Video generation content credibility detection method and system

Abstract

The invention relates to the technical field of multi-modal content security detection, in particular to a video generation content credibility detection method and system. The method comprises the following steps: dividing an input video into non-overlapping segments, independently modeling through three paths of spatial appearance, motion and semantics, and constructing a unified dynamic representation tensor after alignment and fusion; a cross-fragment bidirectional residual evolution mechanism is introduced to analyze time sequence splitting abnormity, Gaussian perturbation is injected to quantify disturbance response, and fragment vulnerability is evaluated in combination with a residual stability index; quantizing a fragile region by combining confrontation disturbance and designing a multi-scale consistency kernel mechanism to generate a credible score, and judging the authenticity of the video by combining an adaptive threshold value; secondary disturbance is adopted to enhance confidence for fuzzy samples, continuous self-learning optimization is realized through pseudo tag guidance, counterfeit style clustering and model fine tuning, and generalization ability for new counterfeit samples is improved. According to the invention, through multi-dimensional dynamic modeling and a self-learning mechanism, the accuracy and robustness of video credible detection are significantly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of multimodal content security detection, and in particular to a method and system for detecting the credibility of video-generated content. Background Art

[0002] With the rapid development of deep fake technology, video fake methods based on generative adversarial networks and diffusion models have continued to make breakthroughs in realism and generation efficiency, resulting in the large-scale dissemination of fake videos on social media and public information platforms, posing a serious threat to the credibility of digital content, the social trust system and national security.

[0003] A Chinese invention patent with announcement number CN112016500B discloses a method and system for identifying abnormal group behavior based on multi-scale temporal information fusion. The scheme includes obtaining a video frame sequence containing only normal samples and extracting an optical flow map between two adjacent frames, sampling the same video clip at different sampling rates to obtain video sequences at different time scales, obtaining a skipped frame video frame sequence of the aforementioned video frame sequence and extracting the skipped frame optical flow map, thereby obtaining video frame sequences and optical flow map sequences at different time scales; using the normal frame, skipped frame video sequence, and optical flow map sequence as input to train a deep convolutional network; using the trained deep convolutional network to detect abnormal behavior, the detection network evaluates the obtained predicted image to obtain an abnormality score, and determines whether the video frame is an abnormal frame based on the score, thereby completing the abnormality detection of the video frame.

[0004] Existing technologies mostly focus on single-dimensional feature analysis (such as spatial artifacts or motion trajectories), and have insufficient complexity when dealing with multi-modal joint forgery attacks (such as local tampering and time-series frame skipping). Further research is needed on continuous adaptive mechanisms in open environments to improve long-term generalization capabilities. Summary of the Invention

[0005] The purpose of the present invention is to address the problems existing in the background technology and propose a method and system for detecting the credibility of video generated content.

[0006] The technical solution of the present invention is a method for detecting the credibility of video generated content, which includes the following specific implementation steps: S1. Divide the input sample into non-overlapping segments and construct three parallel modeling pathways. The spatial appearance pathway uses a convolutional encoder to extract static features, the motion pathway uses an optical flow encoder to extract dynamic features, and the semantic pathway uses a semantic encoder to extract semantic features. After aligning and fusing the features of the three pathways, they are normalized and positionally encoded to construct a unified dynamic representation sequence. S2. Extract temporal evolution differences through cross-segment bidirectional residual evolution modeling, inject Gaussian perturbations to quantify the stability of spatial and motion features, and calculate the residual stability index based on the residual difference before and after perturbation to evaluate the vulnerable areas of the video; S3. By setting a multi-scale time window, the average perturbation residual in the sliding window is calculated. A kernel function based on Euclidean distance is defined for each scale to quantify the similarity between windows and construct a kernel matrix. The kernel symmetry deviation index is extracted and then weightedly fused with the residual stability index to generate a consistent energy field. S4. Construct a credibility tuple by summarizing the multi-scale consistency energy field, disturbance response, and residual stability index and normalize it. Generate a credibility score based on the dynamic weighted fusion of the index variance. Adaptively learn the optimal judgment threshold to distinguish the authenticity of the video. Perform secondary perturbation analysis and fusion scoring on the boundary fuzzy samples. S5. Generate high-confidence pseudo labels based on the initial score and the secondary perturbation score by screening fuzzy samples near the threshold, cluster the perturbation response and consistency energy field features to identify new forgery patterns and construct representation centers, and fine-tune the model parameters by combining binary cross entropy loss and KL divergence distribution alignment loss.

