Short video sensitive content detection method, system and device combining content features and context association and medium

Abstract

The invention provides a short video sensitive content detection method, system and device combining content features and context association and a medium, belongs to the technical field of video processing, and aims to extract multi-modal content features and basic statistical features of a video. And carrying out rapid preliminary screening based on basic statistical characteristics, and filtering obvious normal contents. Suspicious contents are detected by using multi-modal content features, sensitive categories are identified, and suspicious fragments are positioned. And obtaining context features such as user behaviors, propagation networks and spatial and temporal distribution. And integrating the accurate detection result and the context features to carry out comprehensive research and judgment, and generating a comprehensive abnormal score. And judging whether the video is sensitive or not according to the score, and outputting a detection result containing categories and suspicious fragments. Through progressive processing, computing resources are concentrated in high-risk samples, and the detection efficiency is improved. And in combination with correlation analysis of the content and the context, complex violation scenes which are difficult to find by content analysis are effectively identified, and the misjudgment rate is reduced.

Description

Technical Field

[0001] This invention belongs to the field of video processing technology, specifically relating to a method, system, device, and medium for detecting sensitive content in short videos that combines content features and contextual association. Background Technology

[0002] With the rapid development of short video platforms, sensitive content detection has become a key technology for maintaining cybersecurity and social stability. Short videos contain multiple modalities of information, including visual, auditory, and textual information, providing a complex medium for the covert dissemination of sensitive content.

[0003] Current technologies for detecting sensitive content in short videos mainly include the following types of solutions: detection methods based on content analysis and auxiliary judgment methods based on user behavior.

[0004] Content-based detection methods primarily analyze the content features of the video itself, typically focusing on the information sources within the video. For example, image classification methods based on visual features use convolutional neural networks (CNNs) to identify inappropriate content, while speech recognition methods based on audio features use recurrent neural networks (RNNs) to detect sensitive speech. User behavior-based auxiliary judgment methods analyze static features such as users' historical posting records, number of violations, and account age to identify and monitor high-risk users. These methods usually combine user characteristics as supplementary information with the content detection results.

[0005] For the aforementioned content analysis-based detection methods and user behavior-based auxiliary judgment methods, a threshold needs to be set as the judgment standard. The distribution of short video content will change dynamically with time, hot events and other factors. The threshold set by the above methods cannot adapt to this change. When sensitive content appears or data distribution drifts, the false alarm rate or false negative rate will increase, making it difficult to maintain a stable detection effect in the long term.

[0006] The two methods mentioned above ignore contextual information, leading to the failure to detect hidden risks. This mainly targets the content characteristics of short videos themselves. In real-world scenarios, some content may not contain obvious sensitive information, but the publisher may be a frequently violating account, the dissemination path may show signs of manipulation, or the content may be concentrated in sensitive periods or regions. Such hidden risks cannot be identified through content detection alone, creating blind spots.

[0007] Moreover, full-scale accurate detection leads to excessive resource consumption. Performing the same complex accurate detection process on all short videos results in a large number of normal short videos without any sensitive signs occupying most of the computing resources, causing delays in processing potential sensitive content that truly needs to be detected, and failing to meet the performance requirements of real-time detection of large-scale short videos. Summary of the Invention

[0008] This invention provides a method for detecting sensitive content in short videos that combines content features and contextual association. The collaborative analysis of contextual features improves the accuracy of discrimination, the configuration of association representations enhances the ability to identify complex scenes, and hierarchical decision-making and cross-modal interaction ensure the real-time performance of sensitive content detection in short videos, thereby achieving high-precision identification of sensitive content.

[0009] The methods include: S101, Extract multimodal content features and basic statistical features from the short video; wherein, the multimodal content features include visual features, audio features and text features; S102, perform initial screening based on the basic statistical features to obtain an initial screening score; if the initial screening score is lower than the first threshold, it is determined to be normal content and the process ends; otherwise, proceed to step S103. S103, Detect the multimodal content features transmitted in step S102 to obtain preliminary sensitive categories and locate suspicious segments; S104, Obtain the contextual features of the short video; wherein, the contextual features include user behavior features, propagation network features, and spatiotemporal distribution features; S105, based on the preliminary sensitive category obtained in step S103, the suspicious fragment, and the contextual features obtained in step S104, a comprehensive judgment is made to obtain a comprehensive anomaly score; S106, Based on the comparison result between the comprehensive anomaly score and the second threshold, determine whether the short video is sensitive content, and output the detection result including preliminary sensitive category and suspicious segments.

[0010] This invention also provides a short video sensitive content detection system that combines content features and contextual association, the system comprising: The feature extraction module is used to extract multimodal content features and basic statistical features of short videos; wherein, the multimodal content features include visual features, audio features and text features; The statistical initial screening module performs an initial screening based on the basic statistical features to obtain an initial screening score; if the initial screening score is lower than a first threshold, it is determined to be normal content and the process ends; otherwise, the content detection module is executed. The content detection module is used to detect the features of the transmitted multimodal content, obtain preliminary sensitive categories, and locate suspicious segments; The context feature acquisition module is used to acquire the context features of the short video; wherein, the context features include user behavior features, propagation network features, and spatiotemporal distribution features; The comprehensive analysis module performs a comprehensive analysis based on the obtained preliminary sensitive category, the suspicious fragment, and the context features obtained in step S104 to obtain a comprehensive anomaly score. The judgment result output module is used to determine whether the short video is sensitive content based on the comparison result of the comprehensive anomaly score and the second threshold, and output the detection result including the preliminary sensitive category and suspicious segments.

[0011] According to another embodiment of this application, an electronic device is provided, including 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 steps of a short video sensitive content detection method that combines content features with contextual association.

[0012] According to another embodiment of this application, a storage medium is also provided, on which a computer program is stored, which, when executed by a processor, implements the steps of the short video sensitive content detection method that combines content features and context association.

