Network media video sensitive information intelligent identification and early warning system

Abstract

The invention relates to the technical field of network content security, in particular to a network media video sensitive information intelligent identification and early warning system, which comprises a data acquisition module for acquiring network media video data; the video information extraction module performs multi-dimensional analysis on the acquired data to obtain depth features; the AIGC video detection module identifies whether the content is the content generated by the AIGC technology or not according to the depth feature, and if yes, an AI generation mark is marked; the causal inference and evaluation module is used for carrying out causal correlation diagnosis on intention and context, carrying out deep analysis through an anti-fact inference technology, judging whether the intention is sensitive information or not, evaluating dangerousness and obtaining a dangerousness level; and the early warning module takes the minimum intervention cost and the maximum risk control as targets, matches and dynamically generates an optimal intervention scheme from a strategy library in combination with the risk level, and intervenes the sensitive video. Therefore, the problems of incomplete multi-modal information analysis, lack of an AI generation content identification mechanism and the like are solved.

Description

Technical Field

[0001] This application relates to the field of network content security technology, and in particular to an intelligent identification and early warning system for sensitive information in network media videos. Background Technology

[0002] Online video media has become a core medium for the public to access information and entertainment. However, its openness and low barrier to entry for creation have also made it a breeding ground for sensitive and illegal content. The rapid iteration of AIGC technology has made AI-generated fake videos, synthesized speech, and other sensitive content increasingly realistic, posing a serious challenge to online content governance.

[0003] Traditional video content security control systems primarily rely on a "feature matching + manual review" model. This involves initial screening of content using predefined sensitive word libraries, image hash databases, or simple audio-visual templates, followed by pushing suspicious content to a manual review pool. This approach is not only inefficient and costly in terms of manpower, but also struggles to cope with rapidly changing content formats. It lacks the ability to deeply integrate and analyze multimodal features of video (visual, audio, and text), only capturing single-dimensional sensitive traces, making it prone to missed or false positives. Existing technologies generally remain at the surface-level feature matching level, failing to combine publisher metadata and content context for intent and causal relationship analysis, leading to frequent false positives and missed positives. Furthermore, there is no specific identification mechanism for AIGC-generated content, making it unable to effectively identify AI-synthesized fake sensitive content. Summary of the Invention

[0004] This application provides an intelligent identification and early warning system for sensitive information in online media videos to solve problems such as incomplete multimodal information analysis and lack of AI-generated content identification mechanisms in existing technologies.

[0005] This application provides a first aspect of an intelligent identification and early warning system for sensitive information in online media videos, comprising: a data acquisition module, a video information extraction module, an AIGC video detection module, a causal inference evaluation module, and an early warning module. The data acquisition module collects online media video data, including video stream data, bullet screen / comment text, and publisher metadata. The video information extraction module performs multi-dimensional analysis on the collected video stream data and bullet screen / comment text to obtain deep features, including visual, audio, and text dimensions. The AIGC video detection module identifies whether the content is generated by AIGC technology based on the deep features; if so, it marks it with an AI-generated tag. The causal inference evaluation module performs intent and contextual causal relationship diagnosis based on the deep features, AI-generated tags, and publisher metadata, performs deep analysis using counterfactual reasoning techniques, determines whether the information is sensitive, and evaluates the risk level. The early warning module, aiming for minimum intervention cost and maximum risk control, matches and dynamically generates an optimal intervention plan from a strategy library based on the risk level, and intervenes in the sensitive video according to the optimal intervention plan.

[0006] Optionally, the data acquisition module includes: a data acquisition unit and a data preprocessing unit. The data acquisition unit is used to crawl video stream data using web crawler technology and obtain bullet screen / comment text and publisher metadata through the platform's open API. The data preprocessing unit is used to clean and align the acquired data. For video stream data, it uses a frame quality assessment algorithm to filter valid frames, removing stuttering frames, duplicate frames, and black frames. It also uses an improved spectral subtraction method to suppress environmental noise and unifies the sampling rate and encoding format. For text data, it uses regular expressions to filter out garbled characters, meaningless characters, and spamming content. A multi-dimensional association index of video, text, and publisher metadata is established based on timestamps and video IDs to generate structured data with unified identifiers.

[0007] Optionally, the video information extraction module includes: a multi-dimensional parsing unit and a deep feature fusion unit. The multi-dimensional parsing unit is used to extract the spatial and temporal features of the video frame using a pre-trained video understanding model, extract the acoustic features and deep semantic features of the video sound using a pre-trained audio neural network, convert the video sound into text to obtain video subtitle information, process the video subtitle information and bullet screen / comment text to generate text semantic vectors, extract core topics through an LDA model, calculate sentiment polarity and sentiment intensity, and obtain text features. The deep feature fusion unit is used to concatenate the video frame, video sound, and text features to obtain deep features.

[0008] Optionally, the AIGC video detection module includes: an artifact feature detection unit, a semantic consistency verification unit, and a labeling unit. The artifact feature detection unit is used to capture AIGC generation traces from multiple dimensions and perform quantitative evaluation to obtain an artifact confidence score. Specifically, in the visual dimension, it detects common AIGC detail defects based on the spatial features of the video frame, identifies inter-frame motion anomalies based on the temporal features of the video frame, and uses an outlier detection algorithm to quantify the degree of artifacts, generating a visual artifact score. In the audio dimension, it determines whether AI-synthesized speech traces exist based on the acoustic features and deep semantic features of the video sound, generating an audio artifact score. In the text dimension, it identifies typical AIGC text defects based on the text semantic vector and core topics, obtaining a text artifact score. The overall multi-dimensional artifact detection score is then calculated. The system generates an artifact confidence score through weighted calculation. The semantic consistency verification unit verifies video-audio consistency, video-text consistency, and content-comment sentiment consistency, generating a semantic consistency score based on the verification results. Specifically, video-audio consistency verifies whether the video content matches the audio, video-text consistency verifies whether the video content and text semantic features are consistent, and content-comment sentiment consistency verifies whether the emotional tone of the video content matches the sentiment of the bullet comments / comments. The labeling unit outputs a binary decision result and an overall confidence score through a decision engine based on the artifact confidence score and the semantic consistency score. The binary decision result includes AI-generated and non-AI-generated results; if the binary decision result is AI-generated, an AI-generated label is used.

[0009] Optionally, the causal inference assessment module includes: an intent diagnosis unit, a causal analysis unit, and a risk assessment unit. The intent diagnosis unit extracts behavioral features, content features, and attribute features from deep features, AI-generated tags, and publisher metadata. Through a fusion classification model, it outputs four types of intents and corresponding confidence scores. These four types of intents include malicious guidance, active dissemination, unintentional publishing, and AI-generated errors. The causal analysis unit constructs a causal graph, using sensitive features, AI-generated tags, intent, and publishing scenario as cause nodes and whether the information is sensitive as effect nodes. It designs the causal graph topology based on domain rules, performs counterfactual reasoning based on a structural causal model (SCM), eliminates false associations, quantifies the confidence of genuine sensitive associations, and ultimately obtains the sensitive information determination result. The risk assessment unit, based on the sensitive information determination result, intent, and intent confidence score, determines weights and calculates a weighted total risk score from dimensions such as the severity of the sensitive type, the potential scope of dissemination, the degree of malicious intent, and the risk enhancement from AI generation. The unit then determines the risk level based on the total risk score.

[0010] Optionally, the early warning module includes: a dual-objective modeling unit, an early warning strategy generation unit, and an intervention execution unit. The dual-objective modeling unit is used to construct a dual-objective equilibrium model, converting minimum intervention cost and maximum risk control into quantifiable indicators, assigning objective weights, and establishing a comprehensive performance scoring function. The early warning strategy generation unit is used to match and dynamically generate the optimal intervention plan from a strategy library based on the risk level and the comprehensive performance scoring function. The intervention execution unit is used to connect to the core system of the online media platform, execute the optimal intervention plan in real time through a standardized API interface, and synchronously record the entire intervention process log.

