An artificial intelligence generated video authentication method based on multi-view consistency

Abstract

The application discloses an artificial intelligence generated video authentication method based on multi-view consistency. The method comprises the following steps: constructing a large-scale real-world simulation video dataset, using the previous frame as a prompt to generate a future frame video with high consistency in semantics, color and physical laws, and improving data authenticity; designing a detection model based on multi-view consistency physical prior, extracting video features through real-time multi-view matching, combining a time sequence memory module to dynamically store and update inter-frame stereo reconstruction features, and using an attention mechanism to enhance long-time consistency analysis; using a multi-modal large model to generate unified prompts and label data, and improving the model generalization ability; scoring each frame feature through a scorer and globally averaging to output a video authenticity judgment result. The application uses multi-view consistency prior and memory enhancement mechanism to effectively capture the subtle defects in AI generated videos that violate the real physical laws, and significantly improves the detection robustness and accuracy of high-quality generated videos.

Description

Technical Field

[0001] This invention relates to machine learning technology, specifically to AI-generated video authentication technology, specifically an AI-generated video authentication technology based on multi-view consistency. Background Technology

[0002] With the rapid development of generative artificial intelligence technology, video generation models (such as diffusion models and world models) are now capable of synthesizing highly realistic video content. While these technologies bring convenience to film and television production, virtual reality, and other scenarios, they also seriously threaten the credibility of real-world information. Maliciously generated fake videos could be used to spread fake news, commit telecommunications fraud, or manipulate public opinion, posing a significant challenge to social security. Therefore, developing efficient AI-generated video detection technology has become a critical issue that urgently needs to be addressed in the field of computer vision.

[0003] Early research primarily focused on deepfake detection, creating datasets such as DFDC and FaceForensics++. However, these datasets only focused on facial manipulation and could not support the detection needs of videos generated for general scenarios. Subsequently, GVF, GenVideo, and others attempted to build general AI-generated video datasets, but they suffered from the following problems: 1. Lack of real-world scenario relevance: Existing datasets often contain non-realistic content such as animation and game clips, failing to reflect socially impactful video types in the real world (such as driving scenes, human activities, etc.). This data is disconnected from actual detection needs, making it difficult for models to capture the forgery features of realistic simulated videos. 2. Inconsistent generation quality: Existing datasets do not systematically integrate the latest generation models, resulting in generally low video quality. Detection methods relying on generation flaws (such as pixel discontinuities and abnormal lighting) are easily bypassed by high-quality generated videos. 3. Limited modality: Most datasets only support a single generation mode (such as text to video), lacking coverage of cross-modal (text, image, video to video) generated content, limiting the generalization ability of detection models.

[0004] In recent years, some studies have attempted to introduce physical priors to improve detection robustness. For example: 1. 3D reconstruction consistency verification: Dust3R reconstructs 3D structures through multi-view stereo matching, which can detect geometric projection inconsistencies in generated videos. However, its computational complexity is high, making it difficult to directly apply to real-time detection. 2. Dynamic physical law modeling: Navigation world models simulate object motion trajectories through a physics engine, but they rely on precise physical parameter inputs, making them difficult to generalize to open scenes. These methods suggest that physical consistency can serve as a key feature for distinguishing generated videos. Summary of the Invention

[0005] The present invention aims to at least partially solve one of the technical problems in the related art.

[0006] The purpose of this invention is to provide an efficient and general framework that focuses on the physical plausibility and multi-view consistency in video generation, thereby enabling the detection of AI-generated real-world simulated videos. This framework should be able to handle various types of AI-generated videos, including but not limited to those using state-of-the-art generation techniques and highly realistic visual effects. Existing AI-generated video detection methods struggle to detect high-quality AI-generated videos. This is partly due to the lack of high-quality datasets for existing video generation, leading trained models to focus on features unrelated to fake videos, overfitting to defects in generated videos while neglecting physical plausibility. Furthermore, existing methods lack utilization of physical laws such as multi-view consistency present in real-world videos. Therefore, this invention proposes a large-scale, high-quality generated video dataset and utilizes a multi-view consistency prior design model to detect AI-generated videos, effectively improving the robustness and accuracy of AI-generated video detection.

