A method, device, equipment, medium and program product for detecting a fake video

Abstract

The application discloses a counterfeit video detection method, device, equipment, medium and program product. The method comprises the following steps: acquiring at least two target image frames of a to-be-detected video; wherein the target image frames are obtained by screening based on the differences between different image frames in the to-be-detected video; acquiring image description information of the target image frames; performing a preset each authenticity detection task on target detection data of the to-be-detected video to obtain a detection result corresponding to each authenticity detection task; wherein the target detection data at least comprises the image description information; and determining a counterfeit detection result of the to-be-detected video based on each detection result. The application can identify the authenticity of the content of the to-be-detected video, avoid the training set coverage limitation caused by recognizing the visual attributes based on the face and the picture, has better generalization, and can further improve the accuracy of detecting various counterfeit videos.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method, apparatus, electronic device, computer-readable storage medium, and computer program product for detecting fake videos. Background Technology

[0002] Currently, AI (Artificial Intelligence) face-swapping technology, especially deepfake technology, can produce highly realistic videos that are difficult to distinguish from the real thing with the naked eye. This technology is widely used in film production, communication and social networking, and other applications, but it also poses potential security risks such as fraud and infringement of personal privacy. Therefore, effective detection of fake videos is particularly important.

[0003] In existing technologies, facial features and image features are typically extracted from fake and real images to form a training set, which is then used to train a detection model. This model is then used to detect fake videos. However, this method only identifies the visual attributes of faces and images. If the training set is not comprehensive, it can easily lead to deviations in the detection results of fake videos, thus failing to effectively and accurately detect fake videos. Summary of the Invention

[0004] This invention provides a method, apparatus, device, medium, and program product for detecting fake videos, in order to solve the technical problem that existing technologies cannot effectively and accurately detect fake videos.

[0005] To address the aforementioned technical problems, a first aspect of this invention provides a method for detecting forged videos, comprising: At least two target image frames of the video to be detected are obtained; wherein the target image frames are obtained by filtering based on the differences between different image frames in the video to be detected; Obtain the image description information of the target image frame; The target detection data of the video to be detected is subjected to preset authenticity detection tasks to obtain the detection results corresponding to each authenticity detection task; wherein, the target detection data includes at least the image description information; Based on each of the detection results, the forgery detection result of the video to be detected is determined.

[0006] As a preferred embodiment, obtaining the image description information of the target image frame specifically includes: Based on a preset face database, the target image frame is subjected to person recognition to obtain person recognition results; Based on the person recognition result and the target image frame, generate an image description instruction for the target image frame; The image description command is input into a multimodal large model for analysis to obtain the image description information.

[0007] As a preferred embodiment, the authenticity detection task includes at least an event authenticity detection task and an image description information authenticity detection task.

[0008] As a preferred embodiment, the step of performing preset authenticity detection tasks on the target detection data of the video to be detected, and obtaining the detection results corresponding to each authenticity detection task, specifically includes: When the authenticity detection task to be executed is the authenticity detection task of the event, the large language model is used to extract event information from the image description information to obtain the target event information. Based on a preset event database, the authenticity of the target event information is detected to obtain the detection results corresponding to the event authenticity detection task.

[0009] As a preferred embodiment, the step of extracting event information from the image description information using a large language model to obtain target event information specifically includes: Generate an event information extraction instruction according to at least one preset event information type to be acquired and the image description information; The event information extraction command is input into the large language model to extract event information and obtain initial extracted event information. If the initial extracted event information includes event information of each of the event information types to be acquired, then the initial extracted event information is taken as the target event information; If any event information of the type of event information to be acquired is missing in the initial extracted event information, the missing event information in the initial extracted event information is supplemented using the video metadata of the video to be detected, thereby obtaining the target event information.

[0010] As a preferred embodiment, the step of performing preset authenticity detection tasks on the target detection data of the video to be detected, and obtaining the detection results corresponding to each authenticity detection task, specifically further includes: When the authenticity detection task to be executed is the authenticity detection task of the image description information, an image description information authenticity discrimination instruction is generated based on the image description information. The image description information authenticity judgment instruction is input into the large language model to perform image description information authenticity judgment, and the detection result corresponding to the image description information authenticity detection task is obtained.

[0011] As a preferred embodiment, the target detection data also includes audio and video data between two adjacent target image frames.

[0012] As a preferred embodiment, the authenticity detection task also includes a lip shape authenticity detection task; The step of performing preset authenticity detection tasks on the target detection data of the video to be detected, and obtaining the detection results corresponding to each authenticity detection task, specifically includes: When the authenticity detection task to be performed is the lip shape authenticity detection task, each image frame to be tested and its corresponding audio data are obtained from the audio and video data. For each of the image frames to be tested and its corresponding audio data, feature extraction is performed to obtain at least one feature pair; wherein, the feature pair includes the lip shape feature vector corresponding to the image frame to be tested and the audio feature vector corresponding to the audio data; Based on the cosine similarity of each set of feature pairs, lip-sync detection is performed on each image frame to be tested and its corresponding audio data to obtain the detection result corresponding to the lip-sync authenticity detection task.