[0013] As can be seen from the above technical solutions, the present invention has the following advantages: This invention provides a method for detecting sensitive content in short videos that combines content features and contextual association. It extracts multimodal and basic statistical features simultaneously, covering risk signals from different dimensions of short videos. Visual, audio, and textual features are combined to capture semantic information, sound information, and linguistic connotations respectively. Basic statistical features can quickly process the intuitive attributes of visuals, audio, and text. The two types of features complement each other, processing semantically sensitive information and identifying anomalous signals. The hierarchical detection process of this invention allocates detection resources. The initial screening stage quickly filters short videos without sensitive signs using statistical features, sending potentially sensitive content to subsequent precise detection, reducing unnecessary computation and improving overall processing speed. The judgment criteria adapt to data changes in different scenarios. The thresholds for both the initial screening and final judgment stages are dynamically adjusted based on feedback from actual detection data, addressing differences in short video content distribution across different time periods and maintaining stable judgment accuracy. This invention integrates content and context for risk identification, correlating external information such as user behavior, propagation networks, and spatiotemporal distribution with short video content features, enabling the identification of hidden sensitive content that cannot be detected solely by content analysis. The output results improve processing efficiency. Attached Figure Description

[0014] To more clearly illustrate the technical solution of the present invention, the accompanying drawings used in the description will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 Flowchart of a method for detecting sensitive content in short videos that combines content features and contextual association; Figure 2A schematic diagram of a short video sensitive content detection system that combines content features and contextual association; Figure 3 This is a schematic diagram of an electronic device. Detailed Implementation

[0016] The short video sensitive content detection method proposed in this invention combines content features and contextual association to obtain a comprehensive representation of video content through multimodal feature fusion preprocessing, adopts a three-layer progressive detection framework to achieve rapid screening and accurate identification, and conducts in-depth analysis of high-risk content through a contextual association analysis module.

[0017] The following describes in detail the method for detecting sensitive content in short videos that combines content features and contextual association. Specific details, such as particular system structures and techniques, are presented for illustrative purposes rather than limiting, to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application can also be implemented in other embodiments without these specific details.

[0018] It should be understood that, when used in this specification, the term "comprising" indicates the presence of the described feature, integral, step, operation, element, and / or component, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components, and / or collections thereof. The terms "comprising," "including," "having," and variations thereof all mean "including but not limited to," unless otherwise specifically emphasized.

[0019] The terms "one embodiment" or "some embodiments" used in this application mean that one or more embodiments of this application include the specific features, structures, or characteristics described in that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this application do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized.

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Please see Figure 1 The diagram shows a flowchart of a short video sensitive content detection method that combines content features and contextual association in a specific embodiment. The method includes: S101, extract multimodal content features and basic statistical features of the short video; wherein, the multimodal content features include visual features, audio features and text features.

[0022] In some embodiments, visual, audio, and textual features of multimodal content can be extracted.

[0023] Specifically, a multimodal feature fusion preprocessing module can be configured. This module, acting as the front-end processing unit of the detection system, converts the input short video into a structured multimodal feature representation. The multimodal feature fusion preprocessing module receives the raw short video data and, through four processing steps—feature extraction, intra-modal enhancement, cross-modal alignment, and adaptive fusion—outputs a unified multimodal fusion representation vector. and basic statistical characteristics This is for use by subsequent detection modules.

[0024] Specifically, the multimodal feature fusion preprocessing module extracts multimodal information from the input short video, obtaining feature representations in three modalities: visual, audio, and text. Visual Feature Extraction: A 3D convolutional neural network is used to extract spatiotemporal features from a video frame sequence to obtain visual features. It includes semantic information, scene understanding, and action features of the video footage.

[0025] Audio feature extraction: Convert the video audio track into a Mel spectrogram and extract audio features using an audio encoder. It includes information such as voice content, background music, and emotional rhythm; Text Feature Extraction: Text information in the video is extracted through automatic speech recognition and optical character recognition, and encoded into text features using a pre-trained language model. It includes linguistic information, semantic content, etc.

[0026] Basic statistical features are based on calculating a set of lightweight statistical features. Features that can be quickly calculated, including image brightness, color saturation, volume changes, number of sensitive words hit in text, and video duration, are used for subsequent rapid initial screening.

[0027] As can be seen, multimodal content feature extraction utilizes visual image texture, audio frequency variations, and textual semantic information, employing targeted deep learning or signal processing techniques to capture semantic and morphological information at the content level of short videos. Basic statistical features, on the other hand, quickly filter out feature signals with sensitive content by describing the video's fundamental attributes. This approach covers both semantic associations and basic attributes.

[0028] S102, perform initial screening based on the basic statistical features to obtain an initial screening score; if the initial screening score is lower than the first threshold, it is determined to be normal content and the process ends; otherwise, proceed to step S103.

[0029] In some embodiments, basic statistical features are utilized to quickly distinguish between clearly normal and sensitive short videos. Data interference is eliminated through normalization and outlier handling. Optionally, dynamic weights are configured to ensure that the initial screening focuses on the most discriminative features at present, and threshold-based settings can adapt to the distribution of detection data at different time periods, balancing the risks of false positives and false negatives.

[0030] S103, the multimodal content features transmitted in step S102 are detected to obtain preliminary sensitive categories and locate suspicious segments.

[0031] In some embodiments, self-attention ensures that intramodal temporal correlations and intermodal semantic consistency are fully captured. Deep semantic feature processing using a classification network can determine sensitive categories. This embodiment can also locate the time interval that contributes most to sensitivity based on temporal attention, thus locking down suspicious segments. This achieves complementarity between multi-label and single-category determination, covering scenarios where multiple sensitive elements coexist.

[0032] S104, Obtain the contextual features of the short video; wherein, the contextual features include user behavior features, propagation network features, and spatiotemporal distribution features.

[0033] In some embodiments, considering that the risk of short videos lies not only in the content itself, but also in the behavior of the publisher, the method of dissemination, and the spatiotemporal context of the posting, this approach processes user behavior characteristics to extract risk tendencies reflecting the publisher, captures abnormal patterns of artificial manipulation in the dissemination network, and mines posting behavior through time and related characteristics. This supplements risk information from dimensions beyond content, forming a risk assessment dimension. For example, it can accurately identify videos with ambiguous content but abnormal dissemination. The three risk factors of user, dissemination, and time provide a more comprehensive basis for integrated judgment.

[0034] S105, based on the preliminary sensitive category obtained in step S103, the suspicious fragment, and the contextual features obtained in step S104, a comprehensive anomaly score is obtained.

[0035] In some embodiments, by associating and fusing content and contextual features, the inherent correlation between content and contextual features is explored to achieve a comprehensive risk assessment. Bidirectional cross-attention ensures accurate matching between content and contextual information, dynamic weight allocation gives greater weight to risk dimensions with high contribution, and weighted fusion integrates multi-dimensional risk signals to obtain a comprehensive score reflecting the risk level of the short video.

