Multi-modal risk content identification method, device and system, and storage medium

Abstract

The invention discloses a multi-modal risk content identification method, device and system, and a storage medium. The method comprises the following steps: acquiring various modal data; and aligning the risk content of various modal data to a text modal through each modal conversion model to form a unified text abstract form, and meanwhile, carrying out similar case retrieval-based enhanced reasoning. By adopting the technical scheme of the invention, potential risks can be comprehensively and accurately identified, and hierarchical classification can be carried out on the potential risks.

Description

Technical Field

[0001] The present invention belongs to the technical field of risky content detection, and in particular, relates to a multimodal risky content identification method and device, system, and storage medium. Background Art

[0002] With the significant progress of deepfakes in recent years, the emergence of new risky content has posed a serious threat to national security and economic and social stability. However, existing content detection technologies have shown obvious limitations in dealing with increasingly complex, hidden, realistic, and diverse cross-modal risky content:

[0003] 1) They often focus on a single task and can only identify a specific risk, resulting in limited training samples for the model and limited generalization ability for atypical samples.

[0004] 2) It focuses on detecting the presence or absence of a certain risk, lacks the ability to conduct fine-grained classification of specific risk categories, and cannot accurately locate specific risk points.

[0005] 3) Lack of multimodal capabilities, unable to integrate cross-modal content in risk content for detection and analysis to form a comprehensive risk profile, thus limiting the ability to conduct comprehensive analysis of complex risk scenarios.

[0006] 4) It is not based on explicit content understanding. On the one hand, it is impossible to conduct a reasonable analysis of risk content, and on the other hand, it also makes risk detection insufficiently explainable.

[0007] Therefore, in order to overcome the above defects, it is necessary to improve the multi-risk and cross-modal generalization capabilities of the model while ensuring the accuracy of fine-grained classification. At present, in the field of risk content detection, there is a lack of a powerful, privately deployable, multi-modal and universal risk content detection and analysis platform. Summary of the invention

[0008] The technical problem to be solved by the present invention is to provide a multimodal risk content identification method and device, system, and storage medium that can comprehensively and accurately identify potential risks and classify them in a hierarchical manner.

[0009] To achieve the above object, the present invention adopts the following technical solution:

[0010] A multimodal risk content identification method, comprising:

[0011] Acquire multiple modal data;

[0012] The risk content of multi-modal data is aligned to the text modality through each modal conversion model to form a unified text summary form, and enhanced reasoning is performed based on similar case retrieval.

[0013] Preferably, the multiple modal data include: video data, text data, audio data and image data.

[0014] As a preferred approach, a framework that decouples perception and reasoning is adopted to uniformly convert multi-modal data into text descriptions through a multimodal perception layer, and then send the unified text descriptions to a RAG-centered retrieval and reasoning layer to complete similar case retrieval, risk summary generation, risk analysis, and risk avoidance recommendation tasks; wherein, the multimodal perception layer includes: a large visual text model, an OCR model, and a speech-to-text model.

[0015] The present invention also provides a multimodal risk content identification device, comprising:

[0016] Acquisition module, used to obtain multiple modal data;

[0017] The identification module is used to align the risk content of multi-modal data to the text modality through each modal conversion model to form a unified text summary form, and at the same time perform enhanced reasoning based on similar case retrieval.

[0018] Preferably, the multiple modal data include: video data, text data, audio data and image data.

[0019] Preferably, the recognition module adopts a framework of decoupling perception and reasoning, converts multiple modal data into text descriptions through a multimodal perception layer, and then sends the unified text descriptions to the RAG-centered retrieval and reasoning layer to complete similar case retrieval, risk summary generation, risk analysis, and risk avoidance recommendation tasks; among them, the multimodal perception layer includes: a large visual text model, an OCR model, and a speech-to-text model.

[0020] An embodiment of the present invention further provides a multimodal risk content identification system, comprising: a memory and a processor, wherein the memory stores a computer program executed by the processor, and the computer program executes a multimodal risk content identification method when executed by the processor.