[0007] First, this invention proposes a process for constructing real-world simulation data, using the video from the previous few frames as cues to generate future videos, thereby constructing data with highly simulated semantic, color, and physical information, and building a large-scale dataset. Second, it also proposes a method for generating videos using artificial intelligence based on multi-view consistency physical prior detection, achieving higher robustness and accuracy. Finally, this invention designs a temporal memory module in the model to dynamically store and update stereo reconstruction features (such as point cloud projection residuals) between video frames, and combines an attention mechanism to enhance long-term consistency analysis.

[0008] To address this, this invention utilizes a multimodal video understanding large language model to understand existing real-world scene videos. Simultaneously, it designs prompts to mark the text annotations of the current video, thereby unifying the generation of prompts across multiple modalities. This invention constructs a large-scale, high-quality real-world simulation dataset. Secondly, it employs a real-time multi-view matching model to process videos with multiple views, extracting video features and using these features at each moment to interact with and update a memory module. After obtaining the video features for each frame, a scorer is used to score the video, and a global average is performed to obtain the final score to determine the video's authenticity.

[0009] The purpose of this invention is to propose a method for detecting fake videos generated by artificial intelligence based on multi-view consistency, in order to address the insufficient detection capability of existing methods for high-quality generated videos. Specifically, this invention constructs a large-scale simulation dataset and designs a detection model that integrates multi-view consistency priors and temporal memory modules, effectively improving the accuracy and robustness of fake video identification.

[0010] Another objective of this invention is to propose an AI-generated video authentication device based on multi-view consistency.

[0011] A third objective of this invention is to provide a non-transitory computer-readable storage medium.

[0012] To achieve the above objectives, this invention proposes a method for detecting fake videos generated by artificial intelligence based on multi-view consistency, comprising: Acquire real-world scene data and use a large video understanding model to annotate the real-world scene data with data prompts to construct a real-world scene data prompt set; Input real-world scene data cues into generative models with different modal cues to generate videos with varying degrees of simulation. The constructed multi-view reconstruction model is used to extract multi-view reconstruction prior features from real videos and simulated videos; wherein, the real videos are derived from the real scene data; The multi-view reconstruction prior features are fused by the constructed memory module to generate a video temporal prior feature sequence; A spatial decoder and scorer are used to perform a comprehensive analysis of the video's temporal prior feature sequence to obtain the analysis results, which are then used to identify the authenticity of the video.

[0013] The AI-generated video authentication method based on multi-view consistency according to embodiments of the present invention may also have the following additional technical features: Furthermore, in one embodiment of the present invention, the generation model includes text, image, and video generation models; the step of acquiring real-scene data and using a large video understanding model to annotate the real-scene data to construct a real-scene data prompt set further includes: using a text generation model to extract and annotate the text of the video in the real-scene data to generate text modal prompts; using an image generation model to extract and annotate any frame of the image modality of the video in the real-scene data to generate image modal prompts; using a video generation model to extract and annotate the first few frames of the video modality in the real-scene data to generate video modal prompts; and constructing a real-scene data prompt set based on text modal prompts, base image modal prompts, and video modal prompts.

[0014] Furthermore, in one embodiment of the present invention, the step of inputting the real-world scene data cue set into the generation models of different modal cues to generate videos with different degrees of simulation further includes: inputting text modal cues into a text generation model to construct a simulation video that is semantically similar to a real-world scene; inputting image modal cues into an image generation model to construct a simulation video that is semantically, stylistically, and color-wise similar to a real-world scene; and inputting video modal cues into a video generation model to construct a simulation video that is semantically, color-wise, and physically similar to a real-world scene.