[0036] S106, Based on the comparison result between the comprehensive anomaly score and the second threshold, determine whether the short video is sensitive content, and output the detection result including the preliminary sensitive category and the suspicious segment.

[0037] Based on the above-described method for detecting sensitive content in short videos that combines content features and contextual association, in one embodiment of the present invention, based on step S101 of the above method, the following will provide a possible embodiment and its specific implementation plan in a non-limiting manner. S101 specifically includes the following: S1011: Sample short video frame sequences to construct a sample set, use a 3D convolutional neural network to extract features from the sample set, and output visual features.

[0038] In some embodiments, the sampling frequency and duration range of the short video frame sequence are determined. Frame images are extracted from the original short video, and abnormal frames that are blurred, overexposed, or underexposed are removed to construct a sample set containing spatiotemporally continuous information. A 3D convolutional neural network is selected as the feature extraction carrier. The constructed sample set is input into the 3D convolutional neural network, which then undergoes convolution, activation, and pooling operations to simultaneously capture the spatial texture features and temporal dynamic change features of the frame sequence. Finally, the high-dimensional features are mapped to fixed-dimensional vectors through a fully connected layer, thus obtaining the visual features.

[0039] Optionally, since different modalities have different temporal granularities, such as a video with 30 frames per second but possibly corresponding to only the pronunciation of a single word, the modal features are aligned and fused. Intramodal enhancement can be performed: for , , By applying self-attention mechanisms to enhance the temporal correlation within each modality, the following results were obtained. , , .

[0040] Cross-modal alignment: Semantic alignment between modalities is achieved using a cross-attention mechanism. For modality i and modality j, cross-alignment features are calculated: Formula (1) This mechanism allows one modality to dynamically adjust its representation based on the content of another modality, achieving fine-grained semantic alignment.

[0041] S1012: Separate the audio track data of the short video, convert the audio track data into an audio data matrix in the form of a Mel spectrogram, and use an audio encoder to perform feature encoding on the audio data matrix to generate audio features; In some embodiments, single-channel audio track data is extracted from the original short video data using an audio separation algorithm. The audio track data is then resampled to remove high-frequency noise and low-frequency interference. The audio track data is then segmented into frames to obtain a short-time frame sequence. Windowing is applied to each short-time frame to reduce spectral leakage, and a Mel filter bank is used to filter the frequency domain data, generating a Mel spectrogram. An audio encoder based on a Transformer architecture is selected to encode the features of the Mel spectrogram. Alternatively, the Mel spectrogram can be input into the encoder, and a self-attention mechanism can be used to capture audio temporal dependencies. A feedforward neural network then performs feature dimension transformation and nonlinear mapping, ultimately outputting fixed-dimensional audio features.

[0042] Optionally, a gating mechanism can be used to automatically learn the importance weights of each modality. The weights are calculated using a gating network. , , Then weighted fusion: Formula (2) in As a time-series pooling operation, this fusion method can adaptively adjust the contribution of each modality according to the characteristics of different videos, avoiding the limitations of fixed weights.

[0043] S1013: Perform speech-to-text processing on the audio track of the short video and extract the text regions in each frame of the short video. Combine the two processing results to obtain a text information set. Input the text information set into a pre-trained language model for semantic encoding to obtain text features.

[0044] In some embodiments, the audio track data separated in S1012 is processed using an automatic speech recognition technology based on a deep neural network. First, endpoint detection is performed on the audio track data to locate valid speech segments.

[0045] Speech features are mapped to phoneme sequences using an acoustic model, and then the phoneme sequences are converted into text sequences using a language model, resulting in speech-to-text conversion.

[0046] In this embodiment, for each frame of the image extracted from the sample set of S1011, the edge detection algorithm is used to locate the text region. The text region is then binarized, denoised, and segmented. The text region is then input into the optical character recognition model to complete character recognition, and the text extraction result of each frame of the image is obtained. The speech-to-text result and the text extraction result of each frame of the image are deduplicated, concatenated, and grammatically corrected to form a text information set. The BERT pre-trained language model is selected as the semantic encoding carrier. After being converted into word embedding vectors, it is input into the model. The multi-layer bidirectional Transformer encoder of the model captures the semantic association and contextual information of the text, and outputs fixed-dimensional text features from the model. These text features will be used for cross-modal alignment, providing a textual basis for inter-modal semantic matching in formula (1).

[0047] S1014: Calculate the mean brightness and variance of color saturation for each frame of the short video frame sequence, calculate the temporal change rate of volume amplitude for the audio track data, perform sensitive word matching statistics on the text information set obtained in S1013 to obtain the number of sensitive words hit in the text, read the duration parameter in the short video metadata, and integrate the calculation results to form basic statistical features.

[0048] It can be seen that, for the resampled audio track data obtained by S1012, the root mean square value of the audio signal within the frame is calculated, and the temporal change rate of the volume amplitude is obtained by the ratio of the volume amplitude difference between adjacent short frames to the time interval. The maximum value of this change rate is taken as the volume change feature.

[0049] Based on a pre-defined sensitive word dictionary, a string matching algorithm is used to perform word-by-word matching on the text information set obtained from S1013, and the number of successfully matched sensitive words is counted as the text sensitive word hit count. By parsing the metadata fields of the short video file, the duration parameter of the short video is read. The five types of features obtained—brightness, color saturation, volume change, text sensitive word hit count, and video duration—are integrated into a one-dimensional feature vector, which yields the basic statistical features.

[0050] In one embodiment of the present invention, based on step S102, the following will provide a possible embodiment and its specific implementation will be described in a non-limiting manner. S102 specifically includes the following: S1021: Normalize each dimension of the basic statistical features extracted in S101, and generate a standardized basic statistical feature set after removing outlier data points.

[0051] S1022: Retrieve historical detection sample datasets, statistically analyze the proportion of sensitive content exceeding the temporary threshold for each dimension of standardized basic statistical features, obtain feature importance weights, and update the weight vector according to the set sample window.

[0052] In some embodiments, a temporary threshold is preset for each dimension's basic statistical features, which is set to the 75th quantile of the historical sample feature values ​​for the corresponding dimension. The most recent 500 historical detection samples are retrieved, and for each dimension in the standardized basic statistical feature set, the number A of samples whose feature values ​​exceed the temporary threshold is counted, along with the number B of these A samples that are ultimately identified as sensitive content.

[0053] The importance weight of this dimension is calculated according to the weight quantization logic ω=B / A. If A=0, no sample exceeds the threshold, then the weight is set to 0.05.