[0011] The second aspect of this application provides a method for intelligent identification and early warning of sensitive information in online media videos, comprising the following steps: collecting online media video data, including video stream data, bullet screen / comment text, and publisher metadata; performing multi-dimensional analysis on the collected video stream data and bullet screen / comment text to obtain deep features, the multi-dimensional analysis including visual, audio, and text dimensions; identifying whether the content is generated by AIGC technology based on the deep features, and if so, marking it with an AI-generated tag; performing intent and contextual causal relationship diagnosis based on the deep features, AI-generated tags, and publisher metadata, performing deep analysis through counterfactual reasoning technology to determine whether it is sensitive information, and assessing the risk to obtain a risk level; aiming at minimum intervention cost and maximum risk control, matching and dynamically generating an optimal intervention plan from a strategy library based on the risk level, and intervening in the sensitive video according to the optimal intervention plan.

[0012] A third aspect of this application provides an electronic device, including: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being configured to perform a method for intelligent identification and early warning of sensitive information in network media video as described in the above embodiments.

[0013] A fourth aspect of this application provides a computer-readable storage medium having a computer program stored thereon, which is executed by a processor to implement a method for intelligent identification and early warning of sensitive information in network media video as described in the above embodiments.

[0014] The fifth aspect of this application provides a computer program product that stores a computer program that, when executed by a processor, implements a method for intelligent identification and early warning of sensitive information in network media video as described in the above embodiments.

[0015] The beneficial effects of using the present invention are as follows: This application's embodiments collect video streams, bullet screen / comment text, and publisher metadata, combining visual, audio, and text-based multi-dimensional deep feature analysis to specifically identify AI-generated content. This covers sensitive information forms easily overlooked by traditional technologies, effectively addressing the governance challenges of AI-synthesized false sensitive content. Employing causal association diagnosis and counterfactual reasoning techniques, it can deeply mine the publisher's true dissemination intent and the logical connections within the content context, outputting a risk level. This significantly reduces the probability of misjudgments and omissions in sensitive information identification, improving the interpretability and credibility of the identification results. By balancing minimum intervention cost with maximum risk control through a dynamic optimal intervention scheme, it avoids the drawbacks of a one-size-fits-all intervention model, such as high platform operating costs and disruption to normal user experience. It also enables differentiated and precise intervention based on the risk level, effectively curbing the spread of sensitive information and ensuring network content compliance while minimizing unnecessary resource consumption and user interference. Therefore, it solves the problems of incomplete multimodal information analysis and the lack of AI-generated content identification mechanisms in existing technologies.

[0016] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0017] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is a schematic diagram of the structure of an intelligent identification and early warning system for sensitive information in network media videos according to an embodiment of this application; Figure 2 This is a flowchart illustrating a method for intelligent identification and early warning of sensitive information in network media videos according to an embodiment of this application; Figure 3 This is a schematic diagram of the structure of an electronic device provided according to an embodiment of this application. Detailed Implementation

[0018] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.

[0019] The following description, with reference to the accompanying drawings, illustrates an intelligent identification and early warning system for sensitive information in online media videos, according to an embodiment of this application. Addressing the issues of incomplete multimodal information analysis and lack of AI-generated content identification mechanisms mentioned in the background section, this application provides an intelligent identification and early warning system for sensitive information in online media videos. This system collects video streams, bullet screen / comment text, and publisher metadata, and combines this with multi-dimensional deep feature analysis of visual, audio, and textual data to perform targeted identification of AI-generated content. This covers sensitive information forms that are easily overlooked by traditional technologies, effectively addressing the governance challenges of AI-synthesized false sensitive content. By employing causal correlation diagnosis and counterfactual reasoning techniques, this approach can deeply uncover the publisher's true dissemination intentions and the logical connections within the content's context, outputting a risk level. This significantly reduces the probability of misjudgments and omissions in sensitive information identification, enhancing the interpretability and credibility of the results. Through a dynamic optimal intervention scheme that balances minimum intervention cost with maximum risk control, it avoids the drawbacks of a one-size-fits-all approach, such as high platform operating costs and disruption to normal user experience. Furthermore, it enables differentiated and precise intervention based on the risk level, effectively curbing the spread of sensitive information and ensuring the compliance of online content while minimizing unnecessary resource consumption and user interference. This solves the problems of incomplete multimodal information analysis and the lack of AI-generated content identification mechanisms in existing technologies.

[0020] Specifically, Figure 1 This is a schematic diagram of the structure of a network media video sensitive information intelligent identification and early warning system provided in an embodiment of this application.

[0021] like Figure 1 As shown, the intelligent identification and early warning system 10 for sensitive information in network media videos includes: a data acquisition module 100, a video information extraction module 200, an AIGC video detection module 300, a causal inference and evaluation module 400, and an early warning module 500.

[0022] The system includes the following modules: a data acquisition module 100 for collecting online media video data, including video stream data, bullet screen / comment text, and publisher metadata; a video information extraction module 200 for multi-dimensional analysis of the collected video stream data and bullet screen / comment text to obtain deep features, including visual, audio, and text dimensions; an AIGC video detection module 300 for identifying whether the content is generated by AIGC technology based on deep features, and if so, labeling it with an AI-generated tag; a causal inference and evaluation module 400 for diagnosing intent and contextual causal relationships based on deep features, AI-generated tags, and publisher metadata, conducting in-depth analysis using counterfactual reasoning techniques to determine whether the information is sensitive and assessing its risk level; and an early warning module 500 for matching and dynamically generating the optimal intervention plan from a strategy library with the goal of minimizing intervention costs and maximizing risk control, and then intervening in sensitive videos according to the optimal intervention plan.

[0023] It is understood that this application's embodiments, by collecting video streams, bullet screen / comment text, and publisher metadata, combined with multi-dimensional deep feature analysis of visual, audio, and text dimensions, specifically identify AI-generated content. This covers sensitive information forms that are easily overlooked by traditional technologies, effectively addressing the governance challenges of AI-synthesized false sensitive content. Employing causal association diagnosis and counterfactual reasoning techniques, it can deeply mine the publisher's true dissemination intent and the logical relationship between the content context, outputting a risk level. This not only significantly reduces the probability of misjudgment and omission in sensitive information identification but also improves the interpretability and credibility of the identification results. By balancing the minimum intervention cost with the maximum risk control through a dynamic optimal intervention scheme, it avoids the drawbacks of a one-size-fits-all intervention model, such as high platform operating costs and disruption to normal user experience. Furthermore, it enables differentiated and precise intervention based on the risk level, effectively curbing the spread of sensitive information and ensuring the compliance of online content while minimizing unnecessary resource consumption and user interference.

[0024] In this embodiment, the data acquisition module includes: a data acquisition unit and a data preprocessing unit.

[0025] The data acquisition unit is used to crawl video stream data using web crawler technology and obtain bullet screen / comment text and publisher metadata through the platform's open API. The data preprocessing unit is used to clean and align the acquired data. For video stream data, it uses a frame quality assessment algorithm to filter valid frames, remove stuttering frames, duplicate frames, and black frames, uses an improved spectral subtraction method to suppress environmental noise, and unifies the sampling rate and encoding format. For text data, it uses regular expressions to filter out garbled characters, meaningless characters, and spam content. Based on timestamps and video IDs, it establishes a multi-dimensional association index of video, text, and publisher metadata to generate structured data with unified identifiers.

[0026] Specifically, a distributed web crawler architecture is used to capture video stream data, employing an "incremental crawling + timed update" mechanism, and URL queue management to avoid duplicate crawling. The captured raw video stream is temporarily stored in segmented TS format, synchronously recording basic attributes such as video URL, capture timestamp, and platform identifier. Bullet comments and other comment data are obtained in batches through the open API of the online media platform according to the platform's traffic limiting rules, including complete fields such as content text, publication time, publishing user ID, and number of likes. Core metadata of the publisher is obtained through the platform's account information open API, including publisher account information (account ID, nickname, authentication type, registration time), behavioral attributes (number of historically published videos, number of followers, interaction rate in the past 30 days), and risk attributes (historical records of penalties for sensitive content, anonymity status), etc.

[0027] For the acquired video stream data, a multi-index frame quality assessment algorithm is used to perform frame-by-frame detection on the original video stream. This is achieved through sharpness assessment (calculating frame image entropy values; entropy values ​​< 1.5 are considered blurry frames), brightness detection (average grayscale value < 30 is considered a black frame), and inter-frame similarity analysis (SSIM similarity > 0.95 for consecutive frames is considered a duplicate frame). Stuttering frames, blurry frames, black frames, and duplicate frames are removed, retaining the valid frame sequence. For environmental noise (such as background noise and current noise) in the video's built-in audio, an improved spectral subtraction method is used for noise reduction. The processed video frame sequence is then re-encoded into H.264 format, with a uniform audio sampling rate of 44.1kHz, a bit depth of 16bit, and an AAC encoding format.