[0007] Preferably, the three parallel modeling paths include: Path A, spatial appearance path: Apply a lightweight convolutional encoder MobileNet to each video clip to extract spatial appearance static visual features frame by frame and stack them into d A dimensional feature sequence tensor ; Path B, motion path: calculate the inter-frame optical flow sequence for each video clip, use the PWCNet optical flow estimator to obtain the inter-frame optical flow field, and use the motion encoder to perform convolution pooling to generate a feature dimension of d M Motion feature sequence ; Path C, semantic evolution path: Apply the pre-trained visual language model to each video frame to extract the semantic representation vector, stack it into a semantic sequence, and then use Bi-GRU and self-attention mechanism to model the temporal evolution to generate a dynamic representation vector .

[0008] Preferably, the residual difference before and after disturbance is calculated as follows: The compressed representation is obtained by performing temporal average pooling on the dynamic features of the fragments, and the forward residual and reverse residual vectors of adjacent fragments are calculated and concatenated into a residual tensor; Gaussian perturbations are injected into a subset of the spatial and motion features of each fragment to generate perturbed features, which are then concatenated with the unperturbed semantic features to form a new representation. Residual modeling is then repeated to calculate the residual tensor after perturbation. The difference in residuals before and after perturbation is measured based on the Euclidean distance to obtain a perturbation response metric for each fragment pair.

[0009] Preferably, the residual stability index is: ; Among them, RSI represents the residual stability index; Represents a constant to prevent division by zero; N represents the number of time segments, satisfying: ; represents the floor rounding function; T represents the number of video frames; is the length of any video segment; Represents the residual tensor of the forward and reverse residuals concatenated to form a segment pair; It represents the calculation of the Euclidean distance of the residual tensor before and after the perturbation in the i-th segment pair, that is, the perturbation response metric.

[0010] Preferably, the kernel matrix construction process is: Set a finite set of time scales: ; Among them, S is a multi-scale time window set, which contains L sliding window scales of different lengths; Indicates the length of the lth time window in the scale set; For each scale , apply a sliding window across the entire residual sequence and compute the average perturbed residual within each window: ; in, Representation scale Next, the average tensor of the perturbation residuals in the sliding window starting from the jth time position; For any two scale windows and , define the consistency kernel: ; in, Representation scale Next, the kernel similarity value of the perturbation residuals of the j-th and k-th time windows; represents the natural exponential function; Representation scale The bandwidth parameter of the kernel function; For each time scale, an exponential kernel function based on Euclidean distance is defined to calculate the similarity value of the perturbation residual between windows of the same scale, and a kernel matrix is ​​constructed: ; in, Representation scale The perturbation residual kernel similarity matrix under , the rows and columns correspond to different time window indices.

[0011] Preferably, the consistent energy field generation process is: Structural symmetry deviation index: ; in, Representation scale The degree of deviation of nuclear symmetry under ; The KSD at several scales is weighted and fused with the residual stability index RSI to obtain a comprehensive consistency score: ; in, Represents the fusion weight of multi-scale consistency energy field, satisfying ; CEF represents the consistency energy field, which is the final comprehensive score of video credibility.

[0012] Preferably, the process of adaptively learning the optimal judgment threshold to distinguish the authenticity of the video is as follows: By summarizing the multi-scale consistency energy field CEF and disturbance response metrics of each video clip The local credibility tuple and global indicator set are constructed with the residual stability index RSI, and nonlinear normalization is performed based on the statistical mean of the training set: ; in, represents the original global value of the j-th indicator, j=1 is CEF, j=2 is D, j=3 is RSI; represents the statistical mean of the j-th indicator in the training set; represents the scale adjustment coefficient; represents the normalized j-th index value; Define the final credibility score S: ; Automatically generate weights based on indicator stability : ; Among them, S represents the final credibility fusion score; Represents the normalized index The fusion weight of the final score S; represents the variance of the jth normalized indicator at the segment level; By defining the expected misjudgment risk function that integrates the false positive rate and the false negative rate, the optimal decision boundary threshold that minimizes the risk loss is statistically learned on a dynamic sliding window, and the video authenticity classification decision is realized based on this threshold.

[0013] Preferably, the feedback execution process of performing secondary perturbation analysis fusion scoring on boundary fuzzy samples is: when the absolute gap between the final credible score and the adaptive judgment threshold is less than the set value, the feedback process is triggered, the perturbation response and multi-scale consistency analysis are re-executed to generate a secondary score, and the enhanced result is generated by weighted fusion of the first score and the secondary score, and the historical indicators are updated to optimize the threshold learning.