[0054] Optionally, an update cycle is set for processing 500 new detection samples. The weights of each dimension are recalculated and the historical weights are overwritten to form a real-time updated weight vector W=[ω1,ω2,ω3,ω4,ω5], where ω1 to ω5 correspond to the weights of screen brightness, color saturation, volume change, number of sensitive words hit in text, and video duration, respectively.

[0055] S1023: Perform a weighted summation of the standardized basic statistical feature set and the corresponding importance weights, input the result into the sigmoid activation function to complete the normalization, and output the initial screening score.

[0056] In some embodiments, the standardized basic statistical feature set obtained in S1021 is converted into a feature vector X=[x1,x2,x3,x4,x5].

[0057] Where x1 to x5 are the standardized feature values ​​of each dimension. Perform a weighted summation operation S_sum=ω1x1+ω2x2+ω3x3+ω4x4+ω5x5 to obtain the feature weighted sum.

[0058] Input Ssum into the sigmoid activation function to output the initial screening score S1. The score is constrained within the interval [0,1], and the score directly corresponds to the degree of suspicion of the short video; the higher the score, the stronger the suspicion. After calculation, the correspondence between the initial screening score and the feature weighted sum is retained.

[0059] S1024: Based on the changes in false alarm rate and false negative rate from historical detection data, the current first threshold is calculated using the following formula:

[0060] in As the baseline threshold, This represents the change in the false alarm rate relative to the baseline within the most recent time window. The change in the underreporting rate and This is the adjustment coefficient; Compare the initial screening score with the current first threshold. If the score is lower, the content is judged to be normal and the detection process is terminated.

[0061] In some embodiments, a reference threshold for setting the first threshold is defined. =0.3, false alarm rate adjustment coefficient α=0.2, false alarm rate adjustment coefficient β=0.15.

[0062] Retrieve historical detection data from the most recent hour, calculate the false alarm rate (FPR) and false negative rate (FNR) within this time window, and determine the baseline false alarm rate (FPR). base Baseline FNR (Failure Rate) base .

[0063] According to the formula

[0064] Calculate the current first threshold, where FPR(t) = FPR(t) - FNR base This represents the change in the false alarm rate relative to the baseline. FNR(t) = FNR(t) - FPR base This represents the change in the underreporting rate relative to the baseline.

[0065] The initial screening score S1 obtained in S1023 is compared with the current first threshold θ1(t). If S1 < θ1(t), a detection result of normal content is generated, and the subsequent detection steps are terminated.

[0066] If S1≥θ1(t), then the original short video data, standardized basic statistical feature set, initial screening score and weight vector are passed to step S103.

[0067] Optionally, a temporary threshold of the 75th quantile can be selected to balance the discriminative power and coverage of features. This improves the processing efficiency of the detection system and ensures the reasonableness of the initial screening results.

[0068] In one embodiment of the present invention, based on step S103, the following is a possible embodiment and its specific implementation will be described in a non-limiting manner. S103 specifically includes the following: S1031: Perform self-attention enhancement processing on the visual features, audio features, and text features of S102 respectively, complete cross-modal semantic alignment based on the cross-attention mechanism, and then obtain a unified multimodal fusion representation through gating fusion.

[0069] In some embodiments, the visual features Fv, audio features Fa, and text features Ft transmitted in S102 are retrieved, and an independent self-attention calculation unit is configured for each modality feature.

[0070] Optionally, the self-attention unit generates query, key, and value vectors through linear transformation, and obtains self-attention weights through dot product operation and softmax normalization. The weights are then weighted and summed with the V vector to output enhanced features Hv, Ha, and Ht. Three sets of cross-attention units are constructed to achieve bidirectional semantic alignment between Hv and Ha, Hv and Ht, and Ha and Ht, respectively, according to the formula...

[0071] Calculate the cross-alignment features, where i and j are different modal identifiers.

[0072] The three sets of cross-aligned features are concatenated and input into the gated fusion unit. The gated network calculates the modal weights gv, ga, and gt, according to the formula.

[0073] After weighted fusion, Pool employs average pooling to output a multimodal fusion representation with a dimension of 1024. This ensures that the representation covers both unimodal temporal associations and cross-modal semantic associations.

[0074] S1032: Input the multimodal fusion representation into the classification branch of the multi-task learning framework and output the probability distribution of sensitive categories; determine each sensitive category through the multi-label branch and generate a binary multi-label set.

[0075] In some embodiments, a multi-task learning framework is constructed, wherein the classification branch consists of 5 layers of Transformer coding blocks and 1 fully connected layer.

[0076] The multimodal fusion characterization obtained from S1031 The input classification branch, after semantic feature enhancement by the encoding block, is mapped to a preset sensitive category space through a fully connected layer. This preset sensitive category space involves sensitive content such as violence and terror.

[0077] The output class probability distribution Pcontent is normalized using softmax, with the sum of the probabilities of each class being 1. The multi-label branch and the classification branch share the parameters of the first three Transformer encoding blocks. Through independent fully connected layers and sigmoid activation functions, a binary judgment result of 0 or 1 is output for each sensitive class, forming a multi-label set Lmulti. The class corresponding to the maximum probability in Pcontent is recorded as a candidate sensitive class, while all class identifiers with a value of 1 in Lmulti are retained to ensure that subsequent evaluation covers both a single dominant sensitive class and multiple sensitive elements.

[0078] S1033: Based on multimodal fusion representation, reconstruct the temporal feature sequence, input the regression branch of the multi-task learning framework to obtain the sensitivity score; analyze the weight of each time step through the temporal attention mechanism to determine the time interval of suspicious segments.

[0079] In some embodiments, the original temporal sequences of each modality are extracted from Hv, Ha, and Ht in S1031, and the temporal feature sequences are obtained by aligning them according to the video frame rate.

[0080] The regression branch is constructed using a fully connected layer and a linear activation function. It takes the global average pooling result of the time series feature sequence as input and outputs the sensitivity score Sseverity with a value range of [0,1].

[0081] A temporal attention mechanism unit is constructed, taking the temporal feature sequence as input. The cosine similarity between each time step feature and the global feature is calculated, and then normalized using softmax to obtain the temporal attention weight αt. An attention weight threshold of 0.65 is set, and consecutive time steps with weights exceeding the threshold are classified as suspicious time intervals. These intervals are then converted into a readable format based on the video frame rate. The maximum attention weight max(αt) within each interval is recorded.

