security management framework proposed in an embodiment of the present application is shown. DETAILED DESCRIPTION

[0033] The following describes the embodiments of the present application through specific examples, and those skilled in the art can easily understand other advantages and effects of the present application from the contents disclosed in this specification. The present application can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present application. It should be noted that the following embodiments and features in the embodiments can be combined with each other without conflict.

[0034] Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making any creative work shall fall within the scope of protection of this application.

[0035] In the media industry, the traditional content review method based on keyword matching can no longer meet the requirements of existing media production for multimedia content security. On the one hand, the keyword matching method is prone to missed detection or false positives; on the other hand, for non-text content (such as images, audio and video), traditional review methods often have difficulty in accurately identifying potential risk factors. Therefore, how to improve the content review method in the existing technology to meet the requirements of existing media production for multimedia content security has become a research direction for technical personnel in this field.

[0036] In order to solve the above technical problems, the embodiment of the present application proposes a security management framework based on cross-modal semantic understanding, which uses the full-media multi-dimensional conversion technology to convert the input data into the form of a single picture, a set of sequenced pictures, a set of unrelated pictures, audio content and text content. Through an AI reasoning model that includes six major application scenarios and more than 40 audit details, combined with a multi-tenant knowledge base, content security control and intelligent judgment are achieved, effectively improving the level and ability of cultural content security management, and realizing the functions of full content coverage, refined audit, intelligent decision-making, and cross-modal retrieval of media assets, which will be described in detail below.

[0037] Please refer to Figure 1 , Figure 1 A schematic diagram of a security management framework based on cross-modal semantic understanding proposed in an embodiment of the present application is shown, the framework including a content input and output module, a security management and disposal system, and an IaaS layer;

[0038] Content input and output module, which is used for centralized management and analysis of cultural content, including but not limited to video and audio file analysis, image content recognition, electronic newspaper and periodical material review, and comprehensive multimedia manuscript evaluation. It uses an integrated security management and disposal system to convert processed information into a form of output, including classification, annotation and interpretation of identified risk content, and intelligent report generation;

[0039] The security management and disposal system includes AI preprocessing module, AI recognition module, large model reasoning module and rule matching module. The AI ​​preprocessing module converts multimedia input into a standardized data format. The AI ​​recognition module uses a pre-trained deep learning model to comprehensively scan the standardized multimedia content and intelligently identify and classify risk scenarios. The large model reasoning module uses vectorization technology to map data of different modalities such as images, audio and text to the same high-dimensional feature space. The rule matching module automatically performs content security assessment and makes judgments based on the knowledge base and rule set.

[0040] The IaaS layer includes computing resource pool, storage resource pool, network resource pool, security resource pool and container resource pool, which are used to provide computing power, storage space, network configuration, security protection and container services.

[0041] The content input and output module serves as the interface of the system. It is responsible for collecting various cultural content (video, audio, pictures, e-newspapers, multimedia manuscripts, etc.), and after processing by the integrated security management and disposal system, it outputs results that are easy to identify and understand, including classification annotation, explanation and intelligent reporting of risk content.

[0042] AI preprocessing module: converts multimedia data into a standardized format. The AI ​​recognition module uses deep learning models to identify sensitive information in the content, such as pornography, politics, violence and terrorism, to achieve automatic review. The large model reasoning module uses cross-modal vectorization to achieve efficient correlation retrieval between different data types and enhance the depth of understanding. The rule matching module automatically performs content security assessment based on the knowledge base and preset rules to improve judgment accuracy. By comprehensively considering multi-dimensional information, the AI ​​judgment mechanism can respond quickly in complex content environments, reducing the need for manual intervention and improving overall work efficiency.

[0043] The IaaS layer provides the necessary infrastructure services to ensure the stable operation of the system, covering computing, storage, network, security and container services, and providing flexible and reliable technical support for upper-layer applications. The computing resource pool, storage resource pool, network resource pool, security resource pool and container resource pool can all be implemented through general computing devices, which can be integrated into a single computing device or distributed in multiple computing devices and connected through the network.