[0014] Preferably, the implementation process of fine-tuning model parameters by combining binary cross entropy loss and KL divergence distribution alignment loss is: Take the video to be processed V and obtain its final credibility score S. If it satisfies , the video is considered to be a potential high-value blurred sample, and the following fusion pseudo-label generation function is used: ; in, Indicates the enhanced tolerance for pseudo-label determination; Represents the automatically generated pseudo-label value. If it is confirmed to be forged, it is 1, if it is true, it is 0, and if it is uncertain, it is uncertain and does not participate in training; Construct a perturbation response memory cache unit: , using similarity clustering algorithm to analyze disturbance response and consistency Perform cluster analysis to identify new forged style groups with significant structural differences and construct the local representation center vector C for each type of forged style. k ; Among them, M is the perturbation response memory cache unit, which contains the residual tensor, perturbation response, consistency kernel feature and pseudo label of all automatically labeled samples; C k is the cluster center vector of the k-th forged style in the perturbation response space; Select samples with high pseudo-label reliability ( ), enter the fine-tuning set D fine , and lightly fine-tuned using the following loss function: ; in, represents the overall loss function of fine-tuning training; Represents the weighted coefficient of the KL divergence loss term; Indicates that the current model is in the input segment F i The probability distribution of the predicted output; represents the binary cross entropy loss; Indicates the current detection model for the video segment F i The output fake credibility score; represents the KL divergence.

[0015] The technical solution of the present invention is a video-generated content credibility detection system, which is used to execute the above-mentioned video-generated content credibility detection method, including: Memory; processor; a computer program stored in the memory and executable on the processor; The processor executes the computer program to implement the above-mentioned method for detecting the credibility of video-generated content.

[0016] Compared with the prior art, the above technical solution of the present invention has the following beneficial technical effects: The present invention designs a method and system for credible detection of video-generated content. The overall technical solution has made substantial breakthroughs in detection accuracy, anti-interference ability and continuous evolution. Specifically: First, based on the independent modeling and unified fusion of spatial appearance, motion and semantics, it comprehensively captures static texture anomalies, dynamic trajectory fragmentation and semantic evolution contradictions in deep fake videos, effectively solving the limitations of single-dimensional detection; secondly, the cross-segment bidirectional residual evolution mechanism is combined with the quantification of anti-disturbance sensitivity to accurately identify the vulnerability of temporal fragmentation and forged areas, and significantly enhances the ability to identify covert forgeries such as frame skipping replacement and local tampering; thirdly, the multi-scale consistency fluctuation kernel mechanism effectively resists complex interference and improves the generalization of the model by quantifying the structural stability of the video at multiple time scales; finally, the pseudo-label-guided self-learning mechanism dynamically updates the forged style memory library and optimizes the decision boundary, continuously adapts to new forgery techniques, and ensures the long-term effectiveness of the system in an open environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 This is a flow chart of a method for detecting the credibility of video generated content proposed by the present invention. DETAILED DESCRIPTION