[0082] S1034: Integrate the maximum probability of sensitive categories, sensitivity scores, maximum temporal attention weights, and maximum multi-label values, analyze the comprehensive anomaly score, determine the preliminary sensitive categories, and output the specific temporal information of suspicious segments.

[0083] In some embodiments, preset weighting coefficients γ1=0.3, γ2=0.25, γ3=0.25, and γ4=0.2, and the sum of these weighting coefficients is 1.

[0084] Extract the maximum value max(Pcontent) from Pcontent, extract the maximum value max(Lmulti) from Lmulti, and retrieve Sseverity and max(αt) obtained from S1033.

[0085] The comprehensive abnormality score is calculated using the following formula.

[0086]

[0087] Formula (4) This embodiment For boundary samples in the 0.55-0.65 range, confidence is calibrated using temperature scaling with a temperature parameter τ=1.1. After calibration, the final comprehensive anomaly score is obtained. The category corresponding to the maximum probability in Pcontent is determined as the preliminary sensitive category. Combined with the temporal information of the suspicious fragment and max(αt), the detection result is formed. The weighted fusion of formula (4) here integrates classification confidence, sensitivity, temporal anomaly, and multi-label information, improving the reliability of the comprehensive anomaly score.

[0088] In one embodiment of the present invention, based on step S104, the following will provide a possible embodiment and its specific implementation will be described in a non-limiting manner. S104 specifically includes the following: S1041: Collect historical behavior data of users who publish short videos, extract static behavior statistical features and construct a behavior time series, and encode the time series using a Transformer encoder to generate user behavior representations.

[0089] In some embodiments, historical behavioral data of short video posting users is collected, and the extracted static behavioral statistical features can include posting frequency, historical violation count, account age, and interaction rate. A time-series sequence of user behavior is constructed in chronological order, with each sequence node containing four fields: posting timestamp, content type tag, review result tag, and interaction data. A Transformer encoder encodes the time-series sequence, with each layer employing a self-attention mechanism and a feedforward neural network. The time-series sequence is converted into word embedding vectors and input into the encoder. The self-attention mechanism captures long-term dependencies between behaviors, and after layer normalization and linear transformation, a user behavior representation is output.

[0090] S1042: Collect propagation link data of short videos, construct a propagation network graph, analyze the topological features of the graph and fit the propagation curve, and obtain the propagation network representation through feature encoding.

[0091] In some embodiments, data on the propagation chain after the short video is published is collected, including a list of users who forwarded the video, the following relationships between users, and the forwarding timestamp.

[0092] Using forwarding users as nodes and follow or interaction relationships between users as edges, an undirected propagation network graph is constructed. The graph's topological features—degree centrality, betweenness centrality, and clustering coefficient—are calculated. The propagation curve is fitted using the SIR epidemic propagation model. Forwarding timestamp data is input to obtain the theoretical propagation curve, and the mean squared error between the actual and theoretical curves is calculated as a propagation anomaly feature. The topological features and propagation anomaly features are concatenated into a high-dimensional feature vector, which is then compressed and nonlinearly mapped through a fully connected layer to output the propagation network representation.

[0093] S1043: Extract the release time and geographic location information of short videos, analyze the temporal distribution characteristics and spatial clustering characteristics respectively, and obtain the spatiotemporal distribution representation after performing spatiotemporal joint analysis.

[0094] In some embodiments, the release timestamps in the short video metadata are extracted, the release probability of each time period within a day is counted, the time entropy is calculated to quantify the concentration of release time distribution, and whether it is in a sensitive period is marked. Sensitive periods are major holidays, anniversaries, and emergencies.

[0095] This embodiment obtains geographical location information by resolving the IP address of the user who posted the short video, uses the DBSCAN density clustering algorithm to detect geographical location clustering patterns, and calculates the cluster density as a spatial clustering feature.

[0096] This embodiment performs spatiotemporal joint analysis, counts the number of similar short videos published within the same time period and the same geographical cluster, calculates the spatiotemporal clustering coefficient, and concatenates the temporal distribution characteristics, spatial clustering characteristics, and spatiotemporal clustering coefficient, outputting a spatiotemporal distribution representation after linear transformation.

[0097] S1044: The user behavior representation, propagation network representation, and spatiotemporal distribution representation are concatenated and fused through a fully connected layer and layer normalization to output contextual features.

[0098] In some embodiments, the user behavior representation obtained in S1041, the propagation network representation obtained in S1042, and the spatiotemporal distribution representation obtained in S1043 are retrieved and sequentially concatenated into a fused feature vector.

[0099] In this embodiment, the feature vectors are fused and input into the fully connected layer. The first layer has an output dimension of 1024, the second layer has an output dimension of 768, and the third layer has an output dimension of 512. After each fully connected layer, a layer normalization operation is performed to stabilize the feature distribution.

[0100] This embodiment introduces nonlinearity through the ReLU activation function to enhance feature representation capabilities and ultimately output contextual features.

[0101] In one embodiment of the present invention, based on step S105, the following will provide a possible embodiment and its specific implementation will be described in a non-limiting manner. S105 specifically includes the following: S1051: Organize and encode the preliminary sensitive categories and suspicious fragment-related features obtained in S103, perform dimensional alignment and standardization processing with the context features output in S104, and generate feature pairs suitable for association modeling.

[0102] In some embodiments, the preliminary sensitivity category, the temporal attention weights of suspicious segments, and the sensitivity score output by S103 are retrieved. The preliminary sensitivity category is converted into a one-hot encoded vector, and the temporal attention weight sequence of the suspicious segments is compressed into a one-dimensional feature through average pooling. This feature is then concatenated with the sensitivity score to form a content feature vector. .

[0103] The unified context features output by S104 are retrieved, and standardization is performed on both types of features to ensure that the feature vectors have consistent magnitudes. If there is a difference in dimension between the content feature vector and the context feature vector, a fully connected layer is used to map the low-dimensional features to the high-dimensional ones to achieve dimension alignment. Finally, feature pairs (Rfinal, Hcontent) adapted for subsequent association modeling are generated, where Rfinal is the standardized and aligned content feature, and Hcontent is the standardized and aligned context feature.

[0104] S1052: Perform content-guided context feature enhancement and context-guided content feature recalibration, and generate associated representations through multi-level iterative interaction.