[0021] An embodiment of the present invention further provides a storage medium, on which a computer program is stored, and the computer program executes the multimodal risk content identification method when running.

[0022] The present invention adopts multimodal risk content recognition based on large models to automatically identify and analyze multimodal and multi-type risk content. Based on the general understanding ability and modal fusion ability of large models, the platform combines a variety of advanced deepfake detection technologies at different levels, and comprehensively uses a variety of modern large model technologies such as retrieval enhancement generation technology, contextual learning, large model reasoning and planning, and efficient parameter fine-tuning. It can comprehensively identify and analyze risk content in various modalities such as video, text, image, audio, web page links and files, overcoming the shortcomings of previous methods such as insufficient support for multimodality and lack of explainable analysis capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] 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 embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0024] Figure 1 This is a flow chart of a multimodal risk content identification method according to an embodiment of the present invention. DETAILED DESCRIPTION

[0025] 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.

[0026] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0027] Embodiment 1:

[0028] like Figure 1 As shown, the embodiment of the present invention provides a multimodal risk content identification method, which adopts a framework of decoupling perception and reasoning, uniformly converts multiple modal data into text descriptions through a multimodal perception layer, and then sends the unified text descriptions to the RAG-centered retrieval and reasoning layer to complete similar case retrieval, risk summary generation, risk analysis, and risk avoidance recommendation tasks; wherein, the multiple modal data include: video data, text data, audio data, and image data; the multimodal perception layer includes: a large visual text model, an OCR model, and a speech-to-text model. Specifically including:

[0029] Step 1: Develop the Web front end and the Android front end

[0030] Based on the different needs of individual users, an Android APP and a Web front-end interface are developed; based on the relevant needs of enterprise users, a Web front-end interface is built. The Web user interface is built using Gradio.

[0031] The risk identification content output by the front end includes five contents: ① whether there is risk content, ② risk summary, ③ identified risk type, ④ risk content analysis, and ⑤ prevention suggestions.

[0032] For individual users, an APP is developed and deployed on mobile devices. Users keep the APP running in the background. When receiving text messages, the APP obtains the text messages received by the mobile phone in real time through the Android Studio broadcast mechanism, and realizes the transmission of data between the Java program (client) and the Python program (server) through FastAPI. According to the analysis results of the server, it is sent to the mobile client, so that users can check in real time whether the received message contains risks, suggestions for preventing risks, etc. according to the pop-up window of the APP. For the web side of individual users, according to the different types of information carriers, it is divided into risk identification of video, picture, text message, sound, web page, PDF, as well as customized digital human and face-changing functions.

[0033] For video risk identification, users can choose to upload a video, which will be analyzed and then output with risk identification content, pornography detection, video face forgery detection, sound forgery detection, and comprehensive forgery detection; for image risk identification, by uploading images, risk identification content, synthetic detection, and pornography detection will be output; for SMS risk identification, the risk points of risk content can be identified, for example: "You only need to download our APP and fill in relevant information." After detection, "Download our APP" and "Fill in relevant information" will be marked in red, indicating that there will be risk points here. In addition, it outputs risk identification content and similar case analysis; for sound risk identification, it analyzes the audio uploaded by the user, outputs whether the audio is forged and detects risk content; for web page risk identification, it outputs risk identification content and identifies whether the pictures contained in the web page contain risk content through the web page link given by the user; for PDF file risk identification, it gives two parts of text risk identification and image pornography detection based on the PDF file uploaded by the user; for customized digital human generation, users can upload face images, upload a template audio and the text they want to output, and after synthesis, they can output the target video; for the face-changing function, it supports uploading videos and target faces, replacing faces in videos, and outputting the replaced videos. In order to optimize the user experience, for some functions, examples are set up, and users only need to click to quickly upload examples for analysis and experience the output of risk content of the analyzed examples.

[0034] For corporate users, a comprehensive user front-end interface has been developed, which can not only upload information in different forms for analysis and output more detailed content, such as: segmenting fraud types and similar cases, but also count risky content and intuitively display the current risk identification results, i.e. the number of risky videos, the number of risky pictures, etc.