[0018] Example 1, as Figure 1 As shown, the present invention proposes a method for detecting the credibility of video generated content, including the following specific implementation steps: S1. Decompose, model, and uniformly represent the input video from multi-dimensional information paths, construct a dynamic representation tensor that can be used for subsequent forgery detection, and achieve a structured expression of video semantic behavior through temporal segment segmentation, three-path dynamic modeling, and multi-dimensional feature alignment and fusion. The specific implementation process is as follows: S11. Given an input video , divide it into N non-overlapping segments, each of length , get the fragment set: ; Where V represents the original input video sequence, which contains T frames, each frame is an RGB image with a size of H×M; T represents the number of video frames; H and W represent the height and width of the video frame respectively; N represents the number of time segments, satisfying: ; represents the floor function; represents the i-th time segment, which is composed of Δt frames taken continuously from V; S12. To capture the dynamic evolution of video content in different dimensions, three parallel modeling pathways are constructed: the spatial appearance pathway, the motion pathway, and the semantic pathway. Each pathway independently extracts temporal features in its own dimension and aligns them on the time axis. Specifically: (1) Pathway A, the spatial appearance path, models the static changes of video frames at the visual pixel level and is suitable for detecting static camouflage features such as local texture fusion, edge forgery, and facial masking. Specifically: For each paragraph Input lightweight convolutional encoder , extract features from each frame and stack them into tensors: ; in, Represents the sequence of spatial appearance features extracted from the i-th segment, one d per frame A dimensional vector; represents the spatial feature vector of the jth frame in the i-th segment; represents the spatial encoder function, which extracts the static visual features of each frame. In this embodiment, the lightweight convolutional network MobileNet is used; (2) Path B, the motion path, captures the inter-frame motion trend in the video and is used to identify dynamic forgeries such as unreasonable motion, regional drift, and frame skipping. Specifically: Calculate the inter-frame optical flow sequence for each video clip: ; ; in, Represents the optical flow sequence of all frames in the i-th segment, with a length of ; represents the optical flow field between the jth frame and the j+1th frame in the i-th segment, and the two-dimensional vector describes the direction and amplitude of pixel motion; represents the optical flow estimator (PWCNet is used in this embodiment), which extracts the two-dimensional motion vector field between frames; Represents two frames of image and Calculate optical flow between them; Using motion encoders Convolutional pooling encoding of optical flow sequence: ; in, Represents the optical flow encoder, which performs convolution pooling on the optical flow sequence and outputs fixed-length motion features; represents the motion feature sequence extracted from the i-th segment; d M Represents the characteristic dimension of the motor pathway; (3) Path C, the semantic evolution path, models the continuity and evolutionary stability of frame-level semantic concepts, specifically: Apply a visual language model (such as CLIP, BLIP) to each frame to obtain a semantic vector: ; Get the frame-level semantic sequence: ; Input Time Modeler (This embodiment uses Bi-GRU+Self-Attention) to dynamically model the semantic sequence and generate an evolution vector: ; in, The semantic encoder is set as the visual encoding part of the pre-trained graphic model (such as CLIP and BLIP) in this embodiment to extract the semantic representation vector of each frame; represents the semantic vector of the jth frame in the i-th segment; Represents the semantic sequence of the i-th segment (the semantic vectors of each frame are stacked in time); represents the semantic dynamic representation of the i-th segment; represents a semantic modeler that captures the semantic temporal evolution trend; d S Represents the semantic vector dimension of each frame; It should be noted that CLIP (Contrastive Language-Image Pretraining) aligns image and text features through contrastive learning, uses 400 million network image-text pairs for training, supports zero-shot image classification (such as matching images to arbitrary text categories), and is widely used in image-text retrieval, content review, and multimodal research; BLIP (Bootstrapping Language-Image Pretraining) focuses on generation and understanding, combines Captioner (generating descriptions), Filter (cleaning noise data), and Retriever (image-text matching), and uses model bootstrapping technology to optimize the quality of training data. It has outstanding performance in image description generation (such as generating natural language titles for images) and visual question answering (VQA) tasks. S13. Perform unified alignment and normalization fusion to construct a unified representation tensor for subsequent analysis. Specifically: Align the three-channel output to the same number of frames in the time dimension and unify the dimensions through channel transformation: ; ; ; in, Represents the feature sequence after spatial pathway alignment and compression, and all segments are normalized to the same frame number and dimensions ; represents the motor pathway alignment output; represents the semantic pathway alignment output; Indicates the interpolation / sampling of sequences of different pathways in the time dimension to make them of uniform length; Fusion builds the final tensor : ; in, Represents the final fusion feature of the i-th segment, including three types of representations: space, motion, and semantics; Represents tensor concatenation operations (along the channel dimension); Indicates the uniform frame number after alignment; represents the unified dimension to which each pathway is compressed; For the final tensor Channel normalization (LayerNorm) is performed and learnable position encoding is added to improve the model's ability to perceive temporal position. Ultimately, all segments form a sequence representing the overall dynamic representation of the video: ; Among them, F represents the dynamic representation sequence of the entire video, which is used by the subsequent detection model.

[0019] S2: By building a cross-segment bidirectional residual evolution mechanism, introducing adversarial perturbation sensitivity assessment, and defining credibility indicators, we achieve mid- and high-level structural behavioral analysis of the credibility of video generated content, thereby identifying potential forgery signals in deepfakes, such as evolutionary anomalies, motion trajectory fragmentation, and perturbation vulnerability. The specific implementation process is as follows: S21. Construct a bidirectional residual evolution modeling mechanism across segments and introduce reverse residual modeling. Observe temporal fragmentation through the symmetry of forward and reverse evolution, extract temporal evolution differences between segments, and perform symmetry analysis on the forward and reverse residuals to capture abnormal temporal fragmentation. Specifically: For each dynamic feature segment , first perform average pooling on the time dimension to obtain the compressed representation of the fragment: ; in, Representation fragment Average feature representation over the time dimension; Represents the fusion feature vector of the i-th segment at time step t, with dimension ; For adjacent segments and , construct the bidirectional residual vector: Forward residual vector (representing the natural evolution trend): ; Inverse residual vector (indicating the rationality of the inversion): ; Merge the residual information representing the i-th segment pair: ; in, Represents the forward residual vector of the i-th pair of adjacent segments, describing the natural evolution of the video from i to i+1; represents the inverse residual vector of the i-th pair of segments, which is used to judge the rationality of the inversion; Represents the residual tensor of the forward and reverse residuals concatenated to form a segment pair; S22: Perform local perturbation injection and residual response mapping. The perturbation injection is used to test the degree of change in the response of the segment features to the perturbation, thereby quantifying the stability of the segment and determining whether it is a possible forgery region. Specifically: For each fusion segment feature , inject Gaussian distribution perturbations (only the first two paths, space and motion): , construct the perturbation feature, that is, construct the perturbation vector: ; in, Represented as a fragment Generated Gaussian perturbation noise; express The identity matrix of , controls the covariance structure to be diagonal; It represents the hyperparameter that controls the intensity of the disturbance and is set manually; Represents the fused representation fragment after generating the perturbation; Represents the fusion feature subset of the spatial appearance + motion path of each frame of the i-th segment in time, composed of the following sources: It is a static image content, It is the inter-frame optical flow or dynamic motion information; represents the subset of semantic pathway features for each frame in segment i, without perturbation; The concatenation operation of features along the channel dimension (i.e., concatenating two tensors along the last dimension); For all fragments after perturbation , repeat step S21 to calculate the corresponding perturbation residual tensor: ; in, Represents remodeling of the residuals of the perturbed features to obtain the perturbed residuals, i.e., the i-th residual pair after perturbation; represents the forward residual vector of the i-th pair of adjacent segments after perturbation; represents the reverse residual vector of the i-th pair of adjacent segments after perturbation; Calculate the difference between the residual vectors before and after the perturbation to obtain the perturbation response measure for each segment pair: ; in, It represents the calculation of the Euclidean distance of the residual tensor before and after the perturbation in the i-th segment pair, that is, the perturbation response metric; represents the L2 norm, i.e., the Euclidean distance, which measures the residual change caused by the disturbance; S23. Calculate the Residual Stability Index (RSI) and determine its credibility. Convert the disturbance response into an interpretable numerical indicator as the basis for determining the credibility of the video. Specifically: ; Among them, RSI represents the residual stability index, which ranges from [0,1]; Represents a constant that prevents division by zero, enhancing numerical stability.