[0028] For the acquired bullet screen / comment text, multi-dimensional filtering rules are constructed based on regular expressions to batch remove garbled characters (such as abnormal UTF-8 encoded characters), meaningless characters (such as consecutive special symbols "!!!###"), and spam content (such as text with repeated characters more than 5 times, and spam advertisements unrelated to the video). Simultaneously, very short texts (less than 3 characters) are filtered, while retaining valid interactive content. The filtered text is then uniformly encoded (UTF-8), case-converted (lowercase uniform), and punctuation is standardized.

[0029] Using video ID as the core primary key and timestamp as the secondary index, a multi-dimensional association mapping is established for video stream data, bullet screen / comment text (sorted by release time), and publisher metadata. All preprocessed data is integrated into a structured data package in JSON format. Each data package contains a unified identifier, basic attributes (collection time, video ID), and core data (standardized video stream, cleaned text, publisher metadata).

[0030] It is understandable that this application embodiment comprehensively covers video stream data, bullet screen / comment text, and publisher metadata through a hybrid collection mode of web crawling and platform API integration. This ensures that no core data from multiple sources is omitted and that the collection process is compliant and stable. Furthermore, targeted preprocessing operations achieve precise optimization of data quality. Effective frame filtering, environmental noise suppression, and format standardization at the video stream level significantly improve the purity and consistency of video data. Regular expression filtering at the text level effectively removes redundant and interfering information. Combined with timestamps and video IDs, a multi-dimensional association index is established to generate structured data with unified identifiers. This not only solves the pain points of heterogeneous, scattered, and noisy multi-source data, but also achieves spatiotemporal alignment and association integration of video, text, and publisher metadata, providing high-quality and highly adaptable structured input for in-depth analysis in subsequent modules.

[0031] In this embodiment, the video information extraction module includes: a multi-dimensional parsing unit and a deep feature fusion unit.

[0032] The multi-dimensional parsing unit uses a pre-trained video understanding model to extract spatial and temporal features of video images, uses a pre-trained audio neural network to extract acoustic and deep semantic features of video sound, converts video sound into text to obtain video subtitle information, processes video subtitle information and bullet screen / comment text to generate text semantic vectors, extracts core topics through an LDA model, calculates sentiment polarity and sentiment intensity, and obtains text features; the deep feature fusion unit is used to concatenate video images, video sound, and text features to obtain deep features.

[0033] Specifically, a pre-trained video understanding model is used to segment and encode the effective frames after video preprocessing, outputting a 768-dimensional spatial feature vector that covers details such as target shape, color texture, scene layout, and edge contours. The effective frame sequence is divided into video segments of 16 frames per segment, which are then input into a pre-trained I3D model. 3D convolution is used to capture inter-frame motion trajectories, shot transition logic, and action coherence, outputting a 1024-dimensional temporal feature vector to characterize the changing patterns of the dynamic scene. The extracted spatial and temporal features are stored as vectors, simultaneously recording the corresponding frame positions and segment temporal information.

[0034] A pre-trained audio neural network was used to process the standardized audio data into frames (20ms frame length, 10ms frame shift), extracting a 128-dimensional acoustic feature vector containing core indicators such as pitch, loudness, spectral distribution, speech rate, and spectral smoothness. Through the encoding layer of the audio neural network, the audio signal was mapped to a 768-dimensional deep semantic feature vector, capturing the emotional tendency and semantic connotation in the speech. The Whisper-large-v2 model was used to transcribe the audio into text subtitles, supporting multi-language recognition including Chinese and English. After transcription, punctuation was repaired and sentence segmentation was optimized to generate standardized subtitle text. The standardized subtitle text and the pre-processed bullet screen / comment text were combined into a unified text dataset. The text dataset underwent word segmentation and part-of-speech tagging, removing invalid and redundant information while retaining core semantic components. A pre-trained BERT-based-Chinese model is used to encode the processed text, outputting a 768-dimensional text semantic vector. A preset number of 5-10 topics are selected in the LDA topic model. The text semantic vectors are input into the LDA topic model to mine core topics of the text set (such as politically sensitive topics, violence-related topics), and output keywords and probability distributions for each topic. The VADER sentiment analysis algorithm is used, combined with the text semantic vectors, to calculate sentiment polarity (positive / negative / neutral) and sentiment intensity (0-1 points, closer to 1 indicates stronger sentiment). The text semantic vectors, LDA core topics (keywords + probability distribution), sentiment polarity labels, and sentiment intensity scores are associated and stored by video ID to obtain text features.

[0035] It is understood that the embodiments of this application accurately extract the spatial and temporal features of video images and the acoustic and deep semantic features of sound through video understanding models and audio neural networks. They also integrate subtitles and bullet screen / comment text through speech-to-text conversion, and comprehensively cover key identification information in the three dimensions of vision, audio and text through semantic vector generation, LDA core topic extraction and sentiment polarity and intensity calculation. This avoids the limitations of single-modal features. By directly splicing, heterogeneous multimodal features are integrated into unified deep features, which not only retains the core representation information of each modality, but also simplifies the fusion process and improves processing efficiency. This provides comprehensive, complete and highly adaptable feature inputs for the content identification of the AIGC video detection module and the sensitive information discrimination and risk assessment of the causal inference evaluation module.

[0036] In this embodiment, the AIGC video detection module includes: an artifact feature detection unit, a semantic consistency verification unit, and a tagging unit.

[0037] The artifact feature detection unit is used to capture and quantify AIGC generation traces from multiple dimensions to obtain artifact confidence scores. Specifically, in the visual dimension, it detects common AIGC detail defects based on the spatial features of the video frame and identifies inter-frame motion anomalies based on the temporal features of the video frame. An outlier detection algorithm is used to quantify the degree of artifacts, generating a visual artifact score. In the audio dimension, it determines the presence of AI-synthesized speech traces based on the acoustic and deep semantic features of the video sound, generating an audio artifact score. In the text dimension, it identifies typical AIGC text defects based on the text semantic vector and core topics, obtaining a text artifact score. The multi-dimensional artifact scores are combined and weighted to generate an artifact confidence score. The reliability score and semantic consistency verification unit are used to verify video-audio consistency, video-text consistency, and content-comment sentiment consistency. Based on the verification results, a semantic consistency score is generated. Among them, video-audio consistency is used to verify whether the video content matches the audio, video-text consistency is used to verify whether the semantic features of the video content and the text are consistent, and content-comment sentiment consistency is used to verify whether the emotional tone of the video content matches the sentiment of the bullet screen / comment. The labeling unit is used to output a binary decision result and an overall confidence score through the decision engine based on the artifact confidence score and the semantic consistency score. The binary decision result includes AI-generated and non-AI-generated results. If the binary decision result is AI-generated, then an AI-generated label is used.

[0038] Specifically, a pre-defined AIGC high-frequency defect database (such as facial feature misalignment, hand and finger deformities, texture repetition and blurring, and logical inconsistencies in light and shadow reflections) is used. Based on the 768-dimensional spatial feature vector of the video frame, the matching degree between the feature vector and the defect database is calculated using cosine similarity. Simultaneously, the isolated forest algorithm is used to detect feature outliers and identify novel visual distortions not included in the defect database. A spatial artifact score of 0-1 is output; the higher the matching degree and the higher the outlier, the closer the score is to 1. Based on the 1024-dimensional temporal feature vector of the video frame, the feature similarity between adjacent video segments (16 frames / segment) is calculated. If three consecutive segments have a similarity > 0.95 (indicating duplicate frame stacking) or < 0.1 (indicating broken motion trajectory), an anomaly marker is triggered. Kalman filtering is used to predict the object's motion trajectory, which is compared with the actual inter-frame features. If the deviation exceeds a preset threshold, it is recorded as a temporal anomaly, and a temporal anomaly score is obtained. The spatial artifact score and the temporal anomaly score are weighted and summed at a ratio of 6:4, and then normalized to the 0-1 interval using the Sigmoid function to obtain the final visual artifact score.