[0035] Gradio is an open source Python library for quickly building demos and web applications to demonstrate deep learning models and convert AI algorithms into visual interfaces.

[0036] Step 2: Develop the service distribution layer

[0037] Based on the content that needs to be transmitted between the front-end and service layer, and between the modules of the service layer, a service distribution layer based on FastAPI is built to be responsible for the interaction.

[0038] FastAPI is a modern, fast, high-performance web framework for building APIs. Using FastAPI in platform development can encapsulate sub-function logic and models that may be shared by multiple functions as RESTful APIs.

[0039] The interaction process in the platform includes interactions across environments, languages, and versions. That is, during the development process, the Python environments required by each sub-module and model are inconsistent, and some libraries and CUDA versions are incompatible. All models cannot be run in a unified environment. FastAPI can package sub-modules and models as services to support cross-environment interactions. This design also allows interaction between Java and Python, greatly reducing the difficulty of developing Android apps.

[0040] FastAPI supports pre-loading of models and is compatible with modern reasoning engines such as vllm. If the model needs to be imported into the GPU every time it is used, it will take a lot of time. However, using the FastAPI framework to package the model into an API port can import the model into the GPU in advance; when the model needs to be used, just make a request to the corresponding API port, and the server will call the model loaded into the GPU to quickly obtain the model output.

[0041] FastAPI enhances the scalability of the system. When replacing submodules and models, there is no need to modify other parts. When adding new functions and new models, you only need to develop and deploy new subservices without making major changes to the original system, making it easier for the system to continuously iterate and evolve.

[0042] Step 3: Build four risk identification services at the service layer

[0043] Based on the different carriers of risk information, risk content types can be roughly divided into video, text, image, and audio, and different risk identification service modules are developed for the four types. Therefore, the service layer includes the service distribution layer and the risk identification service module, and the risk identification service module includes video risk identification service, text risk identification service, image risk identification service, and audio risk identification service.

[0044] For the video risk identification service module, detection is mainly carried out from five aspects: ① Whether there is a risk in the video content, ② Whether the face in the video is replaced, ③ Whether there is pornographic content in the video, ④ Whether the audio signal of the video is forged, and ⑤ Whether there is a risk in the content of the audio signal of the video.

[0045] The main implementation process is as follows: extract ten frames from the video evenly, and then send the ten pictures to the large visual-text model Qwen2-VL to obtain the risk content of each picture. Then send the risk content of these extracted frames to the large language model Qwen2 to summarize the risk content of the entire video; send the video to the VideoFACT model to detect whether the face in the video has been replaced; send the extracted frame pictures to the opennsfw2 model to detect whether the video contains pornographic content; extract the audio signal of the video and send it to the DeepFake-Audio-Detection-MFCC model to determine whether the audio signal is forged, and send it to the whisper-large-v3 model to determine whether the audio content is risky.

[0046] For the image risk identification service module, detection is mainly carried out from five aspects: ① Whether the image itself is forged (including background, face, etc.), ② Whether the face in the image has been replaced by AI, ③ Whether the image itself has naked or pornographic parts, ④ For text images (Text Image), whether these texts themselves contain risk information, and ⑤ Whether the content of the image itself is risky.

[0047] The main implementation process is as follows: Use the open source model UniversalFakeDetect to help detect forged images, including the detection of various types of forged images, such as deepfake, styletransfer, traditional image tampering, etc.; Use the open source project Combining EfficientNet and VisionTransformers for Video Deepfake Detection to analyze a set of video frame images through temporal clues to detect face-swapped images; Refer to the open source model Opennsfw2, use deep learning technology to judge whether the image contains pornographic content by scoring the image; For text images, use optical character recognition (OCR), select the CnOcr model to extract text information, and then send it to the text channel for risk identification and judgment.

[0048] For the audio risk identification service module, detection is mainly carried out from two aspects: ① Whether the content of the audio contains fraud, pornography, or violence, and ② Whether the audio itself is a generated forged audio.