[0020] S3. A dynamic consistency fluctuation kernel mechanism is proposed to jointly model the disturbance response, residual evolution, and inter-scale variation structure to quantify the consistency characteristics of videos at multiple time scales. The specific implementation process is as follows: S31. Construct multi-scale time segments and generate residual fusion views, and reconstruct the perturbed residual tensor according to different time scales. Specifically: Assume a finite set of time scales: ; Among them, S is a multi-scale time window set, which contains L sliding window scales of different lengths; Indicates the length of the lth time window in the scale set; For each scale , apply a sliding window across the entire residual sequence and compute the average perturbed residual within each window: ; in, Representation scale Next, the average tensor of the perturbation residual in the sliding window starting from the jth time position, ; S32. After obtaining the perturbation residual sequences at different scales, a set of residual consistency kernel functions is defined for each scale, and the law of local structural changes of the video is captured in a nonlinear manner. Specifically: For any two scale windows and , define their consistency core as: ; Among them, j and k represent the j-th and k-th time window positions respectively; Representation scale Next, the kernel similarity value of the perturbation residuals of the j-th and k-th time windows; represents the natural exponential function; Representation scale The bandwidth parameter (standard deviation) of the kernel function, ; For all windows<j,k> , construct a complete scaling kernel matrix: ; in, Representation scale The perturbation residual kernel similarity matrix under , where the rows and columns correspond to different time window indices; S33. Construct two types of quantitative mechanisms to extract structured indicators from the kernel matrix, specifically: Structural symmetry deviation index: ; in, Representation scale Kernel Symmetry Deviation under ; The KSD at multiple scales is weighted and integrated with the residual stability index RSI to obtain a comprehensive consistency score: ; in, Represents the multi-scale consistency energy field fusion weight, which is used to weight the contribution of each scale to meet CEF stands for Consistency Energy Field, which is the final comprehensive score of video credibility.