[0105] In some embodiments, a cross-attention fusion mechanism is configured, and the first stage performs content-guided context feature enhancement: the content features obtained in S1051 As a query, the context feature Hcontent is used as the key and value, and the enhanced context feature is calculated according to formula (5).

[0106]

[0107] Formula (5) For example, if the video content contains specific political symbols, cross-attention will automatically focus on the user's history of posting politically sensitive content; if the video content is of a general entertainment nature, it will reduce the overall weight of attention to the user's history.

[0108] The second phase involves context-guided content feature recalibration: enhancing the context features... Features of the original content Concatenate the input gated network and calculate the gate coefficient. ,according to

[0109] Formula (6) Complete the recalibration of content features, where ⊙ represents element-wise multiplication. Set up a 3-layer iterative interaction cycle, in which the above enhancement and recalibration operations are executed sequentially. After the iteration, construct a quadruple association representation: concatenate the recalibrated content features, the enhanced context features, their element-wise product and element-wise difference, and compress them into a 512-dimensional association representation Hjoint through a fully connected layer.

[0110] S1053: Based on content features, context features, and association representations, analyze the content risk score, context risk score, and association risk score respectively.

[0111] In some embodiments, three identical multilayer perceptron networks are constructed for calculating three types of risk scores, respectively. Content features aligned after S1051. The first perceptron is input, and the output is a content risk score Scontent with a value in the range [0,1]. The score reflects the sensitivity of the video itself. The aligned contextual feature Hcontent is input into the second perceptron, and the output is a contextual risk score Scontext, which reflects the degree of abnormality of the context, such as the publisher's behavior and the propagation environment.

[0112] In this embodiment, the association representation Hjoint obtained in S1052 is input into the third perceptron, and the association risk score Sjoint is output, reflecting the comprehensive association risk after the content and context are combined. Each perceptron output layer uses the sigmoid activation function to ensure that the score is strictly constrained in the [0,1] interval. At the same time, the neuron response values ​​of each perceptron output layer are recorded to provide feature basis for subsequent dynamic weight calculation.

[0113] S1054: The weights of the three risk scores are analyzed using a neural network, and the comprehensive anomaly score is obtained by weighted fusion according to the following formula, and the detection confidence is calculated:

[0114] Formula (7) In the formula, For content risk score, For context risk score, For associated risk scores, For content risk weighting coefficient, For contextual risk coefficient, This refers to the associated risk coefficient.

[0115] In some embodiments, the neural network is configured to concatenate the response values ​​of the output layer neurons of the three multilayer perceptrons in S1053 as input, and output the dynamic weights w1, w2, and w3 of the three risk scores as 1 through the softmax activation function.

[0116] Calculate the comprehensive abnormality score according to formula (7). The detection confidence level is calculated synchronously: the category probability distribution corresponding to the content risk score in S1053 is obtained, and the confidence level is quantified by calculating the entropy value of this distribution. The lower the entropy value, the more certain the model's judgment. The final output is a comprehensive anomaly score. The corresponding detection confidence level forms a comprehensive judgment data. The weighted fusion of formula (7) allows the comprehensive anomaly score to cover content, context and associated risks, and the output of the confidence level provides a reference for the reliability of the score.

[0117] In one embodiment of the present invention, based on step S106, the following is a possible embodiment and its specific implementation will be described in a non-limiting manner. S106 specifically includes the following: S1061: Based on the changes in false alarm rate and false negative rate of historical preset duration detection data, the current second threshold is calculated using the following dynamic threshold formula: θ2(t)= +α2· FPR2(t)-β2· FNR2(t) α2 is the second threshold baseline value, β2 is the false alarm rate adjustment coefficient, β2 is the false alarm rate adjustment coefficient, FPR2(t) is the false alarm rate during the statistical period, and FNR2(t) is the false alarm rate.

[0118] In some embodiments, a second threshold reference value is set. =0.6, combined with the requirement of prioritizing accuracy in the second layer, the false alarm rate adjustment coefficient α2=0.25 and the false negative rate adjustment coefficient β2=0.1 are configured, α2>β2, to strengthen the weight of false alarm rate on threshold control. The historical detection dataset within the most recent preset time period is retrieved, and the false alarm rate FPR2(t) and false negative rate FNR2(t) within the time period are statistically analyzed. The average false alarm rate of the same time period within the most recent preset time period is extracted as the false alarm rate baseline FPR2base, and the average false negative rate is extracted as the false negative rate baseline FNR2base. Referring to the dynamic adjustment logic of formula (3), according to θ2(t)= +α2· FPR2(t)-β2· FNR2(t) calculates the current second threshold.

[0119] in FPR2(t) = FPR2(t) - FPR2base represents the fluctuation of the false alarm rate relative to the baseline. FNR2(t) = FNR2(t) - FNR2base represents the fluctuation of the false negative rate relative to the baseline.

[0120] To avoid the threshold deviating too much from the reasonable range, a threshold constraint interval of [0.5, 0.7] is set. If the calculated θ2(t) is lower than 0.5, it is forcibly corrected to 0.5, and if it is higher than 0.7, it is corrected to 0.7.

[0121] S1062: Verify the validity of the comprehensive anomaly score output by S105, compare the valid score with the current second threshold to determine the content type, and add a manual review mark based on the detection confidence.

[0122] In some embodiments, the comprehensive anomaly score output by S105 is retrieved. And the detection confidence level. Perform score validity check: if Scores outside the range [0,1] are considered invalid, but are marked as normal content by default and an exception log is recorded.

[0123] If the score is within the valid range, it is compared numerically with the current second threshold θ2(t) calculated in S1061. The comparison rule is: if... If the value is greater than or equal to θ2(t), it is considered sensitive content.

[0124] like <θ2(t) indicates normal content. The detection confidence level is used to assist in the judgment. The confidence level is quantified by the entropy value of the category probability distribution. If the entropy value is higher than 0.8, it indicates high uncertainty in the model's judgment, and a mark requiring manual review should be added to the judgment result to indicate that further verification is needed.

[0125] S1063: Organize the information according to the judgment results and output the organized information.

[0126] In some embodiments, if the content is determined to be sensitive, the preliminary sensitivity category, suspicious fragment information, and sensitivity score are retrieved from S103, and the content risk score Scontent, context risk score Scontext, association risk score Sjoint, and detection confidence are retrieved from S105.

[0127] If the content is determined to be normal, then the video ID, the determination result, and the overall anomaly score are compiled. And the detection confidence level, omitting sensitive and related details.