[0049] The main implementation process is as follows: the speech is converted into text through the speech-to-text model whisper-large-v3, and then output to the text channel for content risk identification and judgment; for audio forgery detection, the open source project DeepFake-Audio-Detection-MFCC is referred to, and a new audio forgery recognition dataset is constructed using part of the data in the Codecfake dataset. The original project is retrained to improve the model's ability to identify forged audio, especially audio generated by the current mainstream large models.

[0050] For the text risk identification service module, since text can be the type conversion target of all risk information, that is, pictures, videos, and audio can all be converted into text descriptions, the text detection module has become the most basic and important module among the four modules. Based on the need for text detection, the construction process mainly includes ① two-level text risk classification, ② text risk point extraction, ③ similar case retrieval and risk avoidance suggestion generation, ④ risk summary generation and risk analysis.

[0051] The main implementation process is as follows: use efficient parameter fine-tuning methods to fine-tune the instructions of the large model, use the embedding model to vectorize the input content, then train a classification head for classification, and directly fine-tune the pre-trained BERT model for text risk classification; refer to the common methods of keyword extraction, such as KeyBERT and KeyLLM, adopt the few-shot setting, design the prompt in a targeted manner, guide the model to complete the task through multiple rounds of dialogue, further enhance the instruction following degree of the model, and perform simple post-processing on the model return results to ensure the correct return of the results and realize the extraction of risk points; for the retrieval of similar cases and the generation of risk avoidance suggestions, use the classic framework of retrieval-augmented generation (RAG). First, the risk case library is naturally divided into blocks according to the case, compressed into embedded vectors through the embedding model (such as bge-large-zh-v1.5), and a single-layer or multi-layer index is established according to the data scale. After the input content of multiple modalities is converted into text through the multimodal perception layer, the embedding model is used to compress the input into vectors for retrieval. When the amount of risk case database data is small, the cosine value of the input text vector and the case vector is used as the distance metric to directly recall the top-k risk cases. When the amount of risk case data is large, the efficiency of direct linear scanning is low. First, the approximate nearest neighbor (ANN) algorithm is used to recall more approximate most similar cases, and then the Reranker model (such as bge-reranker-large) is used to fine-rank the recalled cases. Finally, the k cases with the highest relevance after reranking are used as the final similar cases and added to the input of the large language model to assist in the generation of risk avoidance suggestions; for risk summary generation and risk analysis, the user text input, the text description of the user's other modal input converted by the multimodal perception layer, the forgery / synthesis judgment results of each deepfake detection model for each modality of the input content, the coarse-grained risk classification results of the text, the detailed classification results of fraud risk, and the retrieved similar cases are spliced ​​in a certain format, and then the large model is used to unify and analyze the above content to give the final risk analysis conclusion and risk summary. Using the few-shot setting, the style of the model-generated content is controlled through multi-round dialogue guidance to improve the quality of risk summaries. Using the automatic evaluation method based on the large model, the reward model pre-trained with a large amount of human feedback is used to evaluate the quality of risk summaries and risk analyses generated by the large model on FGRC. Using instruction fine-tuning to train the model can generalize to different tasks in the same field.

[0052] At the same time, in order to enhance the system's self-evolutionary capabilities and cope with increasingly complex and hidden multimodal risk content, the platform can, with the user's permission, desensitize and store the user's text input, the text description of the non-text input converted by the multimodal perception layer, and the risk summary output by the corresponding model to build a risk case database. This case database can be used by RAG to enhance the quality of model output and improve the timeliness of the system by manually adding new cases. It can also be constructed as a data set after screening and cleaning for fine-tuning the model to further improve the model's ability to generate risk summaries and risk analysis tasks.

[0053] Step 4: Build the dataset

[0054] Based on the fact that the model requires a large amount of data set training, the platform uses a large language model (LLM) to generate the FGRC-SCD data set for conversation and SMS text fraud risk identification, which is mainly constructed using cases from the CCL 2023 telecommunications network fraud case classification evaluation. The FGRC-SCD data set that includes high-risk fraud texts and risk-free texts is generated through a carefully designed prompt, and the data set is streamlined through multiple effective data screening mechanisms, forming a screening data set with better evaluation performance for the training and evaluation of the conversation fraud detection system.