[0044] The content input and output module not only analyzes the content, but also takes actions based on the analysis results, such as automatically blocking illegal content, notifying administrators, or executing other preset security policies. In this way, a closed loop is formed from detection to response, which improves processing efficiency. At the same time, the analysis results are presented in a user-friendly form for quick identification and understanding, including visual reports, highlighted sensitive content, or concise summaries, ensuring that users can grasp the key points without going into technical details.

[0045] The large model reasoning module uses advanced cross-modal reasoning technology to associate different types of data (such as pictures, videos, text, and audio), and can also achieve accurate retrieval in non-text data. The integration of features improves the comprehensive capabilities of the system, enabling it to adapt to a variety of application scenarios, such as content review, intelligence analysis, personalized recommendations, etc., to meet the needs of different user groups. It can also simplify complex operations, provide instant feedback, and ensure efficient processing, so that users can enjoy a smooth and fast service experience, whether they are content creators, auditors or ordinary users.

[0046] The AI ​​preprocessing module converts multimedia input into single images, sequenced image collections, unrelated image collections, audio content, and text content through full-media multi-dimensional conversion technology.

[0047] The AI ​​preprocessing module splits the video into single-frame images or a collection of images in a time series, making it easier to analyze each frame individually or understand the video content in sequence, and organizes the original images into unrelated collections of images in a unified format for subsequent processing. Convert the audio content into an analyzable audio file format, retaining the original audio information while adapting the analysis process. Extract text content from electronic newspapers and multimedia manuscripts and convert it into structured text in preparation for text analysis. Through efficient conversion algorithms, it ensures that different types of content can be processed consistently, solving the limitations of traditional single-modality processing technology when facing complex multimedia content.

[0048] The AI ​​recognition module performs logo recognition, face recognition, scene recognition, map classification, OCR recognition, ASR recognition of audio content, and segmentation / sentence processing of text content on single images, sets of images with sequences, and sets of unrelated images based on the AI ​​preprocessed data;

[0049] Conduct security review of all media content based on risk scenarios, with review granularity including pornography, violence and terrorism, politics, sensitive flags, bad artists, custom faces, bad behavior, and geography. Figure 2 Classification, disgusting pictures, advertisements, logos and black image databases, risk scenarios include pornography, politics, violence and terrorism, contraband, advertisements, and low quality. By subdividing application scenarios and customizing audit rules, the system can more accurately identify potential risk points in different scenarios, improving the pertinence and accuracy of the audit.

[0050] AI recognition calls the AI ​​atomic capabilities of the intelligent middle platform for review. It supports the access of AI atomic capabilities of different brands, has the ability to uniformly schedule various AI operators, provides the ability to send specific messages to specific addresses at specific times, and provides the ability to query service logs and business information statistics.

[0051] Perform multi-dimensional analysis on single images, sets of sequential images, and sets of unrelated images, including logo recognition to detect brand logos, face recognition to confirm individual identities, scene recognition to analyze image background and activities, map classification to determine geographic information, and OCR technology to extract text content from images.

[0052] Through ASR (automatic speech recognition) technology, the audio is converted into readable text for further text analysis and review.

[0053] Divide text content into segments or sentences to facilitate more detailed semantic understanding and risk screening.

[0054] For all media content, the AI ​​recognition module conducts fine-grained reviews of more than 40 risk types, including but not limited to pornography, violence and terrorism, politics, sensitive symbols, public figures with bad records, specific faces, inappropriate behavior, map violations, disgusting images, illegal advertising, copyright-infringing logos, and blacklisted images, to ensure the compliance and security of the content.

[0055] The large model reasoning module establishes an information retrieval mechanism through vectorized representation of images, audio and semantics, and uses vector space models to associate cross-modal information.