[0015] Furthermore, in one embodiment of the present invention, the step of fusing the multi-view reconstruction prior features through the constructed memory module to generate a video temporal prior feature sequence further includes: constructing a memory model; obtaining the current frame prior information of the multi-view reconstruction prior features; using the memory model and a temporal cross-attention mechanism to perform feature fusion of the current frame prior information and the existing information of the previous frame to obtain the temporal fusion information of the current frame; and fusing the temporal fusion information of all frames to obtain the video temporal prior feature sequence.

[0016] Furthermore, in one embodiment of the present invention, the step of using a spatial decoder and a scorer to perform a comprehensive analysis of the video temporal prior feature sequence to obtain the analysis result further includes: constructing a decoder based on an attention mechanism, using the decoder to aggregate the spatial information of each frame of the video temporal prior feature sequence to obtain aggregated information; and using a global scorer of a multilayer perceptron to perform fusion analysis on the aggregated information to obtain a fusion analysis result of the video authenticity.

[0017] To achieve the above objectives, another aspect of the present invention proposes an artificial intelligence-generated video authentication device based on multi-view consistency, comprising: The Real Scene Data Hint Set Generation Module is used to use a large video understanding model to annotate real scene data with data hints in order to construct a real scene data hint set. The simulation video generation module is used to input real-world scene data cue sets into generation models with different modal cues to generate videos with different levels of simulation. The video temporal prior feature sequence generation module is used to extract multi-view reconstruction prior features of real videos and simulated videos using the constructed multi-view reconstruction model; and to fuse the multi-view reconstruction prior features through the constructed memory module to generate a video temporal prior feature sequence.

[0018] The video authenticity comprehensive analysis module is used to perform comprehensive analysis of the temporal prior feature sequence of the video using a spatial decoder and scorer to obtain analysis results, so as to identify the authenticity of the video based on the analysis results.

[0019] To achieve the above objectives, a third aspect of this application provides a non-transitory computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the method described in the first aspect.

[0020] The AI-generated video authentication method and apparatus based on multi-view consistency in this invention can effectively improve the accuracy and robustness of identifying fake videos.

[0021] The beneficial effects of this invention are as follows: 1) The large-scale real-world simulated video dataset constructed in this invention generates future frames that are highly consistent with physical laws through preceding frame prompts, which significantly improves the quality and authenticity of the training data and provides a reliable training foundation for the detection model.

[0022] 2) The proposed detection method based on multi-view consistent physical priors uses a real-time multi-view matching model to extract stereo features from the video and combines a temporal memory module to dynamically store and update inter-frame reconstruction residuals (such as point cloud projection errors), effectively capturing the inconsistencies in spatial geometry and temporal evolution of AI-generated videos, and significantly enhancing the accuracy and robustness of detection.

[0023] 3) In addition, a multimodal large language model was introduced to generate unified text prompts and labeled data, which improved the model's generalization ability to diverse inputs; long-term dependency analysis was strengthened through attention mechanism, and then a scorer was used for global scoring and judgment, which enabled accurate identification of high-quality fake videos.

[0024] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0025] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 A flowchart for an AI-generated video authentication method based on multi-view consistency; Figure 2 Example images labeled with real-world scene data; Figure 3 A flowchart illustrating the construction process for generating videos with multiple modal cues; Figure 4 This is a schematic diagram of the structure of an AI-generated video authentication device based on multi-view consistency according to an embodiment of the present invention. Detailed Implementation

[0026] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0027] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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 should fall within the scope of protection of the present invention.

[0028] The following description, with reference to the accompanying drawings, describes an AI-generated video authentication method and apparatus based on multi-view consistency, according to embodiments of the present invention.

[0029] This invention mainly includes a process for constructing a large-scale, high-quality video dataset and a multi-view... Figure 1 A model for identifying prior knowledge of consistency. Figure 2 To build a process for generating video datasets, for existing real-world scene datasets, a large video understanding model is used for text annotation, thereby unifying the process of generating videos based on multiple modal cues. Figure 3 For this model design, the present invention employs a real-time multi-view matching model to process multi-view videos, accurately extracting frame features and dynamically updating them in conjunction with a memory module. Each frame's features are independently evaluated by a scorer, and a global average is used to generate the final score, thereby efficiently determining the authenticity of the video.