[0055] The initial data set used by the platform comes from the authoritative data set of CCL 2023 telecommunications network fraud case classification evaluation. This data set has been strictly screened and covers a wide range of telecommunications network fraud cases. It is highly representative and authoritative in the field of telecommunications fraud research. At the same time, in order to ensure the privacy rights of users, all cases have undergone preliminary desensitization processing.

[0056] For the data generation method, we used the initial data set, considered the cooperation between humans and AI, and used the powerful task model deepseek to recreate a data set that is closer to the task requirements; in addition, referring to the AttrPrompt method, we gave LLM multi-dimensional and random attribute prompts in the prompts as constraints to generate training data without the original data set or pre-trained task model. In addition, when generating data, we also used a multi-round dialogue system to guide and standardize the generation of the model, and used a multi-process parallel processor to batch synthesize data to improve the efficiency of data generation.

[0057] For data evaluation, in order to verify whether the generated dataset has sufficient reliability and validity, and to ensure the accuracy and reliability of the model training results and analysis, the quality of the generated dataset is quantified and evaluated from three aspects: data diversity, task relevance, and human preference score. Among them, the quantification of data diversity mainly uses word frequency distribution statistics and sample correlation calculations; the compute-score method of the BGE Reranker-large model is used to quantify the relevance between the generated dataset and the research question; the get-scores method of the InternLM2-7B-Reward model is used to evaluate the quality of the generated text by calculating the average human preference score of the dataset, including the relevance, coherence, innovation of the content, and whether it conforms to human values.

[0058] Finally, filtering is performed based on the generated data: In order to improve the quality of the generated data, the data is filtered based on four quantitative indicators including cosine similarity, Euclidean distance, question relevance, and human preference score, from the aspects of data diversity, data relevance to the task, and human preference score.

[0059] For the two indicators of cosine similarity and Euclidean distance, data screening is performed using the following two methods: Method 1 is a selection method based on row and matrix addition; Method 2 is a selection method based on maximum value selection. Task relevance and human preference rating are both positive indicators. They can be directly calculated for each piece of data in the original data set, and then the high quantile data can be selected as the screening data set.

[0060] After optimizing and screening the above multiple indicators layer by layer (the screening ratio of each layer is set to 50%), the final screening data set is obtained. Then, each indicator of the screening data set is evaluated and compared with the original data set.

[0061] Step 5: Database selection and construction

[0062] Since enterprises need to process large amounts of data, and these data can also be used by RAG to enhance the quality of model output, the platform uses a database for storage. Since large files such as images and videos are not suitable for storage in a database, only data in textual form is stored in the database.

[0063] In terms of database selection, MongoDB is selected as the database management system. On the one hand, for inputs of different modes, the content finally sent to the text module has different structures. MongoDB is a NoSQL database with flexible storage types. It can store data without defining a schema, and it is highly scalable and suitable for storing heterogeneous data. On the other hand, considering that on the enterprise side, massive amounts of data need to be processed continuously; and for the data in the database, we need to use RAG to enhance the quality of model output, so we need to query continuously. MongoDB supports efficient query and indexing on data, supports processing large-scale data and efficient concurrent access.

[0064] So far, the general functions of the platform have been realized. However, in order to improve the user experience, the platform still needs to improve its efficiency and accuracy to make the platform more perfect.

[0065] Step 6: System acceleration solution

[0066] Latency is an important indicator for measuring system performance and is also an important factor in user experience.

[0067] Currently, the commonly used delay indicator is the first token delay (Time To First Token, TTFT). The platform adopts a variety of optimization methods to reduce TTFT.

[0068] First, we need to analyze the critical path of the delay and determine that the main sources of delay are file IO, ffmpeg video processing, and large model reasoning. The delay of file IO mainly comes from the large number of image files after the video frame extraction, which need to interact between different modules through the file system. The solution avoids image IO by directly passing image objects as much as possible; for unavoidable IO, use multiple processes to reduce the delay caused by multiple IO. ffmpeg video processing can use GPU to accelerate the frame extraction process.