[0056] Images, audio and semantic information are converted into high-dimensional vectors, and multimedia data that is difficult to compare intuitively is converted into a format that can be calculated and compared in mathematical space. Then, using the vector space model, high-dimensional vectors are used to bridge different types of media data and achieve cross-modal association. Even when facing non-textual data (such as pictures and audio), it can effectively perform retrieval and matching, which improves the breadth and depth of information processing.

[0057] The rule matching module builds a knowledge base and supports adding custom faces, custom texts, custom behavioral actions, custom audio information, and blacklists according to tenants. It adjusts the audit strategy according to needs, selects the audit category according to the audit scenario, and supports adjusting the confirmation threshold and suspected threshold of the audit item.

[0058] Customized face is the facial recognition information of a specific individual, which is used to meet the needs of different customers for monitoring specific people. Customized text is a specific keyword or phrase that helps identify text content that may involve sensitive topics or proper nouns. Customized behavior actions define certain specific behavior patterns, such as gestures and action sequences, to identify potential violations. Customized audio information represents specific audio clips or keywords to strengthen the targeted monitoring of audio content. Black Library maintains a specific database of taboo content to strictly filter out content that is not allowed to appear.

[0059] In addition, this module also supports flexible adjustment of strategies according to audit scenarios, including selecting audit categories and setting confidence thresholds (confirmation thresholds and suspected thresholds) for different audit items, thereby realizing an audit mechanism that can both strictly control and adapt to different sensitivity requirements.

[0060] After obtaining the semantic vectorized data, the large model inference module determines whether it is necessary to detect the existence of illegal content in the semantic blacklist. If it is necessary to compare with the sentence blacklist, the semantic vectorized data is calculated with the vectors of each sensitive word in the sentence blacklist. The similarity threshold is used to determine whether the input semantic vectorized data is similar to the sensitive words in the blacklist, and cross-modal semantic retrieval of the text is performed to determine whether the semantic vectorized data contains sensitive word information.

[0061] First, evaluate whether the data needs to be sent to the semantic blacklist for further inspection. The decision is based on preset rules, context, or user-defined strategies. If it is determined that detection is necessary, match the semantic vectorized data with the sensitive word vectors in a pre-built sentence blacklist library, which contains the semantic representations of various sensitive words. By calculating the similarity between vectors, such as using cosine similarity and other methods, the proximity between the input data and the sensitive words in the blacklist library is quantified. Based on the pre-set similarity threshold, it is determined whether the input semantic vectorized data has a sufficiently high similarity with one or some sensitive words in the blacklist to confirm whether it contains sensitive information. It can effectively identify and mark possible illegal text content, and complete cross-modal semantic retrieval and sensitive content screening of text.

[0062] In a possible implementation, the AI ​​preprocessing module is further used for:

[0063] When the multimedia input is a single type of business object, a single type of business object analysis service is executed. If the multimedia input is audio or image, it is input into the AI ​​recognition module. If the multimedia input is video, it is converted to low code, separated from video and audio, and input into the AI ​​recognition module.

[0064] When the multimedia input is a rich text service, rich text service processing is performed. If the rich text service is a multimedia manuscript, in-depth processing is performed using the HTML parsing service. If the rich text service is an e-newsletter, the e-newsletter is processed through the e-newsletter script rule parsing service, and text and images are extracted from the e-newsletter for preprocessing.

[0065] In a possible implementation, the AI ​​recognition module is further used to receive the preliminarily preprocessed video, text, picture, and audio for classification processing, transcode the video and extract frames into a picture set, convert the audio into text, and standardize the format of the picture and text;

[0066] Apply AI models to images for recognition, including logo recognition, face recognition, scene recognition, and map classification;

[0067] Convert audio content into text through speech recognition technology.

[0068] In a possible implementation, the large model reasoning module is further used to:

[0069] The images are converted into embedding vectors through the deep learning model, the texts are embedded into words through the deep learning model, the semantic features of each word or sentence are extracted, and the audio data is extracted and converted into vectors reflecting the audio characteristics;

[0070] Integrate the information obtained through recognition and reasoning into structured data, including but not limited to facial information, behavioral actions, and audio information;

[0071] Based on structured data, compare with the black database to identify potential illegal content or sensitive information;

[0072] Outputs the classification information, coordinates and timestamp of the offending content, and the sensitivity score.