[0030] Figure 1 This is a flowchart of an AI-generated video authentication method based on multi-view consistency.

[0031] like Figure 1 As shown, the method includes, but is not limited to, the following steps: S1. Acquire real-world scene data and use a video understanding model to annotate the real-world scene data with data prompts to construct a real-world scene data prompt set.

[0032] Understandably, this invention first constructs a real-world scene data cue set as the basic input to drive the generation of high-quality simulated videos. To adapt to multimodal video generation models (including text-to-video, image-to-video, and video-to-video modalities), this invention adopts a unified data cue construction strategy: utilizing a multimodal video understanding large language model to perform semantic parsing and content understanding of real-world scene videos, automatically generating accurate text annotations to form text modal cue; for image modal cue, single-frame images are randomly extracted from real videos as visual guidance for the generation model; for video modal cue, the first few frames of the video are extracted as input to preserve initial temporal dynamic information. Based on the above methods, a large-scale, high-quality real-world scene cue dataset covering text, image, and video modalities is constructed, ensuring that the generated videos possess high realism in semantics, visual style, and spatiotemporal dynamics (e.g., Figure 2 ).

[0033] S2 inputs a set of real-world scene data cues into the generation model of different modal cues to generate videos with different levels of simulation.

[0034] Specifically, based on the constructed prompt dataset, this invention utilizes ten of the most advanced AI video generation models to generate simulated videos, covering all mainstream prompt modalities. For videos generated from text prompts, the semantic content is consistent with the real scene, but visual details such as color and texture are determined by the model's own style, resulting in physical inconsistencies. For videos generated from image prompts, the visual features of the first frame are highly consistent with real images, and the style is highly controllable, but due to the lack of complete temporal guidance, the motion process is prone to incoherence or violations of physical laws. For videos generated from video prompts, because the input contains multi-frame dynamic information, the model can learn real motion patterns well, achieving high fidelity in semantics, appearance, and physical behavior, posing a greater challenge to detection models. Through multimodal and multi-model generation strategies, a simulated video dataset covering different levels of forgery and generation mechanisms (such as...) was constructed. Figure 3 This effectively improves the generalization ability of the detection model.

[0035] S3, using the constructed multi-view reconstruction model to extract multi-view reconstruction prior features of real videos and simulated videos; wherein, the real videos come from the real scene data.

[0036] Furthermore, this invention conducts multi-view consistency analysis on real videos and AI-generated videos, finding that the consistency of geometric structure, motion trajectory, and lighting in real videos is significantly better than that in generated videos under different viewpoints, indicating that the generation model cannot fully simulate the multi-view physical constraints in the real world. Inspired by this, this invention introduces a multi-view reconstruction model to extract multi-view reconstruction prior features of videos, such as 3D point cloud projection residuals and disparity consistency errors, as key indicators reflecting physical rationality. This feature can effectively expose inconsistencies in depth structure and viewpoint transformations in generated videos:

[0037] S4, the multi-view reconstruction prior features are fused through the constructed memory module to generate a video temporal prior feature sequence.

[0038] Furthermore, to model the long-term temporal dependencies of videos, this invention designs a temporal memory module. This module maintains a dynamically updated feature sequence, and for each frame, the extracted multi-view prior features are fused with historical features through a temporal cross-attention mechanism to capture the consistent evolution between frames. The final output is a video temporal prior feature sequence containing spatiotemporal consistency. .

[0039] S5 uses a spatial decoder and scorer to perform a comprehensive analysis of the video's temporal prior feature sequence to obtain the analysis results, which are then used to identify the authenticity of the video.

[0040] Specifically, the decoder aggregates spatial features of each frame based on an attention mechanism, and the scorer uses a multilayer perceptron to fuse temporal features, outputting a comprehensive score for video authenticity, thus achieving accurate identification of AI-generated videos.

[0041] .