[0039] Based on the 128-dimensional acoustic feature vector of the video, three key indicators are extracted: spectral smoothness, fundamental frequency stability, and the proportion of breathing sounds. AI-synthesized speech typically exhibits a spectral smoothness > 0.9 (natural speech is generally < 0.75), a fundamental frequency fluctuation range < 50Hz (natural speech > 80Hz), and a breathing sound proportion < 0.05 (natural speech > 0.1). Meeting any one of these indicators is considered suspicious. Combining the 768-dimensional audio semantic feature vector, semantic coherence is calculated using the BERT model. Semantic discontinuities (coherence score < 0.3) are used to strengthen artifact detection. The number of acoustic feature anomalies and the semantic coherence score are fused with a 1:1 weight. Each additional anomaly adds 0.25 points, and the semantic coherence score is calculated inversely. Finally, the summation is normalized to the 0-1 range to obtain the final audio artifact score.

[0040] Based on the 768-dimensional text semantic vector, vector sparsity is calculated (AIGC text sparsity is typically >0.6, while natural text sparsity is <0.4). Combined with LDA core topics, if the matching degree between text keywords and core topics is <0.5, or the topic switching frequency is >5 times / minute, it is judged as topic drift and marked as text artifact. At the same time, through sentiment entropy calculation, sentiment polarity fluctuation amplitude >0.8 (natural text is generally <0.5) is also included in the defect range. Vector sparsity, topic matching degree, and sentiment fluctuation amplitude are fused with a weight of 4:3:3, and the topic matching degree is reversed (1-matching degree). After summing, the results are normalized to the 0-1 interval to obtain the final text artifact score.

[0041] The visual artifact scores, audio artifact scores, and text artifact scores are weighted and summed to obtain the artifact confidence score.

[0042] The CLIP cross-modal matching model is employed, inputting the spatial feature vectors of the video frame and the semantic feature vectors of the audio frame into the model to calculate the matching similarity. The temporal synchronization between lip movements and speech is verified by using a dynamic time warping algorithm to calculate the deviation between lip movements and speech rhythm; a deviation <0.2 is considered acceptable. Finally, the matching similarity score and the acceptable score are fused with a 7:3 weighting to obtain a video-audio consistency score of 0-1.

[0043] Input the visual spatial feature vector and the text semantic vector into the cosine similarity calculator. If the similarity is less than 0.4, it is considered inconsistent. The video-text consistency score is obtained as 0 or 1.

[0044] The emotional tone of the video content is extracted (calculated using the VADER algorithm based on audio semantics and visual color features) and compared with the emotional polarity and intensity of the bullet comments / comments. If the emotional polarity is opposite (e.g., the content is sad but the comments are happy) or the intensity difference is >0.6, it is judged as inconsistent. The consistency rate (number of consistent comments / total number of comments) is the score for this dimension, which is normalized to 0-1 to obtain the content-comment emotional consistency score.

[0045] The semantic consistency score is obtained by weighting and summing the video-audio consistency score, the video-text consistency score, and the content-comment sentiment consistency score.

[0046] A gradient boosting decision tree is used as a classifier to receive artifact confidence scores and semantic consistency scores. Through complex nonlinear calculations, the deep relationship between these scores and AI generation is learned, outputting a probability value between 0% and 100%. When the probability value is greater than a target threshold, the video is determined to be AI-generated and labeled with an AI-generated tag; otherwise, it is considered non-AI-generated.

[0047] It is understood that this application's embodiments achieve multi-dimensional, high-precision identification of AI-generated content by constructing a dual discrimination system of "microscopic artifact analysis + macroscopic semantic verification." The artifact feature detection unit can accurately capture the generation traces of AIGC technology from multi-dimensional data such as vision, audio, and text, and form a comprehensive artifact confidence score through quantitative evaluation, effectively identifying the underlying technical defects of various generative AIs. The semantic consistency verification unit performs in-depth verification at the cross-modal logical association level, effectively identifying inherent logical contradictions and traces of human manipulation in the content by analyzing the semantic consistency between video, audio, text, and comments. By fusing the two types of evidence for collaborative decision-making, the accuracy and robustness of AIGC content discrimination are significantly improved.

[0048] In this embodiment, the causal inference assessment module includes: an intent diagnosis unit, a causal analysis unit, and a risk assessment unit.

[0049] The intent diagnosis unit extracts behavioral, content, and attribute features from deep features, AI-generated tags, and publisher metadata. By fusing classification models, it outputs four types of intents and their corresponding confidence scores. These four types of intents include malicious guidance, active dissemination, unintentional publication, and AI-generated errors. The causal analysis unit constructs a causal graph, using sensitive features, AI-generated tags, intents, and publication scenarios as cause nodes and whether the information is sensitive as effect nodes. It designs the causal graph topology based on domain rules, performs counterfactual reasoning based on a structural causal model (SCM), eliminates false associations, quantifies the confidence of genuine sensitive associations, and finally obtains the sensitive information determination result. The risk assessment unit determines the weights and calculates the total risk score based on the sensitive information determination result, intent, and intent confidence score, considering the severity of the sensitive type, the potential scope of dissemination, the degree of malicious intent, and the risk enhancement from AI generation. The risk level is then determined based on the total risk score.

[0050] Specifically, based on the number of videos posted, number of followers, and interaction rate over the past 30 days, behavioral features are extracted for the past 30 days. The frequency of AI-generated content postings within the past 30 days is statistically analyzed, and behavioral sequences are captured using an LSTM model to output a behavioral feature vector. Content features are extracted based on deep features. If the AI-generated label is 1, an attention mechanism is used to increase the weight of AI-related content features, resulting in a content feature vector. The AI-generated label is then co-encoded with fields such as anonymity status, historical sensitive records, and account authentication type to obtain an attribute feature vector.

[0051] The behavioral feature vector, content feature vector, and attribute feature vector are concatenated, and weights are dynamically allocated through a cross-modal attention layer: when the AI-generated label is 1, the weight of the AI ​​label dimension in the attribute features is increased to 0.4, and the weight of the AI ​​synthesis trace dimension in the content features is increased to 0.3, thus enhancing the recognition of AI-related intents. A BERT-base model is used to capture deep semantic relationships, and the fully connected layer outputs confidence scores for the four types of intents.

[0052] Construct a cause-effect graph and design its topology based on domain rules. The topological relationships are as follows: Direct causal edges: Sensitive features → Sensitive information (sensitive features directly drive the determination of sensitive information, core edge); Intent → Sensitive information (intent directly affects the authenticity of risk, such as malicious intent strengthening the determination of sensitivity); Release scenario → Sensitive information (release scenario directly affects the spread of risk, high scenario increases the sensitivity weight); AI-generated tags → Sensitive information (AI-generated tags directly affect the risk type, such as AI batch generating sensitive content).

[0053] Indirect causal edge: AI video → high dissemination (AI videos are easily recommended by algorithms to enter high dissemination scenarios); intent → high dissemination (publishers who maliciously guide intent are more inclined to choose high dissemination scenarios for publication).

[0054] Co-causal edge: (Sensitive feature + intent) → Sensitive information (Sensitive feature + malicious intent form a synergistic effect, and the sensitive risk increases exponentially); (AI generation + sensitive feature) → Sensitive information (AI generation + sensitive feature form a "synthetic sensitivity" synergy, and the risk is higher than that of non-AI + sensitive feature).

[0055] The core of SCM is to define structured functions for result nodes, integrating the influence of the four major cause nodes. The function form is as follows: ; in, Confidence level for sensitive information; It is a sensitive feature, and it is a binary variable. Its weight; Generate labels for AI, which are binary variables. Its weight; For the purpose of, Its weight; For the publishing scenario, it is a binary variable. Its weight; The synergistic weighting of sensitive features and intent; Generate collaborative weights for AI based on labels and sensitive features; For bias terms; This is a sigmoid function that maps the output to the 0-1 range. Different weights are applied to AI-tagged videos and non-AI-tagged videos.

[0056] The core of counterfactual reasoning is to "assume that the state of a cause node changes and observe whether the effect node changes," thereby eliminating spurious associations.