[0073] Figure 2 A flow chart of the management and disposal method of the security management framework proposed in the embodiment of the present application is shown, wherein cultural content is pre-processed by AI through content input to complete the sequence splitting and type classification of cultural content. The content is analyzed, transformed and vectorized according to the AI ​​algorithm model and the cross-modal large model. Then, through the rule matching of the face information library, text information library, behavior action library, audio information library, custom library and black library, the analysis and understanding of the cultural content is completed. Finally, a cross-modal information retrieval framework and the export of understanding information after AI analysis are constructed to achieve security control and intelligent judgment of the content, thereby effectively improving the level and ability of cultural content security management. The management and disposal methods are as follows:

[0074] 1. On the content input page, this embodiment supports the input types of single-type business objects and rich-text business objects. Among them, single-type business objects include videos, audios, and pictures, and rich-text business objects include multimedia manuscripts and e-newspapers. It should be noted that rich-text business objects contain at most multiple single-type business objects such as videos, audios, and pictures. This embodiment mainly provides three content input methods: manual uploading of local files, input of business object online addresses, and interface calls. It should be noted that this embodiment constructs AI preprocessing as the input content processing center, and all input content will enter AI preprocessing. After the content is entered, the business object type of the input content will first be determined by the intelligent business object analysis service to determine whether it belongs to a single class or a rich text class.

[0075] 2. When the content input is determined to be a single-type business object, the single-type business object analysis service in the AI ​​preprocessing will further process the input content. In one embodiment, the single-type business object analysis service will automatically generate corresponding metadata information for each single-type business object, including whether the type of the input content is video, audio, or image.

[0076] 3. If the input content type of a single-category business object is determined to be audio, the audio content is used as the input source to enter AI recognition & reasoning.

[0077] 4. If it is determined that the input content type of a single type of business object is a video, the video will be converted to a low-code format first; through the preset configuration, the input video will be converted into the encoding and packaging format with the highest efficiency and compatibility for the subsequent AI reasoning process. For the video material after low-code conversion, the audio and video separation process will continue, and the video screen will be output as a video track, and the audio part will be output as an audio track. It should be further explained that the processing mode of the video content is to extract key frames according to the established rules by extracting frames every 5 frames of the video; it supports increasing or decreasing the frame extraction frequency, and correspondingly, the overall system resource overhead and the accuracy of the final content recognition results will change accordingly. In one embodiment, all key frames extracted from the same input video will become a set of pictures with a sequence as the input source to enter AI recognition & reasoning. The audio track output will use the extracted audio content as the input source to enter AI recognition & reasoning.

[0078] 5. If the input content type of a single business object is determined to be a picture, the single picture is used as the input source to enter AI recognition & reasoning.

[0079] 6. The rich text business object analysis service in AI preprocessing will further process the input content. In one embodiment, the rich text business object analysis service will automatically generate corresponding metadata information for each rich text business object, including the type of the input content, such as multimedia manuscript or electronic newspaper.

[0080] 7. When the content input is judged to be a rich text business object, if the specific type is further judged to be a multimedia manuscript, the multimedia manuscript will be processed through the HTML parsing service; in this embodiment, multimedia manuscript content such as video, text, pictures, etc. are usually defined by different HTML tags. Each tag has its specific function, which enables the multimedia content to be organized in a structured manner. The multimedia manuscript is split according to these tags to extract specific types of elements. It should be further explained that <video>The elements in the tag will be extracted as a video for output. The elements in the tag will be extracted and output as images. and <h>Elements within tags are extracted as text for output; input content is further preprocessed.

[0081] 8. If the specific type is determined to be an electronic newspaper, the electronic newspaper will continue to be processed through the electronic newspaper script rule parsing service; the independent text and image will be extracted from the electronic newspaper document through the script as input content for further preprocessing.