[0021] S4. Build a multi-dimensional fusion credibility scoring mechanism, combining the multi-scale consistency energy field (CEF) with the disturbance response metric and residual stability index. By designing a dynamic adaptive fusion model and a credible boundary inference algorithm, we can accurately determine whether a video is forged. The specific implementation process is as follows: S41, using the consistency energy field CEF of step S3 as the core trust indicator, and collecting the disturbance response measurement of step S2 at the same time Together with the residual stability index RSI, we summarize the three-dimensional indicators of each video segment i and construct its local credibility tuple , for the entire video, build a global trustworthy indicator set I: ; In order to eliminate dimensional differences, all indicators are uniformly normalized nonlinearly: ; in, Represents the credibility tuple of the i-th video clip, which contains the multi-dimensional evaluation indicators of the clip; It represents the multi-scale consistency energy field index, reflecting the degree of structural alignment of the segment in the multi-scale residual analysis; represents the perturbation response metric, which quantifies the degree of change in the dynamic performance of the segment after the adversarial perturbation injection; represents the residual stability index, which measures the stability of the perturbation residual distribution in the time series; I represents the set of global trust indicators of the entire video, which respectively represent the global levels of multi-scale consistency, perturbation responsiveness and stability; It represents the mathematical expectation of the indicators of all segments i, that is, the average value of the video level; represents the raw global value of the j-th indicator (j=1 is CEF, j=2 is D, j=3 is RSI); represents the statistical mean of the j-th indicator in the training set; Represents the scale adjustment coefficient, which is used to control the sensitivity of the normalization function; represents the normalized j-th indicator value, which serves as the basis for the fusion score; S42. Design a fusion model based on dynamic weight adjustment of multi-dimensional indicator space to generate the final credibility score S, specifically: Define the final credibility score S: ; Automatically generate weights based on indicator stability : ; Among them, S represents the final credibility fusion score, which ranges from (0, 1). The larger the value, the more likely it is a real video. Represents the normalized index The fusion weight of the final score S depends on its cross-segment variance, that is, its credibility in video judgment; It represents the variance of the jth normalized index at the segment level, reflecting its volatility in the temporal dimension. The smaller it is, the more stable it is, and the more important this index is in the judgment; S43. Construct an adaptive credible boundary inference mechanism to dynamically adjust the credible decision threshold T, and introduce a confidence region decision function to determine whether the sample falls into the boundary fuzzy band, thereby achieving a more reliable recognition strategy. Specifically: Define the expected misjudgment risk function , combining the false positive rate (FPR) and the false negative rate (FNR), derive the optimal credible decision boundary T: ; By performing statistical learning on a dynamic sliding window: ; Use the calculated threshold , make credibility judgment on the test sample: ; in, represents the false positive rate, that is, the proportion of real videos that are mistakenly identified as forged videos; represents the false negative rate, that is, the proportion of forged videos that are mistakenly identified as real videos; represents the risk loss function; represents the risk balance coefficient, which controls the relative importance of FPR and FNR in the loss function; T represents the decision boundary for dividing the video into real and fake; The optimal value of the decision boundary for classifying videos into real and fake, learned by the adaptive algorithm; S44. Add confidence assessment and self-feedback update mechanism to handle uncertain samples near the decision boundary. Specifically: If the final score is too close to the decision boundary, that is: , it is considered that the judgment confidence is insufficient, that is, the model has no confidence in the judgment of this sample, and then enters the feedback optimization process; in, represents the confidence bound tolerance constant; Perform secondary perturbation on the video and the process of steps S2 and S3 to generate a second round of perturbation response and consistency analysis, and re-weight the fusion based on the first score: , and update historical statistical indicators to optimize future threshold learning; in, Represents the scoring result after confidence enhancement, integrating the first judgment score S and the secondary perturbation analysis score get; represents the credibility score calculated after the secondary perturbation of the fuzzy boundary sample in the feedback process; Based on this: the ability to distinguish uncertain samples is enhanced through dynamic perturbation contrast, simulating the "secondary judgment of the human eye" process, greatly improving the model's performance on fuzzy boundaries.

[0022] S5. Construct a self-learning mechanism that combines pseudo-label enhancement guidance, dynamic pseudo-memory module update, and positive and negative sample boundary correction feedback to achieve continuous adaptation to new pseudo-samples. The specific implementation process is as follows: S51. Discover and use video samples that are not yet clear to the model (with scores close to the threshold) as potential learning samples, and automatically generate pseudo labels for soft training. Specifically: Take the video to be processed V and obtain its final credibility score S; If satisfied , then the video is considered to be a potential high-value fuzzy sample; Use the following fused pseudo-label generation function: ; in, Indicates the enhanced tolerance for pseudo-label determination, used in the initial score S and the secondary score On the basis of , a stricter credible range is set to improve the reliability of pseudo labels; Indicates the automatically generated pseudo-label value. If it is confirmed to be forged, it is 1, if it is true, it is 0, and if it is uncertain, it is "uncertain" and does not participate in training; S52. Identify the disturbance behavior patterns of unknown types of forged videos and dynamically enhance the detection system, specifically: Construct a perturbation response memory cache unit: ; Use similarity clustering algorithm (DBSCAN is used in this example) to analyze disturbance response and consistency Conduct cluster analysis to identify new groups of forgery styles with significant structural differences; Construct the local representation center vector C of each type of forged style k , used to update the attention mechanism of the discriminator or to forge distribution priors; Among them, M is the perturbation response memory cache unit, which is a set containing the residual tensor, perturbation response, consistency kernel feature and pseudo label of all automatically labeled samples; C k is the cluster center vector of the k-th forgery style in the perturbation response space, which is used to model new forgery types; S53. Use the credible boundary samples (automatically generated pseudo labels) and the identified new forgery style features to fine-tune the original discriminant model to maintain robustness and generalization in new scenarios. Specifically: Select samples with high pseudo-label reliability ( ), enter the fine-tuning set D fine ; Light fine-tuning is performed using the following loss function: ; in, Represents the overall loss function of fine-tuning training, which consists of two parts: cross entropy loss + distribution alignment loss; Represents the weighting coefficient of the KL divergence loss term, which is used to adjust the weight balance between cross entropy loss and forged behavior alignment; Indicates that the current model is in the input segment F i The probability distribution of the predicted output is used to compare the matching degree between the true pseudo label and the model output; represents the binary cross entropy loss, which is used to measure the degree to which the detector correctly determines the pseudo label; Indicates the current detection model for the video segment F i The output fake credibility score; represents the KL divergence.