[0057] For videos without AI tags, two types of core counterfactual queries are designed to verify the authenticity of sensitive associations. These are: Scenario 1: "If sensitive features exist but the publishing intent is 'unintentional publishing,' is it still considered sensitive information?", fixing "sensitive features," "publishing scenario," and "AI-generated tag" (0), only intervening in "intent" as "unintentional publishing," and calculating the confidence level. Scenario 2: "If sensitive features exist but the publishing scenario is a low-propagation scenario, is it still considered sensitive information?", fixing "sensitive features," "intent," and "AI-generated tag" (0), only intervening in "publishing scenario" as "low-propagation scenario," and calculating the confidence level.

[0058] Three core counterfactual queries were designed for videos with AI-generated tags to verify the authenticity of sensitive associations. These are: Scenario 1: "If the video is not AI-generated, and the sensitive features, intent, and publishing scenario remain unchanged, is it still considered sensitive information?", fixing "sensitive features," "intent," and "publishing scenario," and only intervening with "AI-generated tag" to 0, and calculating the confidence level. Scenario 2: "If the publishing intent is 'AI misgenerated,' even with the presence of sensitive features, is it still considered sensitive information?", fixing "sensitive features," "AI-generated tag (1)," and "publishing scenario," and only intervening with "intent" to "AI misgenerated," and calculating the confidence level. Scenario 3: "If the sensitive features are removed, will the AI-generated tag still lead to a sensitive determination?", fixing "AI-generated tag (1)," "intent," and "publishing scenario," and only intervening with "sensitive features" to 0, and calculating the confidence level.

[0059] If the probability does not decrease significantly in the counterfactual scenario, the counterfactual inference verification coefficient is set to 1; if the probability decreases by 20%-40% in the counterfactual scenario, the counterfactual inference verification coefficient is set to 0.7; if the probability decreases by ≥50% in the counterfactual scenario, the counterfactual inference verification coefficient is set to 0.3. The confidence score of the true sensitive association is the product of the original probability calculated by the SCM function and the counterfactual inference verification coefficient.

[0060] For videos with AI-tagged tags, if the confidence level of the true sensitive association is greater than 0.6, the video is considered sensitive information. For videos without AI-tagged tags, if the confidence level of the true sensitive association is greater than 0.5, the video is considered sensitive information.

[0061] Sensitivity type severity: The severity of sensitivity type is determined based on the confidence level of the true sensitivity association. A confidence level of the true sensitivity association greater than 0.8 scores 4 points; a confidence level of the true sensitivity association between 0.5 and 0.8 scores 3 points; a confidence level of the true sensitivity association between 0.3 and 0.5 scores 2 points; and a confidence level of the true sensitivity association less than 0.3 scores 1 point.

[0062] Potential reach: The potential reach is determined based on the number of followers of the publisher and the publishing scenario. Publisher with more than 100,000 followers = 4 points × publishing scenario coefficient; publisher with 50,000-100,000 followers = 3 points × publishing scenario coefficient; publisher with 10,000-50,000 followers = 2 points × publishing scenario coefficient; publisher with less than 10,000 followers = 1 point × publishing scenario coefficient. The publishing scenario coefficient is 1 for high reach and 0.5 for low reach.

[0063] Malicious intent level: Calculated by weighting four types of intent: Malicious guidance = 4 points × confidence level, Active dissemination = 3 points × confidence level, Unintentional release = 1 point × confidence level, AI-generated error = 1 point × confidence level.

[0064] AI-generated risk bonus: 1 point is added when AI-generated tag = 1, 0 points are added when it is not AI-generated, and an additional 0.5 points are added if it is AI-generated and determined to be sensitive information.

[0065] Total risk score = Severity of sensitive type × 0.4 + Potential scope of spread × 0.25 + Malicious intent × 0.2 + AI-generated risk bonus × 0.15 The original score range was 0-4 points, which was converted to a 0-10 score range through linear mapping.

[0066] A score of 0-3 indicates a low risk level. A score of 3-6 indicates a medium risk level. A score of 6-8 indicates a high risk level. A score of 8-10 indicates an extremely high risk level.

[0067] It is understood that the embodiments of this application achieve in-depth understanding and precise quantification of video content risks by constructing a progressive analysis framework of "intent diagnosis - causal analysis - risk assessment". The intent diagnosis unit can accurately identify the underlying motivation for content dissemination, providing key dimensions for risk assessment; the causal analysis unit, through causal graphs and counterfactual reasoning, breaks through the limitations of traditional correlation analysis, effectively distinguishes between real sensitive associations and false statistical associations, and significantly improves the interpretability and accuracy of sensitive information identification; the risk assessment unit integrates multiple factors and generates a quantitative risk level through weighted fusion, providing a scientific basis for early warning decisions that considers both objective impacts and subjective malice, and comprehensively improves the system's risk assessment capabilities in complex information environments.

[0068] In this embodiment, the early warning module includes: a dual-objective modeling unit, an early warning strategy generation unit, and an intervention execution unit.

[0069] The dual-objective modeling unit is used to construct a dual-objective equilibrium model, which transforms the minimum intervention cost and the maximum risk control into quantifiable indicators, assigns target weights, and establishes a comprehensive performance scoring function. The early warning strategy generation unit is used to match and dynamically generate the optimal intervention plan from the strategy library based on the risk level and the comprehensive performance scoring function. The intervention execution unit is used to connect to the core system of the network media platform, execute the optimal intervention plan in real time through a standardized API interface, and synchronously record the entire intervention process log.

[0070] Specifically, intervention costs include technology implementation costs, operational management costs, and user experience costs. Risk control includes risk reduction rate and transmission blocking rate. The comprehensive effectiveness score function is: ; Where R is the overall performance score, ranging from 0 to 10; Risk control weights; Weighting for intervention costs; For risk reduction rate, Its weight; For the transmission blocking rate, Its weight; For the cost of technology implementation, Its weight; For operating and management costs, Its weight; For user experience cost, Assign weights to them.

[0071] A strategy library is constructed using a two-tiered architecture of "basic strategies + scenario-adaptive strategies" to cover intervention needs across different risk levels and video types (AI / non-AI). Based on the video's risk level, a corresponding set of basic strategies is extracted from the strategy library. If the video is generated by AI, AI-adaptive strategies are overlaid to form a candidate strategy list. For each strategy in the candidate strategy list, its comprehensive effectiveness score S is calculated by substituting it into the comprehensive effectiveness score function of the bi-objective modeling unit. If any candidate strategy has an S greater than 8, the strategy with the highest score is directly selected as the optimal intervention plan. If the S values ​​of all candidate strategies are less than 8, strategy combination generation is triggered, the S values ​​of the combined strategies are recalculated, and the optimal combination is selected as the optimal intervention plan.

[0072] It is understood that the embodiments of this application construct a dual-objective equilibrium model to quantify the minimum intervention cost and the maximum risk control into calculable indicators, thereby achieving a precise balance between governance effectiveness and operating costs. The early warning strategy generation unit dynamically generates the optimal intervention plan based on the risk level and comprehensive efficiency function, ensuring that the response measures are both accurate and economical. The intervention execution unit achieves seamless connection with the platform's core system through standardized interfaces, ensuring the real-time nature of the response while building an auditable and traceable decision-making closed loop based on full-process log recording. Ultimately, an intelligent early warning governance system with scientific decision-making, rapid response, and continuous optimization capabilities is formed.

[0073] The following is a detailed description of a network media video sensitive information intelligent identification and early warning system through a specific embodiment: The analysis focused on a short video from a certain online media platform. The video was posted by an anonymous account (52,000 followers) during peak evening traffic hours on the platform's public recommendation feed. The specific processing steps are as follows: Data Acquisition Module: This module uses a distributed web crawler architecture to capture the original TS format video stream, simultaneously recording the video URL, acquisition timestamp, and platform identifier. It also obtains the video's bullet screen / comment text (236 in total, including many inflammatory comments) and publisher metadata via the platform's open API: the account is anonymous, registered 3 months ago, has published 12 AI-generated videos in the past 30 days, has no history of sensitive penalties, has 52,000 followers, and an interaction rate of 18% in the past 30 days. Preprocessing Stage: A multi-index frame quality assessment algorithm is used to inspect the video stream frame by frame. Calculating frame image entropy values ​​reveals 3 blurry frames and 2 black frames, which are removed, leaving 156 valid frame sequences. An improved spectral subtraction method is used to suppress background noise in the video audio, unifying the audio to a 44.1kHz sampling rate and AAC encoding format. Regular expression filtering is applied to the bullet screen / comment text to remove garbled characters, continuous special symbols, and extremely short texts (<3 characters), ultimately retaining 189 valid comments. Using the video ID (VID20251125001) as the core primary key and the timestamp as the secondary index, a multi-dimensional association mapping is established between video stream, text, and publisher metadata, generating a JSON-formatted structured data package containing a unified UUID identifier, collection attributes, and standardized core data.