[0082] 9. After the initial AI preprocessing of rich text business objects, the input content will be organized into video, text, and pictures; the video will be further preprocessed and converted to a low bit rate service; the obtained low bit rate video will be extracted to the video track, and the frames will be extracted into a collection of pictures with a sequence, and the audio track will be output as audio content. Unrelated picture collections, text content, audio content, and picture collections with sequences are used as input sources to enter AI recognition & reasoning.

[0083] 10. After AI preprocessing is completed, AI recognition & reasoning will receive a message that preprocessing is complete and accept the input content. First, the input source data is received and aggregated for classification processing. By dividing the input source into a set of pictures with sequences, a set of unrelated pictures, a single picture, audio content, and text content, the next step is driven and the above content is input into the corresponding AI algorithm model.

[0084] 11. In the AI ​​algorithm model step, the content type provided by the input source will be identified first, and different types of content will be assigned to the corresponding algorithm model. In this embodiment, the audio content will first undergo ASR recognition to output text content; this example constructs an intelligent processing flow for a sequence of picture sets, an unrelated picture set, and a single picture. First, the input picture is preprocessed, resized, cropped, rotated, etc., so that the algorithm can better recognize features. Use the Logo recognition model to output the location and category information of the Logo; use the face recognition model to detect faces in the image, and identify or verify them, including face detection, feature extraction, and comparison with known faces in the database; use the scene recognition model to output the scene type or description represented by the image. If further explanation is required, if the image contains a map, the map classification model will output the map type (such as a topographic map, a traffic map, etc.) and key information on the map. After all image recognition tasks are completed, the results are integrated and processed to form a structured data set. This data set is segmented and sentence-processed according to the meaning relevance with the ASR recognition output results and the text content in the input source to obtain a process text content set; this set continues to undergo NLP structured processing to identify entities in the text, analyze the sentiment of the text, understand the semantics of the text, extract keywords from the text, identify the relationship between entities in the text and generate a text summary; the processed data is integrated into a structured format and output to rule matching.

[0085] 12. In this example, the AI ​​pre-processed data includes single images, serialized images, text, and audio, which will be output to the large model for inference at the same time. If the input content type is determined to be an image, it will be further determined whether it is a single image or a serialized image. By performing feature extraction (Embedding) on ​​the image, the local features in the image are captured, and finally a vector of a fixed length is generated to represent the image, and the vectorized data result is output. If an image sequence is being processed, each image will undergo the above process separately and obtain its own Embedding vector. Combined with the attention mechanism, different weights can be assigned to the features of different images to highlight important image information and output a vectorized set result.

[0086] 13. If the input content type is determined to be text, the input text content will continue to be cleaned and segmented, stop words will be removed, and stemming or word form restoration will be performed. The results will be subjected to feature extraction (Embeddings). Each word or sentence can be converted into a fixed-length vector, which captures the semantic information of the word and its role in the context, and finally outputs the vectorized data results.

[0087] 14. If the input content type is determined to be audio, the input audio content will be subjected to sampling rate conversion and noise reduction. Through feature extraction, each audio clip will be converted into a vector of fixed length, and the vectorized data result will be output.

[0088] 15. In this embodiment, when processing structured data, it is necessary to judge the data type. The data type judgment can help us understand the data and select the corresponding matching rules. The structured data output by AI recognition and reasoning is divided into three categories: face information, behavior action, and audio information.

[0089] 16. If it is structured data of facial information, take the personId in the structured data, compare it with the face ID in the default face database, and return the classification information of the face in the structured data in the default database.

[0090] 17. If it is structured data of behavioral action, the action type label in the structured data is taken, compared with the action database, and the information on whether the action in the structured data is illegal is returned.

[0091] 18. If it is structured data of audio information type, take the audio information feature value in the structured data, compare it with the audio base library, and return the sound feature attributes in the structured data.