[0042] To achieve the above embodiments, such as Figure 4 As shown, this embodiment also provides an AI-generated video authentication device 10 based on multi-view consistency. The device 10 includes: a real scene data prompt set generation module 100, a simulation video generation module 200, a video temporal prior feature sequence generation module 300, and a video authenticity comprehensive analysis module 400.

[0043] The real-scene data prompt set generation module 100 is used to use a large video understanding model to annotate real-scene data with data prompts to construct a real-scene data prompt set; The real-scene data simulation video generation module 200 is used to input the real-scene data prompt set into the generation model of different modal prompts to generate videos with different degrees of simulation. The video temporal prior feature sequence generation module 300 is used to extract multi-view reconstruction prior features of real videos and simulated videos using the constructed multi-view reconstruction model; and to fuse the multi-view reconstruction prior features through the constructed memory module to generate a video temporal prior feature sequence. The video authenticity comprehensive analysis module 400 is used to perform video authenticity comprehensive analysis on the temporal prior feature sequence of the video using a spatial decoder and scorer to obtain analysis results, so as to identify the authenticity of the video based on the analysis results; Furthermore, the real-world scenario data prompt set generation module is also used for: Text generation models are used to extract and annotate text from videos in real-world scene data, generating text modal cues. The image generation model is used to extract and annotate any frame of the image modality in a video from real-world scene data, generating image modality cues. The video generation model is used to extract and annotate the first few frames of video modalities in real-world scene data to generate video modal cues. A set of real-world scene cues was constructed based on text modal cues, base image modal cues, and video modal cues.

[0044] Furthermore, the simulation video generation module is used for: Text modal prompts are input into a text generation model to construct a simulated video that is semantically similar to a real-world scene; Image modality cues are input into the image generation model to construct simulated videos that are semantically, stylistically, and color-wise similar to real-world scenes; Video modal cues are input into the video generation model to construct simulated videos that resemble real-world scenes in terms of semantics, color style, and physical laws.

[0045] Furthermore, the video temporal prior feature sequence generation module is also used for: Build a memory model; Obtain the current frame prior information of the multi-view reconstruction prior features; The temporal fusion information of the current frame is obtained by using a memory model and a temporal cross-attention mechanism to fuse the prior information of the current frame with the existing information of the previous frame. The video temporal prior feature sequence is obtained by fusing temporal fusion information from all frames.

[0046] Furthermore, the video authenticity comprehensive analysis module is also used for: A decoder is constructed based on an attention mechanism, and the spatial information of each frame of the video is aggregated by the decoder using the temporal prior feature sequence of the video to obtain aggregated information. The global scorer of a multilayer perceptron is used to perform fusion analysis on the aggregated information to obtain the fusion analysis results of video authenticity.

[0047] The AI-generated video authentication device based on multi-view consistency according to embodiments of the present invention can effectively improve the detection accuracy of high-quality forged videos. It innovatively utilizes a multimodal large language model to generate unified prompts and construct a high-fidelity dataset, enhancing the model's generalization ability. Through real-time multi-view matching and interaction with a temporal memory module, it dynamically updates inter-frame stereo features and, combined with an attention mechanism and global scoring, accurately identifies forged videos that violate physical laws.

[0048] To implement the above embodiments, this application also proposes a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the method described in the foregoing embodiments.

[0049] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0050] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0051] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for detecting fake videos generated by artificial intelligence based on multi-view consistency, characterized in that, Includes the following steps: Acquire real-world scene data and use a large video understanding model to annotate the real-world scene data with data prompts to construct a real-world scene data prompt set; Real-world scene data cues are input into generative models with different modal cues to generate videos with varying degrees of simulation. The constructed multi-view reconstruction model is used to extract multi-view reconstruction prior features from real videos and simulated videos; wherein, the real videos are derived from the real scene data; The multi-view reconstruction prior features are fused by the constructed memory module to generate a video temporal prior feature sequence; A spatial decoder and scorer are used to perform a comprehensive analysis of the video's temporal prior feature sequence to obtain the analysis results, which are then used to identify the authenticity of the video.