[0074] The video information extraction module: Effective frames are segmented and encoded using a pre-trained ViT model, outputting a 768-dimensional spatial feature vector to capture visual details containing false political symbols in the video. Effective frames are divided into 9 video segments of 16 frames per segment, input into the I3D model, and output a 1024-dimensional temporal feature vector to identify abnormal features such as broken motion trajectories between frames. Standardized audio is segmented using a pre-trained AudioBERT model, extracting a 128-dimensional acoustic feature vector with a spectral smoothness of 0.92 (far exceeding the natural speech threshold of 0.75). A 768-dimensional audio deep semantic feature vector is also output, capturing inflammatory semantics in the speech. The Whisper-large-v2 model is used to transcribe the audio into standardized subtitle text, containing multiple instances of false political statements. The subtitle text and the cleaned bullet screen / comment text were integrated into a text dataset. After jieba word segmentation and stop word filtering, the data was input into the BERT-base-chinese model, which output a 768-dimensional text semantic vector. The core topic "false political information" was extracted using the LDA topic model (preset number of topics 6), with keywords "inaccurate policy interpretation" and "incitement to confrontation". The VADER algorithm was used to calculate that the sentiment polarity of the comments was mainly negative, with a mean sentiment intensity of 0.9. The deep feature fusion unit concatenates visual (spatial + temporal), audio (acoustic + semantic), and text features to generate cross-modal deep features, providing core input for subsequent detection and evaluation.

[0075] In the AIGC video detection module, the artifact feature detection unit shows that the visual dimension achieves a matching degree of 0.85 between spatial features calculated using cosine similarity and the AIGC high-frequency defect library. The isolated forest algorithm detects novel visual distortion features, and the temporal features identify two segments of overlapping repeated frames (similarity 0.97). The final visual artifact score, after 6:4 weighting and Sigmoid normalization, is 0.88. In the audio dimension, the extracted spectral smoothness, fundamental frequency stability, and proportion of breathing sounds all conform to AI synthesized speech features, with a semantic coherence score of 0.2 (indicating semantic discontinuity). The audio artifact score, after normalization, is 0.91. In the text dimension, the semantic vector sparsity is 0.72, the matching degree between text keywords and core topics is 0.35 (indicating topic drift), and the emotional fluctuation amplitude is 0.85. The text artifact score, after 4:3:3 weighting, is 0.82. The weighted sum of the visual, audio, and text artifact scores yields an artifact confidence score of 0.87. In the semantic consistency verification unit, the video-audio matching similarity calculated using the CLIP model is 0.3, the lip-sync and speech rhythm deviation is 0.35 (unacceptable), and the video-audio consistency score is 0.2; the cosine similarity between visual spatial features and text semantic vectors is 0.28, and the video-text consistency score is 0; the emotional tone of the video content is inflammatory and negative, while some comments are malicious echoes (emotional intensity difference 0.7), and the content-comment sentiment consistency score is 0.15. The weighted average of the three scores results in a semantic consistency score of 0.12. The labeling unit uses a gradient boosting decision tree to input the artifact confidence score and semantic consistency score, outputting an overall confidence score of 92%, classifying it as AI-generated content, and labeling the video with AI-generated tags.

[0076] The causal inference assessment module: The intent diagnosis unit extracts behavioral features, content features, and attribute features from deep features, AI-generated tags, and publisher metadata. After feature concatenation and cross-modal attention weighting, it is input into the BERT-base model to output four types of intent confidence scores: malicious guidance 0.92, active dissemination 0.05, unintentional publication 0.02, and AI-generated error 0.01. The causal analysis unit constructs a causal graph, setting sensitive features = 1, AI-generated tags = 1, intent = 0.92, and publication scenario = 1 (high dissemination). The original sensitive association confidence score is calculated to be 0.95 using the SCM function. For videos with AI tags, three types of counterfactual queries are designed: Scenario 1: Intervention AI-generated tag = 0, calculated confidence score 0.68 (decreased by 28%); Scenario 2: Intervention intent is AI-generated error, confidence score 0.21 (decreased by 78%); Scenario 3: Intervention sensitive features = 0, confidence score 0.18 (decreased by 81%). The verification coefficient was set to 0.7, and the final confidence level of the true sensitive association was 0.95 × 0.7 = 0.665, which was determined to be sensitive information. In the risk assessment unit, the severity of the sensitivity type was scored as 3 points based on a confidence level of 0.665; the potential scope of dissemination was calculated as 52,000 followers (3 points) × the posting scenario coefficient 1 = 3 points; the degree of malicious intent was 4 × 0.92 = 3.68 points; and the AI-generated risk bonus was 1 + 0.5 = 1.5 points. The total risk score was 3 × 0.4 + 3 × 0.25 + 3.68 × 0.2 + 1.5 × 0.15 = 1.2 + 0.75 + 0.736 + 0.225 = 2.911 (raw score), which, after linear mapping (×2.5), yielded a standard score of 7.277 points, which was determined to be high risk.

[0077] A dual-objective equilibrium model was constructed, with a risk control weight of 0.8 and an intervention cost weight of 0.2. For this high-risk AI video, the candidate strategy was "video removal (visible only to the publisher) + account traffic restriction for 14 days + AI-generated features added to the database." The intervention cost was quantified as follows: technical implementation cost 6 points, operational management cost 5 points, and user experience cost 3 points, for a total cost of 0.4×6 + 0.3×5 + 0.3×3 = 4.8, resulting in a cost-benefit ratio of 5.2. The risk control effectiveness was assessed as follows: risk reduction rate 9 points, dissemination blocking rate 9 points, for a total risk control effectiveness of 0.5×9 + 0.5×9 = 9. The overall effectiveness score S = 0.8×9 + 0.2×5.2 = 8.24, satisfying the condition S≥8, thus this strategy was selected as the optimal intervention plan. The intervention execution unit connected to the platform's content management system and user management system via a standardized API interface to perform real-time video removal (visible only to the publisher) + account traffic restriction for 14 days, and reported the sensitive content information to the regulatory authorities via a data reporting API. The entire intervention process is logged synchronously, including information such as video ID, risk level, intervention plan, execution time, execution result, and actual effect of risk control. The logs are stored in the platform's distributed database to complete the entire process of early warning and intervention.

[0078] In summary, this application's embodiments ensure the integrity and usability of video, text, and publisher information through multi-source data collection and standardized preprocessing; accurately identify AI-generated attributes and sensitive core features through feature extraction and AI artifact detection; eliminate false associations by combining counterfactual reasoning, accurately determine sensitive attributes and high-risk levels, and avoid misjudgments and omissions; balance risk control and intervention costs through dual-objective modeling, and the matched "removal + traffic restriction + feature storage" strategy not only achieves risk downgrading but also avoids resource waste caused by excessive intervention; and simultaneously completes AI feature tracing and full-process log retention, meeting the traceability and compliance requirements of network content governance.

[0079] Secondly, referring to the accompanying drawings, a method for intelligent identification and early warning of sensitive information in network media videos, based on an embodiment of this application, is described.

[0080] Specifically, Figure 2 A flowchart illustrating an intelligent identification and early warning method for sensitive information in network media videos provided in this application embodiment.

[0081] like Figure 2 As shown, this intelligent identification and early warning method for sensitive information in online media videos includes the following steps: In step S101, online media video data is collected, including video stream data, bullet screen / comment text, and publisher metadata.