[0092] 19. The results obtained from 16, 17, and 18 above are filtered according to the threshold set by the system. Different data types have different thresholds set. If the result does not meet the threshold, the structured data is directly discarded. If it meets the threshold, it is determined whether it needs to be compared with the custom library according to the audit template set by the system. If the custom library is not needed, the result in (24) is output.

[0093] The audit template supports different customizations for associating different audit type data. If there is a custom library associated with the configuration, it means that the structured data needs to be customized. If there is no relevant custom library configured, then the custom library is not required.

[0094] 20. If it needs to be processed by a custom library, different custom libraries are selected for comparison according to the data type. For facial data, the face id in the returned result is matched with the face id in the custom face library, and the classification information corresponding to the face is output; for text data, the text data to be processed is compared with the custom text library, and similarity, pattern matching and other algorithmic techniques are used to output the sensitive information contained in the text data; for audio data, the audio feature data is matched with the custom audio library, and pornographic and violent information in the audio is output.

[0095] 21. In this embodiment, the semantic vectorized data obtained through large model reasoning first needs to be judged whether it needs to be passed through the semantic black library to further detect whether there is any illegal content. If the semantic black library is not needed, the result of the subsequent 24 is directly output. If the semantic black library is needed, the type of black library needs to be determined. Different types of semantic vectorized data need to be compared with different semantic black libraries, and judgment and selection are made according to system settings.

[0096] 22. If a comparison with a sentence blacklist library is required, the obtained semantic vectorized data will be compared with the sensitive word vectors in the sentence blacklist library, and the similarity threshold will be used to determine whether the input semantic vectorized data is similar to the sensitive words in the blacklist library, thereby realizing cross-modal semantic retrieval of text and determining whether the semantic vectorized data contains sensitive word information.

[0097] 23. If a black database of behavioral action images is required, the obtained semantic vectorized data will be similar to each sensitive image vector in the black database of behavioral action images, and the similarity threshold will be used to determine whether the input semantic vectorized data is similar to the sensitive actions in the black database, thereby realizing multimodal retrieval of image similarity and determining whether the semantic vectorized data contains sensitive action information.

[0098] 24. The final output results include violation classification information, coordinates and timestamp information corresponding to the violation content, and sensitivity score.

[0099] Compared with the prior art, the embodiments of the present application have the following beneficial effects:

[0100] By building a method to manage the security of cultural content using AI technology and cross-modal semantic understanding. This method first converts the format and types of cultural content through AI preprocessing, and then uses the AI ​​recognition module to conduct in-depth analysis of images, audio, text and other content, covering multiple aspects from logo recognition to voice transcription, and can conduct detailed security reviews for a variety of potential violations such as pornography, political sensitivity, violence and terrorism. Furthermore, cross-media information retrieval is achieved through large model reasoning, which enhances the system's efficiency in processing diverse data. In addition, the rule matching mechanism allows flexible adjustment of audit strategies according to regulatory needs and builds a personalized audit knowledge base to meet the specific needs of different customers. The entire system architecture is designed based on a cloud computing environment, making full use of computing, storage, network and other resource pools, which not only improves the efficiency of cultural content security management, but also ensures the high flexibility and scalability of services.

[0101] The above description is only a preferred embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present application should be included in the protection scope of the present application.< / h> < / video>

Claims

1. A security management framework based on cross-modal semantic understanding, characterized in that: The framework includes content input and output modules, security management and disposal systems, and IaaS layers; Content input and output module, which is used for centralized management and analysis of cultural content, including but not limited to video and audio file analysis, image content recognition, electronic newspaper and periodical material review, and comprehensive multimedia manuscript evaluation. It uses an integrated security management and disposal system to convert processed information into a form of output, including classification, annotation and interpretation of identified risk content, and intelligent report generation; The security management and disposal system includes AI preprocessing module, AI recognition module, large model reasoning module and rule matching module. The AI ​​preprocessing module converts multimedia input into a standardized data format. The AI ​​recognition module uses a pre-trained deep learning model to comprehensively scan the standardized multimedia content and intelligently identify and classify risk scenarios. The large model reasoning module uses vectorization technology to map data of different modalities to the same high-dimensional feature space. The rule matching module automatically performs content security assessment and makes judgments based on the knowledge base and rule set. The IaaS layer includes computing resource pool, storage resource pool, network resource pool, security resource pool and container resource pool, which are used to provide computing power, storage space, network configuration, security protection and container services.