[0069] In terms of model reasoning acceleration, the solution uses the vllm reasoning engine to integrate multiple reasoning optimization methods. For models not supported by vllm, specific acceleration libraries or onnx are used for reasoning acceleration. Finally, we return the reasoning results in a streaming manner, further reducing TTFT and improving user experience.

[0070] Step 7: System robustness enhancement plan

[0071] Since the system integrates multiple sub-services, sub-modules, and models, and different components may need to communicate with each other, the system complexity is high and the probability of error is greatly increased. On the other hand, some models in the system are only trained on research data sets or smaller data sets, and may give wrong judgment results for out-of-distribution samples, which may cause the final output of the system to deviate greatly from the expected output.

[0072] Improving system robustness The primary task of the system is to allow the system to operate normally even when some components crash, and to try to recover by automatically restarting services, reloading models, etc. Based on the above situation, the solution is as follows:

[0073] (1) Separated services: Use a separated architecture to design sub-services to prevent the crash of a single sub-service from causing the system to crash.

[0074] (2) Exception input processing: Use exception handling mechanisms and explicit processing of null values, invalid values, etc. to prevent code crashes due to illegal input.

[0075] (3) Exponential delay restart: Use a daemon process to monitor the status of each service, adopt an exponential delay restart strategy to try to restart the crashed sub-service, and promptly report the services that cannot be restarted within the specified number of times under the maximum delay.

[0076] On the other hand, the system should be able to give relatively reasonable outputs when some models output incorrect results. The solution mainly uses the following two methods to achieve this goal:

[0077] (1) Model integration: For difficult tasks such as face-swapping detection in videos, multiple models with different focus domains (time domain, spatial domain, and frequency domain) are used to perform detection simultaneously, and the outputs are integrated into the final output in different ways according to the needs, so as to minimize the impact of single model errors on the system output and increase the system fault tolerance.

[0078] (2) Missing example strategy: In the unified modal reasoning stage, some items in some example inputs are randomly replaced with missing values ​​or wrong values, such as replacing the image text description in the example input with none, or replacing the audio forgery detection result in the example input with no detection or the opposite value, so that the large model has a certain tolerance and adaptability to the erroneous output of other models through context learning. Furthermore, after collecting more multimodal training data on the system, this strategy can be introduced into the training stage to further improve the robustness of the system by mixing a small number of missing samples into the training data.

[0079] The embodiments of the present invention have the following technical effects:

[0080] 1. The project functional framework can be established through risk identification and classification, summary generation, risk point marking, risk analysis, risk avoidance suggestions, etc.;

[0081] 2. The front end of the system can be built through the personal user interface including the Web end, Android end and the enterprise user interface of the Web end;

[0082] 3. You can use the FastAPI-based service forwarding layer, as well as video, text, image, and audio risk identification services, to build the system’s service layer;

[0083] 4. The system model layer can be constructed through the unified module management layer, multimodal perception layer and various risk detection models;

[0084] 5. Can be combined with CodeFormer (image face cropping and enhancement), GPT-SoVits (voice cloning), and V-Express (video generation) to achieve customized digital human generation technology;

[0085] 6. It can combine separated services, model integration, abnormal input processing, missing example strategy and exponential delay restart to enhance the robustness of the system;

[0086] 7. The vllm inference engine can be used to integrate multiple inference optimization methods, onnx can improve the inference speed, and improve the system acceleration solution;

[0087] 8. Use MongoDB database management system for data storage;

[0088] 9. It can restore text through case description, customize fraud through attribute prompts, guide the construction of data sets through multiple rounds of dialogue, and conduct data evaluation and screening;

[0089] 10. A smaller 7B model can be used, which is convenient for small and medium-sized enterprises and individual users to deploy on local servers;

[0090] 11. A series of system acceleration solutions and streaming return output risk content feedback can be provided to minimize the possible losses of users;

[0091] 12. Modular design and development can be used to reduce the coupling between modules and facilitate maintenance and development;

[0092] 13. It can realize flexible replacement of large models or APIs, support horizontal expansion and load balancing, facilitate flexible expansion of sub-services, and ensure the strong scalability of the platform;

[0093] 14. The platform's continuous evolution capability can be enhanced based on non-training updates, training data collection and system iteration;

[0094] 15. Based on advanced AI technology, the detection efficiency and accuracy can be improved while ensuring the safety and reliability of the technology.