[0023] In a second embodiment, the present invention provides a system for detecting the credibility of video-generated content, which is used to execute the method for detecting the credibility of video-generated content proposed in the first embodiment, and includes: Memory; processor; a computer program stored in the memory and executable on the processor; The processor executes the computer program to implement the method for detecting credibility of video-generated content in the first embodiment.

[0024] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited thereto. Various changes can be made within the scope of knowledge possessed by those skilled in the art without departing from the spirit of the present invention.

Claims

1. A method for detecting the credibility of video generated content, characterized in that: The specific implementation steps include the following: S1. Divide the input sample into non-overlapping segments and construct three parallel modeling pathways. The spatial appearance pathway uses a convolutional encoder to extract static features, the motion pathway uses an optical flow encoder to extract dynamic features, and the semantic pathway uses a semantic encoder to extract semantic features. After aligning and fusing the features of the three pathways, they are normalized and positionally encoded to construct a unified dynamic representation sequence. S2. Extract temporal evolution differences through cross-segment bidirectional residual evolution modeling, inject Gaussian perturbations to quantify the stability of spatial and motion features, and calculate the residual stability index based on the residual difference before and after perturbation to evaluate the vulnerable areas of the video; S3. By setting a multi-scale time window, the average perturbation residual in the sliding window is calculated. A kernel function based on Euclidean distance is defined for each scale to quantify the similarity between windows and construct a kernel matrix. The kernel symmetry deviation index is extracted and then weightedly fused with the residual stability index to generate a consistent energy field. S4. Construct a credibility tuple by summarizing the multi-scale consistency energy field, disturbance response, and residual stability index and normalize it. Generate a credibility score based on the dynamic weighted fusion of the index variance. Adaptively learn the optimal judgment threshold to distinguish the authenticity of the video. Perform secondary perturbation analysis and fusion scoring on the boundary fuzzy samples. S5. Generate high-confidence pseudo labels based on the initial score and the secondary perturbation score by screening fuzzy samples near the threshold, cluster the perturbation response and consistency energy field features to identify new forgery patterns and construct representation centers, and fine-tune the model parameters by combining binary cross entropy loss and KL divergence distribution alignment loss.

2. A method for detecting credibility of video generated content according to claim 1, characterized in that: The three parallel modeling paths include: Path A, spatial appearance path: Apply a lightweight convolutional encoder MobileNet to each video clip to extract spatial appearance static visual features frame by frame and stack them into d A dimensional feature sequence tensor ; Path B, motion path: calculate the inter-frame optical flow sequence for each video clip, use the PWCNet optical flow estimator to obtain the inter-frame optical flow field, and use the motion encoder to perform convolution pooling to generate a feature dimension of d M Motion feature sequence ; Path C, semantic evolution path: Apply the pre-trained visual language model to each video frame to extract the semantic representation vector, stack it into a semantic sequence, and then use Bi-GRU and self-attention mechanism to model the temporal evolution to generate a dynamic representation vector .

3. The method for detecting credibility of video generated content according to claim 2, wherein: The calculation process of the residual difference before and after disturbance is: The compressed representation is obtained by performing temporal average pooling on the dynamic features of the fragments, and the forward residual and reverse residual vectors of adjacent fragments are calculated and concatenated into a residual tensor; Gaussian perturbations are injected into a subset of the spatial and motion features of each fragment to generate perturbed features, which are then concatenated with the unperturbed semantic features to form a new representation. Residual modeling is then repeated to calculate the residual tensor after perturbation. The difference in residuals before and after perturbation is measured based on the Euclidean distance to obtain a perturbation response metric for each fragment pair.

4. The method for detecting credibility of video generated content according to claim 3, wherein: The residual stability index is: ; Among them, RSI represents the residual stability index; Represents a constant to prevent division by zero; N represents the number of time segments, satisfying: ; represents the floor rounding function; T represents the number of video frames; is the length of any video segment; Represents the residual tensor of the forward and reverse residuals concatenated to form a segment pair; It represents the calculation of the Euclidean distance of the residual tensor before and after the perturbation in the i-th segment pair, that is, the perturbation response metric.