2. The security management framework according to claim 1, characterized in that: The AI ​​preprocessing module converts multimedia input into single images, sequenced image collections, unrelated image collections, audio content, and text content through full-media multi-dimensional conversion technology.

3. The security management framework according to claim 1, characterized in that: The AI ​​recognition module performs logo recognition, face recognition, scene recognition, map classification, OCR recognition, ASR recognition of audio content, and segmentation / sentence processing of text content on single images, sets of images with sequences, and sets of unrelated images based on the AI ​​preprocessed data; All media content is subject to security review based on risk scenarios. The review granularity includes pornography, violence and terrorism, politics, sensitive flags, bad artists, custom faces, bad behavior, binary map classification, disgusting pictures, advertisements, logos, and black image libraries. Risk scenarios include pornography, politics, violence and terrorism, contraband, advertisements, and low quality.

4. The security management framework according to claim 1, characterized in that: The large model reasoning module establishes an information retrieval mechanism through vectorized representation of images, audio and semantics, and uses vector space models to associate cross-modal information.

5. The security management framework according to claim 1, characterized in that: The rule matching module builds a knowledge base and supports adding custom faces, custom texts, custom behavioral actions, custom audio information, and blacklists according to tenants. It adjusts the audit strategy according to needs, selects the audit category according to the audit scenario, and supports adjusting the confirmation threshold and suspected threshold of the audit item.

6. The security management framework according to claim 1, characterized in that: After obtaining the semantic vectorized data, the large model inference module determines whether it is necessary to detect the existence of illegal content in the semantic blacklist. If it is necessary to compare with the sentence blacklist, the semantic vectorized data is calculated with the vectors of each sensitive word in the sentence blacklist. The similarity threshold is used to determine whether the input semantic vectorized data is similar to the sensitive words in the blacklist, and cross-modal semantic retrieval of the text is performed to determine whether the semantic vectorized data contains sensitive word information.

7. The security management framework according to claim 1, characterized in that: The AI ​​preprocessing module is also used to: When the multimedia input is a single type of business object, a single type of business object analysis service is executed. If the multimedia input is audio or image, it is input into the AI ​​recognition module. If the multimedia input is video, it is converted to low code, separated from video and audio, and input into the AI ​​recognition module. When the multimedia input is a rich text service, rich text service processing is performed. If the rich text service is a multimedia manuscript, in-depth processing is performed using the HTML parsing service. If the rich text service is an e-newsletter, the e-newsletter is processed through the e-newsletter script rule parsing service, and text and images are extracted from the e-newsletter for preprocessing.

8. The security management framework according to claim 7, characterized in that: The AI ​​recognition module is also used to receive the preliminarily preprocessed videos, texts, pictures, and audios for classification processing, transcode the videos and extract frames into picture sets, convert the audios into texts, and standardize the formats of pictures and texts; Apply AI models to images for recognition, including logo recognition, face recognition, scene recognition, and map classification; Convert audio content into text through speech recognition technology.

9. The security management framework according to claim 8, characterized in that: The large model inference module is also used to: The images are converted into embedding vectors through the deep learning model, the texts are embedded into words through the deep learning model, the semantic features of each word or sentence are extracted, and the audio data is extracted and converted into vectors reflecting the audio characteristics; Integrate the information obtained through recognition and reasoning into structured data, including but not limited to facial information, behavioral actions, and audio information; Based on structured data, compare with the black database to identify potential illegal content or sensitive information; Outputs the classification information, coordinates and timestamp of the offending content, and the sensitivity score.

Images