2. The method according to claim 1, characterized in that, Generative models, including models for generating text, images, and videos; the acquisition of real-world scene data and the use of a large video understanding model to annotate the real-world scene data with data prompts to construct a real-world scene data prompt set, also includes: Text generation models are used to extract and annotate text from videos in real-world scene data, generating text modal cues. The image generation model is used to extract and annotate any frame of the image modality in a video from real-world scene data, generating image modality cues. The video generation model is used to extract and annotate the first few frames of video modalities in real-world scene data to generate video modal cues. A set of real-world scene cues was constructed based on text modal cues, base image modal cues, and video modal cues.

3. The method according to claim 2, characterized in that, The step of inputting a set of real-world scene data cues into a generation model for different modal cues to generate videos with different levels of simulation also includes: Text modal prompts are input into a text generation model to construct a simulated video that is semantically similar to a real-world scene; Image modality cues are input into the image generation model to construct simulated videos that are semantically, stylistically, and color-wise similar to real-world scenes; Video modal cues are input into the video generation model to construct simulated videos that resemble real-world scenes in terms of semantics, color style, and physical laws.

4. The method according to claim 3, characterized in that, The step of fusing the multi-view reconstructed prior features through the constructed memory module to generate a video temporal prior feature sequence further includes: Build a memory model; Obtain the current frame prior information of the multi-view reconstruction prior features; The temporal fusion information of the current frame is obtained by using a memory model and a temporal cross-attention mechanism to fuse the prior information of the current frame with the existing information of the previous frame. The video temporal prior feature sequence is obtained by fusing temporal fusion information from all frames.

5. The method according to claim 4, characterized in that, The method of using a spatial decoder and scorer to perform a comprehensive analysis of the video's temporal prior feature sequence to obtain the analysis results also includes: A decoder is constructed based on an attention mechanism, and the spatial information of each frame of the video is aggregated by the decoder using the temporal prior feature sequence of the video to obtain aggregated information. The global scorer of a multilayer perceptron is used to perform fusion analysis on the aggregated information to obtain the fusion analysis results of video authenticity.

6. The method according to claim 1, characterized in that, The multi-view reconstruction prior features of real and simulated videos are extracted using the constructed multi-view reconstruction model, which also includes: Multi-view reconstruction models are used to extract prior features from videos, and video authenticity analysis is performed based on these prior features. 。 7. The method according to claim 4, characterized in that, The method further includes fusing the multi-view reconstruction prior features through a constructed memory module to generate a video temporal prior feature sequence, and also includes: 。 8. The method according to claim 5, characterized in that, The analysis results are obtained by using a spatial decoder and scorer to perform a comprehensive analysis of the video temporal prior feature sequences, and also include: For the obtained temporal prior feature sequence of the video, a decoder is constructed based on an attention mechanism to analyze the video features and aggregate the spatial information of each frame of the video: A global scorer based on a simple multilayer perceptron is used to fuse the temporal information of the video, thereby obtaining a comprehensive analysis of the video's authenticity: 。 9. A device for detecting fake videos generated by artificial intelligence based on multi-view consistency, characterized in that, include: The Real Scene Data Hint Set Generation Module is used to use a large video understanding model to annotate real scene data with data hints in order to construct a real scene data hint set. The simulation video generation module is used to input real-world scene data cue sets into generation models with different modal cues to generate videos with different levels of simulation. The video temporal prior feature sequence generation module is used to extract multi-view reconstruction prior features of real and simulated videos using the constructed multi-view reconstruction model. The multi-view reconstruction prior features are fused using a constructed memory module to generate a video temporal prior feature sequence. The video authenticity comprehensive analysis module is used to perform comprehensive analysis of the temporal prior feature sequence of the video using a spatial decoder and scorer to obtain analysis results, so as to identify the authenticity of the video based on the analysis results.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the AI-generated video authentication method based on multi-view consistency as described in any one of claims 1-8.

Images