[0128] All processed data must undergo format validation: time information must be uniformly formatted as hour:minute:second, fractions must be retained to 4 decimal places, and sensitive category names must strictly match the preset specifications to ensure that there is no missing information or incorrect format.

[0129] Optionally, a detection log is generated, the log content of which includes the video ID, processing timestamp, current second threshold θ2(t), and comprehensive anomaly score. Judgment result, confidence level.

[0130] This invention achieves high-precision identification and efficient processing of sensitive content in short videos by constructing a detection method that integrates multimodal feature fusion, hierarchical detection, and contextual correlation analysis. Through a cross-attention mechanism, it achieves deep correlation modeling between content features and contextual features, enabling the identification of hidden sensitive content that is ambiguous but has anomalies in context, thus reducing the false negative rate. The hierarchical detection framework implements a graded processing method that enables rapid screening, accurate identification, and comprehensive judgment, meeting real-time detection requirements. By introducing multi-dimensional features such as user behavior, propagation networks, and spatiotemporal distribution through contextual correlation analysis, it can distinguish between normal and abnormal propagation patterns, effectively reducing false positives. A risk tracing report is generated, providing content risk, contextual risk, and the contribution of key factors, offering quantitative evidence for manual review and management decisions.

[0131] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0132] The following are embodiments of the short video sensitive content detection system that combines content features and context association provided in this disclosure. This system and the short video sensitive content detection method that combines content features and context association in the above embodiments belong to the same inventive concept. For details not described in detail in the embodiments of the short video sensitive content detection system that combines content features and context association, please refer to the embodiments of the short video sensitive content detection method that combines content features and context association.

[0133] like Figure 2 As shown, the system includes: The feature extraction module is used to extract multimodal content features and basic statistical features of short videos; wherein, the multimodal content features include visual features, audio features and text features; The statistical initial screening module performs an initial screening based on the basic statistical features to obtain an initial screening score; if the initial screening score is lower than a first threshold, it is determined to be normal content and the process ends; otherwise, the content detection module is executed. The content detection module is used to detect the features of the transmitted multimodal content, obtain preliminary sensitive categories, and locate suspicious segments; The context feature acquisition module is used to acquire the context features of the short video; wherein, the context features include user behavior features, propagation network features, and spatiotemporal distribution features; The comprehensive analysis module performs a comprehensive analysis based on the obtained preliminary sensitive category, the suspicious fragment, and the context features obtained in step S104 to obtain a comprehensive anomaly score. The judgment result output module is used to determine whether the short video is sensitive content based on the comparison result of the comprehensive anomaly score and the second threshold, and output the detection result including the preliminary sensitive category and suspicious segments.

[0134] like Figure 3 As shown, this application also provides an electronic device, including a display module 103, a memory 102, a processor 101, a communication module 104, and a computer program stored in the memory and executable on the processor 101. When the processor 101 executes the program, it implements the steps of a short video sensitive content detection method that combines content features and context association.

[0135] In embodiments of the present invention, electronic devices include, but are not limited to, laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. Electronic devices may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the embodiments described and / or claimed herein.

[0136] In this embodiment, processor 101 may be implemented using at least one of an application-specific integrated circuit, a programmable logic device, a field-programmable gate array, a processor, a controller, a microcontroller, a microprocessor, or an electronic unit designed to perform the functions described herein. In some cases, such an implementation may be implemented within a controller. For software implementation, implementations such as processes or functions may be implemented with separate software modules that allow the performance of at least one function or operation. Software code may be implemented by a software application (or program) written in any suitable programming language, and the software code may be stored in memory and executed by the controller.

[0137] The display module 103 is used to display information input by the user or information provided to the user. The display module 103 may include a display panel, which may be configured in the form of a liquid crystal display, an organic light-emitting diode, or the like.

[0138] The memory 102 can be used to store software programs and various data. The memory 102 may include high-speed random access memory, and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0139] The communication module 104 transmits radio signals to and / or receives radio signals from at least one of a base station, an external terminal, and a server. Such radio signals may include voice call signals, video call signals, or various types of data sent and / or received according to text and / or multimedia messages.

[0140] The present invention also provides a storage medium storing a computer program thereon, wherein the computer program, when executed by a processor, implements the steps of the short video sensitive content detection method that combines content features and context association.

[0141] The storage medium may be any combination of one or more readable media. A readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may be, for example,, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples (a non-exhaustive list) of readable storage media include: electrical connections having one or more wires, portable disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0142] The storage medium stores a program product capable of implementing the methods described above in this specification. In some possible implementations, various aspects of this disclosure may also be implemented as a program product comprising program code that, when run on a terminal device, causes the terminal device to perform the steps described in the "Exemplary Methods" section of this specification according to various exemplary embodiments of this disclosure.

[0143] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for detecting sensitive content of short videos in combination with content features and context association, characterized in that, The method comprises: S101, extracting the multi-modal content features and the basic statistical features of the short video; wherein the multi-modal content features comprise visual features, audio features and text features; S102, performing preliminary screening based on the basic statistical features to obtain a preliminary screening score; if the preliminary screening score is lower than a first threshold value, it is determined that the content is normal and the process is ended; otherwise, step S103 is executed; S103, detecting the multi-modal content features passed from step S102 to obtain a preliminary sensitive category and a positioning suspicious segment; S104, obtaining the context features of the short video; wherein the context features comprise user behavior features, propagation network features and spatio-temporal distribution features; S105, comprehensively judging based on the preliminary sensitive category obtained in step S103, the suspicious segment and the context features obtained in step S104 to obtain a comprehensive abnormal score; S106, determining whether the short video is sensitive content according to the comparison result of the comprehensive abnormal score and a second threshold value, and outputting the detection result containing the preliminary sensitive category and the suspicious segment.

2. The method for short video sensitive content detection combined with content features and context association according to claim 1, characterized in that, S101 specifically comprises the following contents: S1011: performing spatio-temporal dimension sampling on the short video frame sequence, constructing a sample set, performing feature extraction operation on the sample set by using a 3D convolutional neural network, and outputting visual features; S1012: separating the short video audio track data, converting the audio track data into an audio data matrix in the form of a mel spectrum graph, performing feature encoding on the audio data matrix by using an audio encoder, and generating audio features; S1013: performing speech-to-text processing on the short video audio track, and also extracting text regions in each frame of the short video picture, fusing to obtain a text information set, inputting the text information set into a pre-trained language model for semantic encoding to obtain text features; S1014: calculating the mean value of picture brightness and the variance of color saturation for each frame of the short video frame sequence, calculating the time sequence change rate of the volume amplitude of the audio track data, performing sensitive word matching statistics on the obtained text information set to obtain the number of text sensitive word hits, reading the time length parameter in the short video metadata, and integrating the calculation results to form the basic statistical features.