[0082] Specifically, video stream data is crawled using web scraping technology, and bullet screen / comment text and publisher metadata are obtained through the platform's open API. The acquired data is cleaned and aligned. For video stream data, a frame quality assessment algorithm is used to filter valid frames, removing stuttering frames, duplicate frames, and black frames. An improved spectral subtraction method is used to suppress environmental noise, and the sampling rate and encoding format are standardized. For text data, regular expressions are used to filter out garbled characters, meaningless characters, and spam content. A multi-dimensional association index of video, text, and publisher metadata is established based on timestamps and video IDs to generate structured data with unified identifiers.

[0083] It is understood that this application's embodiments achieve comprehensive collection of video streams, bullet comments / reviews, and publisher metadata through a combination of automated crawling and API interfaces. Signal processing and text cleaning techniques effectively remove invalid frames, environmental noise, and meaningless text, significantly improving the quality of the original data. Through multi-dimensional association indexing based on timestamps and video IDs, unstructured data is transformed into structured data with unified identifiers. This not only achieves data alignment but also provides complete, clean, and clearly correlated data input for subsequent feature extraction, AIGC detection, and causal inference.

[0084] In step S102, the collected video stream data and bullet screen / comment text are analyzed in multiple dimensions to obtain deep features. The multiple dimensions include visual dimension, audio dimension, and text dimension.

[0085] Specifically, a pre-trained video understanding model is used to extract spatial and temporal features of the video frame, and a pre-trained audio neural network is used to extract acoustic and deep semantic features of the video sound. The video sound is converted into text to obtain video subtitle information. The video subtitle information and bullet screen / comment text are processed to generate text semantic vectors. The core topic is extracted using an LDA model, and sentiment polarity and sentiment intensity are calculated to obtain text features. The video frame, video sound, and text features are concatenated to obtain deep features.

[0086] It is understood that the embodiments of this application utilize pre-trained models to extract spatial and temporal features of the image and acoustic and semantic features of the sound, respectively. Using speech recognition and natural language processing technologies, subtitles and comment texts are transformed into semantic vectors, topic distributions, and sentiment quantification indicators, achieving multi-level information mining from pixels and sound waves to semantics. By concatenating features, heterogeneous data is unified into a fusion representation, not only fully preserving the complementary information of each modality but also forming deep feature vectors with rich semantic connotations.

[0087] In step S103, the content is identified as being generated by AIGC technology based on the depth features. If it is, an AI-generated tag is added.

[0088] Specifically, by capturing and quantifying traces of AIGC generation from multiple dimensions, an artifact confidence score is obtained. In the visual dimension, common AIGC defects are detected based on the spatial features of the video frame, and inter-frame motion anomalies are identified based on the temporal features of the video frame. An outlier detection algorithm is used to quantify the degree of artifacts and generate a visual artifact score. In the audio dimension, the presence of AI-synthesized speech traces is determined based on the acoustic features and deep semantic features of the video sound, generating an audio artifact score. In the text dimension, typical defects of AIGC text are identified based on the text semantic vector and core topics, generating a text artifact score. The artifact confidence score is generated by combining the multi-dimensional artifact scores and calculating them by weight.

[0089] Verify video-audio consistency, video-text consistency, and content-comment sentiment consistency. Generate a semantic consistency score based on the verification results. Among them, video-audio consistency is used to verify whether the video content matches the sound, video-text consistency is used to verify whether the semantic features of the video content and the text are consistent, and content-comment sentiment consistency is used to verify whether the emotional tone of the video content matches the sentiment of the bullet screen / comment.

[0090] Based on the artifact confidence score and semantic consistency score, the decision engine outputs a binary decision result and an overall confidence score. The binary decision result includes AI-generated and non-AI-generated results. If the binary decision result is AI-generated, it is labeled as AI-generated.

[0091] It is understood that the embodiments of this application, by constructing a dual discrimination system of "microscopic artifact analysis + macroscopic semantic verification", have achieved multi-dimensional and high-precision recognition of AI-generated content, providing reliable discrimination basis and confidence support for subsequent risk assessment and early warning.

[0092] In step S104, intent and contextual causal relationship diagnosis is performed based on deep features, AI-generated tags, and publisher metadata. In-depth analysis is conducted using counterfactual reasoning technology to determine whether the information is sensitive and assess its risk level.

[0093] Specifically, behavioral features, content features, and attribute features are extracted from deep features, AI-generated tags, and publisher metadata. By fusing classification models, four types of intents and their corresponding confidence scores are output. The four types of intents include malicious guidance, active dissemination, unintentional publishing, and AI-generated errors.

[0094] A causal graph is constructed, with sensitive features, AI-generated tags, intent, and publishing scenarios as cause nodes and whether the information is sensitive as effect nodes. The topology of the causal graph is designed in combination with domain rules. Counterfactual reasoning is performed based on the structural causal model (SCM) to eliminate false associations, quantify the confidence of true sensitive associations, and finally obtain the result of sensitive information determination.

[0095] Based on the results of sensitive information determination, intent, and intent confidence score, the weights of the sensitivity type severity, potential scope of dissemination, degree of malicious intent, and AI-generated risk enhancement are determined and the total risk score is calculated. The risk level is then determined based on the total risk score.

[0096] It is understood that the embodiments of this application accurately identify the content dissemination intent through multi-feature fusion, providing a key dimension for risk assessment. By utilizing causal graphs and counterfactual reasoning techniques, it overcomes the limitations of traditional correlation analysis, effectively distinguishing between genuine sensitive associations and false statistical associations, and improving the interpretability and accuracy of sensitive information identification. Based on the weighted calculation of multi-dimensional risk factors, a quantitative risk level is generated, providing a scientific basis for early warning decisions.

[0097] In step S105, with the goal of minimizing intervention costs and maximizing risk control, the optimal intervention plan is dynamically generated from the strategy library by combining the risk level, and the sensitive video is intervened according to the optimal intervention plan.

[0098] Specifically, a dual-objective equilibrium model is constructed, transforming minimum intervention cost and maximum risk control into quantifiable indicators, assigning target weights, and establishing a comprehensive effectiveness scoring function. Based on the risk level and the comprehensive effectiveness scoring function, the optimal intervention plan is dynamically generated from a strategy library. The system then interfaces with the core system of the online media platform, executing the optimal intervention plan in real time via a standardized API interface, and simultaneously recording the entire intervention process log.

[0099] It is understood that this application's embodiments, by constructing a dual-objective equilibrium model, transform the minimum intervention cost and maximum risk control into quantifiable indicators, achieving a scientific balance between governance effectiveness and operational costs. Based on the risk level and comprehensive performance function, the optimal intervention plan is dynamically generated from the strategy library, ensuring that the response measures are both accurate and economical. Seamless integration with the platform system through standardized interfaces ensures real-time handling capabilities while relying on full-process log recording to construct an auditable and optimizable decision-making closed loop.

[0100] This application proposes an intelligent identification and early warning method for sensitive information in online media videos. By collecting video streams, bullet screen / comment text, and publisher metadata, and combining multi-dimensional deep feature analysis of visual, audio, and textual data, it can specifically identify AI-generated content, covering sensitive information forms that are easily overlooked by traditional technologies. This effectively addresses the governance challenges of AI-synthesized false sensitive content. Employing causal association diagnosis and counterfactual reasoning techniques, it can deeply mine the publisher's true dissemination intentions and the logical connections within the content context, outputting a risk level. This significantly reduces the probability of misjudgment and omission in sensitive information identification, improving the interpretability and credibility of the identification results. By balancing the minimum intervention cost with the maximum risk control through a dynamic optimal intervention scheme, it avoids the drawbacks of a one-size-fits-all intervention model, such as high platform operating costs and disruption to normal user experience. It also enables differentiated and precise intervention based on the risk level, effectively curbing the spread of sensitive information and ensuring the compliance of online content while minimizing unnecessary resource consumption and user interference. Therefore, it solves the problems of incomplete multimodal information analysis and the lack of AI-generated content identification mechanisms in existing technologies.

[0101] Figure 3 A schematic diagram of the structure of an electronic device provided in an embodiment of this application. The electronic device may include: The memory 301, the processor 302, and the computer program stored on the memory 301 and capable of running on the processor 302.

[0102] When the processor 302 executes the program, it implements the intelligent identification and early warning method for sensitive information in network media video provided in the above embodiments.

[0103] Furthermore, electronic devices also include: Communication interface 303 is used for communication between memory 301 and processor 302.

[0104] The memory 301 is used to store computer programs that can run on the processor 302.