[0095] The embodiments of the present invention use various large and small models to assist and combine to construct video, text, image, and audio risk identification modules, thereby further constructing a multimodal risk detection platform, which not only helps corporate users and individual users identify and avoid different types of risk content, but also enriches the user experience and meets their personalized needs through customized functions.

[0096] Embodiment 2:

[0097] The embodiment of the present invention further provides a multimodal risk content identification device, comprising:

[0098] Acquisition module, used to obtain multiple modal data;

[0099] The identification module is used to align the risk content of multi-modal data to the text modality through each modal conversion model to form a unified text summary form, and at the same time perform enhanced reasoning based on similar case retrieval.

[0100] As an implementation manner of an embodiment of the present invention, the multiple modal data include: video data, text data, audio data and image data.

[0101] As an implementation method of an embodiment of the present invention, the recognition module adopts a framework of decoupling perception and reasoning, and uniformly converts multiple modal data into text descriptions through a multimodal perception layer, and then sends the unified text description to the RAG-centered retrieval and reasoning layer to complete similar case retrieval, risk summary generation, risk analysis, and risk avoidance recommendation tasks; wherein, the multimodal perception layer includes: a large visual text model, an OCR model, and a speech-to-text model.

[0102] Embodiment 3:

[0103] An embodiment of the present invention further provides a multimodal risk content identification system, comprising: a memory and a processor, wherein the memory stores a computer program executed by the processor, and the computer program executes a multimodal risk content identification method when executed by the processor.

[0104] Embodiment 4:

[0105] An embodiment of the present invention further provides a storage medium, on which a computer program is stored, and the computer program executes the multimodal risk content identification method when running.

[0106] The embodiments described above are only descriptions of the preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. A multimodal risk content identification method, characterized in that: include: Acquire multiple modal data; The risk content of multi-modal data is aligned to the text modality through each modal conversion model to form a unified text summary form, and enhanced reasoning is performed based on similar case retrieval.

2. The multimodal risk content identification method according to claim 1, characterized in that: The multi-modal data includes: video data, text data, audio data and image data.

3. The multimodal risk content identification method according to claim 1, characterized in that: By adopting a framework of decoupling perception and reasoning, multi-modal data are uniformly converted into text descriptions through the multimodal perception layer, and then the unified text descriptions are sent to the RAG-centered retrieval and reasoning layer to complete similar case retrieval, risk summary generation, risk analysis, and risk avoidance recommendation tasks; among them, the multimodal perception layer includes: a large visual text model, an OCR model, and a speech-to-text model.

4. A multimodal risk content identification device, characterized in that: include: Acquisition module, used to obtain multiple modal data; The identification module is used to align the risk content of multi-modal data to the text modality through each modal conversion model to form a unified text summary form, and at the same time perform enhanced reasoning based on similar case retrieval.

5. The multimodal risk content identification device according to claim 4, characterized in that: The multi-modal data includes: video data, text data, audio data and image data.

6. The multimodal risk content identification device according to claim 5, characterized in that: The recognition module adopts a framework of decoupling perception and reasoning, converts multi-modal data into text descriptions through the multimodal perception layer, and then sends the unified text descriptions to the RAG-centered retrieval and reasoning layer to complete similar case retrieval, risk summary generation, risk analysis, and risk avoidance recommendation tasks; among them, the multimodal perception layer includes: a large visual text model, an OCR model, and a speech-to-text model.

7. A multimodal risk content identification system, characterized in that: include: A memory and a processor, wherein the memory stores a computer program executed by the processor, and when the computer program is executed by the processor, the multimodal risk content identification method according to any one of claims 1 to 3 is executed.

8. A storage medium, characterized in that: The storage medium stores a computer program, which executes the multimodal risk content identification method according to any one of claims 1 to 3 when running.

Images