5. The method for detecting credibility of video generated content according to claim 4, characterized in that: The kernel matrix construction process is: Set a finite set of time scales: ; Among them, S is a multi-scale time window set, which contains L sliding window scales of different lengths; Indicates the length of the lth time window in the scale set; For each scale , apply a sliding window across the entire residual sequence and compute the average perturbed residual within each window: ; in, Representation scale Next, the average tensor of the perturbation residuals in the sliding window starting from the jth time position; For any two scale windows and , define the consistency kernel: ； in, Representation scale Next, the kernel similarity value of the perturbation residuals of the j-th and k-th time windows; represents the natural exponential function; Representation scale The bandwidth parameter of the kernel function; For each time scale, an exponential kernel function based on Euclidean distance is defined to calculate the similarity value of the perturbation residual between windows of the same scale, and a kernel matrix is ​​constructed: ; in, Representation scale The perturbation residual kernel similarity matrix under , the rows and columns correspond to different time window indices.

6. A method for detecting credibility of video generated content according to claim 5, characterized in that: The process of generating a consistent energy field is as follows: Structural symmetry deviation index: ; in, Representation scale The degree of deviation of nuclear symmetry under ; The KSD at several scales is weighted and fused with the residual stability index RSI to obtain a comprehensive consistency score: ; in, Represents the fusion weight of multi-scale consistency energy field, satisfying ; CEF represents the consistency energy field, which is the final comprehensive score of video credibility.

7. The method for detecting credibility of video generated content according to claim 6, characterized in that: The judgment process of adaptively learning the optimal judgment threshold to distinguish the authenticity of videos is as follows: By summarizing the multi-scale consistency energy field CEF and disturbance response metrics of each video clip The local credibility tuple and global indicator set are constructed with the residual stability index RSI, and nonlinear normalization is performed based on the statistical mean of the training set: ; in, represents the original global value of the j-th indicator, j=1 is CEF, j=2 is D, j=3 is RSI; represents the statistical mean of the j-th indicator in the training set; represents the scale adjustment coefficient; represents the normalized j-th index value; Define the final credibility score S: ; Automatically generate weights based on indicator stability : ; Among them, S represents the final credibility fusion score; Represents the normalized index The fusion weight of the final score S; represents the variance of the jth normalized indicator at the segment level; By defining the expected misjudgment risk function that integrates the false positive rate and the false negative rate, the optimal decision boundary threshold that minimizes the risk loss is statistically learned on a dynamic sliding window, and the video authenticity classification decision is realized based on this threshold.

8. The method for detecting credibility of video generated content according to claim 7, wherein: The feedback execution process of performing secondary perturbation analysis and fusion scoring on boundary fuzzy samples is as follows: when the absolute gap between the final credible score and the adaptive judgment threshold is less than the set value, the feedback process is triggered, the perturbation response and multi-scale consistency analysis are re-executed to generate a secondary score, and the enhanced result is generated by weighted fusion of the first score and the secondary score, and the historical indicators are updated to optimize the threshold learning.

9. The method for detecting credibility of video generated content according to claim 1, wherein: The implementation process of fine-tuning model parameters by combining binary cross entropy loss and KL divergence distribution alignment loss is as follows: Take the video to be processed V and obtain its final credibility score S. If it satisfies , the video is considered to be a potential high-value blurred sample, and the following fusion pseudo-label generation function is used: ； in, Indicates the enhanced tolerance for pseudo-label determination; Represents the automatically generated pseudo-label value. If it is confirmed to be forged, it is 1, if it is true, it is 0, and if it is uncertain, it is uncertain and does not participate in training; Construct a perturbation response memory cache unit: , using similarity clustering algorithm to analyze disturbance response and consistency Perform cluster analysis to identify new forged style groups with significant structural differences and construct the local representation center vector C for each type of forged style. k ; Among them, M is the perturbation response memory cache unit, which contains the residual tensor, perturbation response, consistency kernel feature and pseudo label of all automatically labeled samples; C k is the cluster center vector of the k-th forged style in the perturbation response space; Select samples with high pseudo-label reliability ( ), enter the fine-tuning set D fine , and lightly fine-tuned using the following loss function: ； in, represents the overall loss function of fine-tuning training; Represents the weighted coefficient of the KL divergence loss term; Indicates that the current model is in the input segment F i The probability distribution of the predicted output; represents the binary cross entropy loss; Indicates the current detection model for the video segment F i The output fake credibility score; represents the KL divergence.

10. A video-generated content credibility detection system, configured to execute a video-generated content credibility detection method according to any one of claims 1 to 9, characterized in that: include: Memory; processor; a computer program stored in the memory and executable on the processor; The processor executes the computer program to implement the method for detecting the credibility of video-generated content as described in any one of claims 1 to 9.

Images