3. The method of claim 1, wherein the method further comprises: S102 specifically comprises the following contents: S1021: performing normalization processing on each dimension of the basic statistical features extracted in S101, generating a standardized basic statistical feature set after excluding abnormal data points; S1022: calling historical detection sample data set, counting the sensitive content proportion when each dimension of the standardized basic statistical features exceeds the threshold value, obtaining feature importance weight, and updating weight vector according to the set sample window; S1023: performing weighted summation on the standardized basic statistical feature set and the corresponding importance weight, inputting the result into a sigmoid activation function to complete normalization, and outputting the preliminary screening score; S1024: calculating the current first threshold value based on the false alarm rate and the missed alarm rate change of the historical detection data according to the following formula: wherein is a baseline threshold, is a change in false positive rate relative to the baseline for the most recent time window, is a change in false negative rate, and is a tuning coefficient; Comparing the preliminary screening score with the current first threshold value, if the score is lower, it is determined that the content is normal and the detection process is terminated.

4. The method of claim 1, wherein the method further comprises: S103 specifically comprises the following contents: S1031: Perform self-attention enhancement processing on the visual features, audio features, and text features of S102 respectively, complete cross-modal semantic alignment based on cross-attention mechanism, and then obtain unified multi-modal fusion representation through gating fusion; S1032: Input the multi-modal fusion representation into the classification branch of the multi-task learning framework to output sensitive category probability distribution; Determine each sensitive category through the multi-label branch to generate a binary multi-label set; S1033: Reconstruct the time sequence feature sequence based on the multi-modal fusion representation, input it into the regression branch of the multi-task learning framework to obtain the sensitive degree score; Analyze the weight of each time step through the time sequence attention mechanism to determine the time interval of the suspicious segment; S1034: Integrate the maximum value of the sensitive category probability, the sensitive degree score, the maximum value of the time sequence attention weight, and the maximum value of the multi-label to analyze the comprehensive abnormal score, determine the preliminary sensitive category, and output the specific time sequence information of the suspicious segment.

5. The method of claim 1, wherein, S104 specifically includes the following contents: S1041: Collect the historical behavior data of the short video publishing user, extract the static behavior statistical features and construct the behavior time sequence, encode the time sequence through the Transformer encoder to generate the user behavior representation; S1042: Collect the propagation link data of the short video, construct the propagation network graph, analyze the graph topology features and fit the propagation curve, and obtain the propagation network representation through feature encoding; S1043: Extract the publishing time information and geographic location information of the short video, analyze the time distribution features and spatial aggregation features respectively, and obtain the spatio-temporal distribution representation after performing spatio-temporal joint analysis; S1044: Concatenate the user behavior representation, the propagation network representation, and the spatio-temporal distribution representation, complete fusion encoding through the fully connected layer and layer normalization, and output the context features.

6. The method of claim 1, wherein the method further comprises: S105 specifically includes the following contents: S1051: Organize the preliminary sensitive category and suspicious segment related features obtained from S103, and perform dimension alignment and standardization processing with the context features output from S104 to generate feature pairs suitable for associated modeling; S1052: Perform content-guided context feature enhancement and context-guided content feature re-calibration, and generate associated representation through multi-layer iterative interaction; S1053: Analyze the content risk score, context risk score, and associated risk score based on the content features, context features, and associated representation respectively; S1054: Analyze the weights of the three types of risk scores through the neural network, and calculate the comprehensive abnormal score by weighted fusion according to the following formula to calculate the detection confidence: Formula (7) wherein, is a content risk score, is a contextual risk score, is a correlation risk score, is a content risk weight coefficient, is a contextual risk coefficient, is a correlation risk coefficient.

7. The method of claim 1, wherein the method further comprises: S106 specifically includes the following contents: Based on the change amount of the false positive rate and the false negative rate of the historical preset time length detection data, calculate the current second threshold value based on the following formula: θ2(t) = θ2(t - 1) + α2 · FPR2(t) - β2 · FNR2(t) - γ2 · FNR2(t) is a second threshold reference value, a2 is a false positive rate adjustment coefficient, b2 is a false negative rate adjustment coefficient, FPR2(t) is a false positive rate in a statistical time period, and FNR2(t) is a false negative rate. FPR2(t) = FPR2(t) - FPR2base is the fluctuation in false positive rate relative to baseline, FNR2(t) = FNR2(t) - FNR2base is the fluctuation in false negative rate relative to baseline; Verify the validity of the comprehensive abnormal score output by S105, compare the valid score with the current second threshold value to determine the content type, add artificial review marks combined with the detection confidence; Classify and organize the information according to the determination result, and output the organized information. 8.A short video sensitive content detection system combining content features with context association, characterized in that, The system is used to implement the short video sensitive content detection method combining content features and context association as claimed in any one of claims 1 to 7; The system comprises: The feature extraction module is configured to extract multi-modal content features and basic statistical features of the short video; wherein the multi-modal content features include visual features, audio features and text features; The statistical preliminary screening module is configured to perform preliminary screening based on the basic statistical features to obtain a preliminary screening score; if the preliminary screening score is lower than a first threshold, the short video is determined as normal content and the process ends; otherwise, the content detection module is executed; The content detection module is configured to detect the delivered multi-modal content features to obtain a preliminary sensitive category and a suspicious segment; The context feature acquisition module is configured to acquire context features of the short video; wherein the context features include user behavior features, propagation network features and spatio-temporal distribution features; The comprehensive analysis module is configured to comprehensively analyze the preliminary sensitive category, the suspicious segment and the context features obtained in step S104 to obtain a comprehensive abnormal score; The determination result output module is configured to determine whether the short video is sensitive content according to a comparison result of the comprehensive abnormal score and a second threshold, and output a detection result including the preliminary sensitive category and the suspicious segment.

9. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the program to implement the steps of the short video sensitive content detection method combining content features with context according to any one of claims 1 to 7.

10. A storage medium having stored thereon a computer program, characterized in that The computer program is executed by the processor to implement the steps of the short video sensitive content detection method combining content features with context according to any one of claims 1 to 7.

Images