[0105] The memory 301 may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk storage device.

[0106] If the memory 301, processor 302, and communication interface 303 are implemented independently, then the communication interface 303, memory 301, and processor 302 can be interconnected via a bus to complete communication between them. The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of representation, Figure 3 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0107] Optionally, in a specific implementation, if the memory 301, processor 302, and communication interface 303 are integrated on a single chip, then the memory 301, processor 302, and communication interface 303 can communicate with each other through an internal interface.

[0108] Processor 302 may be a central processing unit (CPU), an application specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of this application.

[0109] This application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the above-described method for intelligent identification and early warning of sensitive information in network media videos.

[0110] This application also provides a computer program product, which stores a computer program that, when executed by a processor, implements the above-described method for intelligent identification and early warning of sensitive information in network media videos.

[0111] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0112] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "N" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0113] Any process or method described in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more N executable instructions for implementing custom logic functions or processes, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as should be understood by those skilled in the art to which embodiments of this application pertain.

[0114] It should be understood that the various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, the N steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0115] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

Claims

1. A smart identification and early warning system for sensitive information in network media videos, characterized in that, include: The module includes a data acquisition module, a video information extraction module, an AIGC video detection module, a causal inference and evaluation module, and an early warning module. The data acquisition module is used to collect online media video data, which includes video stream data, bullet screen / comment text, and publisher metadata. The video information extraction module is used to perform multi-dimensional analysis on the collected video stream data and bullet screen / comment text to obtain deep features. The multi-dimensional features include visual dimension, audio dimension, and text dimension. The AIGC video detection module is used to identify whether the content is generated by AIGC technology based on the deep features. If so, it is marked with an AI generation tag. The causal inference assessment module is used to diagnose the causal relationship between intent and context based on the deep features, AI-generated tags, and publisher metadata. It performs in-depth analysis through counterfactual reasoning technology to determine whether the information is sensitive and assess the risk level. The early warning module is used to match and dynamically generate the optimal intervention plan from the strategy library with the goal of minimizing intervention cost and maximizing risk control, and to intervene in sensitive videos according to the optimal intervention plan.

2. The intelligent identification and early warning system for sensitive information in network media videos according to claim 1, characterized in that, The data acquisition module includes: a data acquisition unit and a data preprocessing unit, wherein... The data acquisition unit is used to crawl video stream data using web crawler technology and obtain bullet screen / comment text and publisher metadata through the platform's open API; The data preprocessing unit is used to clean and align the acquired data. For video stream data, it filters valid frames using a frame quality assessment algorithm, removes stuttering frames, duplicate frames, and black frames, suppresses environmental noise using an improved spectral subtraction method, and unifies the sampling rate and encoding format. For text data, it filters garbled characters, meaningless characters, and spam content using regular expressions. Based on timestamps and video IDs, it establishes a multi-dimensional association index of video, text, and publisher metadata, generating structured data with unified identifiers.

3. The intelligent identification and early warning system for sensitive information in network media videos according to claim 1, characterized in that, The video information extraction module includes: a multi-dimensional parsing unit and a deep feature fusion unit, wherein... The multi-dimensional parsing unit is used to extract the spatial and temporal features of video images using a pre-trained video understanding model, extract the acoustic and deep semantic features of video sound using a pre-trained audio neural network, convert video sound into text to obtain video subtitle information, process the video subtitle information and bullet screen / comment text to generate text semantic vectors, extract core topics through the LDA model, calculate sentiment polarity and sentiment intensity, and obtain text features. The deep feature fusion unit is used to stitch together video images, video audio, and text features to obtain deep features.

4. The intelligent identification and early warning system for sensitive information in network media videos according to claim 1, characterized in that, The AIGC video detection module includes: an artifact feature detection unit, a semantic consistency verification unit, and a labeling unit, wherein... The artifact feature detection unit is used to capture and quantify AIGC generation traces from multiple dimensions to obtain an artifact confidence score. In the visual dimension, it detects common AIGC detail defects based on the spatial features of the video frame and identifies inter-frame motion anomalies based on the temporal features of the video frame. An outlier detection algorithm is used to quantify the degree of artifacts and generate a visual artifact score. In the audio dimension, it determines whether there are traces of AI synthesized speech based on the acoustic features and deep semantic features of the video sound and generates an audio artifact score. In the text dimension, it identifies typical defects of AIGC text based on the text semantic vector and core topics and obtains a text artifact score. The artifact confidence score is generated by combining the multi-dimensional artifact scores through weighted calculation. The semantic consistency verification unit is used to verify video-audio consistency, video-text consistency, and content-comment sentiment consistency, and generates a semantic consistency score based on the verification results. The video-audio consistency is used to verify whether the video content and the sound match, the video-text consistency is used to verify whether the video content and the text semantic features are consistent, and the content-comment sentiment consistency is used to verify whether the emotional tone of the video content matches the sentiment of the bullet screen / comment. The labeling unit is used to output a binary decision result and an overall confidence score through the decision engine based on the artifact confidence score and the semantic consistency score. The binary decision result includes AI-generated and non-AI-generated results. If the binary decision result is AI-generated, then an AI-generated label is used.

5. The intelligent identification and early warning system for sensitive information in network media videos according to claim 1, characterized in that, The causal inference assessment module includes: an intent diagnosis unit, a causal analysis unit, and a risk assessment unit, wherein... The intent diagnosis unit is used to extract behavioral features, content features, and attribute features from deep features, AI-generated tags, and publisher metadata. By fusing classification models, it outputs four types of intents and corresponding confidence scores. The four types of intents include malicious guidance, active dissemination, unintentional publishing, and AI-generated errors. The causal analysis unit is used to construct a causal graph, with sensitive features, AI-generated tags, intent, and publishing scenarios as cause nodes and whether it is sensitive information as effect nodes. It designs the topology of the causal graph in combination with domain rules, performs counterfactual reasoning based on the structural causal model (SCM), eliminates false associations, quantifies the confidence of real sensitive associations, and finally obtains the sensitive information judgment result. The risk assessment unit is used to determine the weights and calculate the total risk score based on the sensitive information judgment result, intent, and intent confidence score, from the dimensions of sensitivity type severity, potential scope of dissemination, degree of malicious intent, and AI-generated risk bonus, and to determine the danger level based on the total risk score.

6. The intelligent identification and early warning system for sensitive information in network media videos according to claim 1, characterized in that, The early warning module includes: a dual-objective modeling unit, an early warning strategy generation unit, and an intervention execution unit, wherein... The dual-objective modeling unit is used to construct a dual-objective equilibrium model, which transforms the minimum intervention cost and the maximum risk control into quantifiable indicators, assigns target weights, and establishes a comprehensive performance scoring function. The early warning strategy generation unit is used to match and dynamically generate the optimal intervention plan from the strategy library based on the risk level and comprehensive effectiveness score function; The intervention execution unit is used to connect to the core system of the online media platform, execute the optimal intervention plan in real time through a standardized API interface, and record the entire intervention process log simultaneously.

7. A method for intelligent identification and early warning of sensitive information in network media videos, characterized in that, Includes the following steps: Collect online media video data, which includes video stream data, bullet screen / comment text, and publisher metadata; The collected video stream data and bullet screen / comment text are analyzed in multiple dimensions to obtain deep features. The multiple dimensions include visual dimension, audio dimension, and text dimension. Based on the deep features, determine whether the content is generated by AIGC technology; if so, mark it with an AI-generated tag. Based on the aforementioned deep features, AI-generated tags, and publisher metadata, the intent and contextual causal relationship are diagnosed. In-depth analysis is performed using counterfactual reasoning techniques to determine whether the information is sensitive and to assess its risk level. With the goal of minimizing intervention costs and maximizing risk control, the optimal intervention plan is dynamically generated from the strategy library by combining the risk level, and the sensitive video is intervened on according to the optimal intervention plan.

8. An electronic device, characterized in that, include: The system includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the intelligent identification and early warning method for sensitive information in network media videos as described in claim 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, The program is executed by the processor to implement the intelligent identification and early warning method for sensitive information in network media videos as described in claim 7.

10. A computer program product, comprising a computer program or instructions, characterized in that, When the computer program or instructions are executed, they implement the intelligent identification and early warning method for sensitive information in network media videos as described in claim 7.

Images