[0013] As a preferred embodiment, the authenticity detection task also includes a subtitle matching detection task; The step of performing preset authenticity detection tasks on the target detection data of the video to be detected, and obtaining the detection results corresponding to each authenticity detection task, specifically includes: When the authenticity detection task to be performed is the subtitle matching detection task, subtitle text and audio text are extracted from the audio and video data; The subtitle text and the audio text are respectively transformed into feature vectors to obtain the subtitle text feature vector and the audio text feature vector; Based on the cosine similarity between the feature vectors of the subtitle text and the feature vectors of the audio text, the subtitle text and the audio text are matched and detected to obtain the detection result corresponding to the subtitle matching and detection task.

[0014] As a preferred embodiment, acquiring at least two target image frames of the video to be detected specifically includes: At least two key image frames are extracted from the video to be detected according to a preset extraction time interval. Based on the intensity value distribution difference between two adjacent key image frames, at least two target image frames are selected from one or more groups of adjacent key image frames according to the filtering order of the intensity value distribution difference from large to small.

[0015] As a preferred embodiment, determining the forgery detection result of the video to be detected based on each of the detection results specifically includes: Based on each of the detection results and the preset weight values, the weighted confidence score corresponding to each of the detection results is obtained; Based on the weighted confidence scores and the preset total confidence score threshold, the forgery detection result of the video to be detected is determined.

[0016] A second aspect of the present invention provides a forged video detection device, comprising: The target image frame acquisition module is used to acquire at least two target image frames of the video to be detected; wherein, the target image frames are obtained by filtering based on the differences between different image frames in the video to be detected; Image description information acquisition module, used to acquire image description information of the target image frame; The authenticity detection module is used to perform preset authenticity detection tasks on the target detection data of the video to be detected, and obtain the detection results corresponding to each authenticity detection task; wherein, the target detection data includes at least the image description information; The forgery detection result determination module is used to determine the forgery detection result of the video to be detected based on each of the detection results.

[0017] As a preferred embodiment, the image description information acquisition module is used to acquire the image description information of the target image frame, specifically including: Based on a preset face database, the target image frame is subjected to person recognition to obtain person recognition results; Based on the person recognition result and the target image frame, generate an image description instruction for the target image frame; The image description command is input into a multimodal large model for analysis to obtain the image description information.

[0018] As a preferred embodiment, the authenticity detection task includes at least an event authenticity detection task and an image description information authenticity detection task.

[0019] As a preferred embodiment, the authenticity detection module is used to perform preset authenticity detection tasks on the target detection data of the video to be detected, and obtain the detection results corresponding to each authenticity detection task, specifically including: When the authenticity detection task to be executed is the authenticity detection task of the event, the large language model is used to extract event information from the image description information to obtain the target event information. Based on a preset event database, the authenticity of the target event information is detected to obtain the detection results corresponding to the event authenticity detection task.

[0020] As a preferred embodiment, the authenticity detection module is used to extract event information from the image description information using a large language model to obtain target event information, specifically including: Generate an event information extraction instruction according to at least one preset event information type to be acquired and the image description information; The event information extraction command is input into the large language model to extract event information and obtain initial extracted event information. If the initial extracted event information includes event information of each of the event information types to be acquired, then the initial extracted event information is taken as the target event information; If any event information of the type of event information to be acquired is missing in the initial extracted event information, the missing event information in the initial extracted event information is supplemented using the video metadata of the video to be detected, thereby obtaining the target event information.

[0021] As a preferred embodiment, the authenticity detection module is used to perform preset authenticity detection tasks on the target detection data of the video to be detected, and obtain the detection results corresponding to each authenticity detection task, specifically including: When the authenticity detection task to be executed is the authenticity detection task of the image description information, an image description information authenticity discrimination instruction is generated based on the image description information. The image description information authenticity judgment instruction is input into the large language model to perform image description information authenticity judgment, and the detection result corresponding to the image description information authenticity detection task is obtained.

[0022] As a preferred embodiment, the target detection data also includes audio and video data between two adjacent target image frames.

[0023] As a preferred embodiment, the authenticity detection task also includes a lip shape authenticity detection task; The authenticity detection module is used to perform preset authenticity detection tasks on the target detection data of the video to be detected, and obtain the detection results corresponding to each authenticity detection task. Specifically, it also includes: When the authenticity detection task to be performed is the lip shape authenticity detection task, each image frame to be tested and its corresponding audio data are obtained from the audio and video data. For each of the image frames to be tested and its corresponding audio data, feature extraction is performed to obtain at least one feature pair; wherein, the feature pair includes the lip shape feature vector corresponding to the image frame to be tested and the audio feature vector corresponding to the audio data; Based on the cosine similarity of each set of feature pairs, lip-sync detection is performed on each image frame to be tested and its corresponding audio data to obtain the detection result corresponding to the lip-sync authenticity detection task.

[0024] As a preferred embodiment, the authenticity detection task also includes a subtitle matching detection task; The authenticity detection module is used to perform preset authenticity detection tasks on the target detection data of the video to be detected, and obtain the detection results corresponding to each authenticity detection task. Specifically, it also includes: When the authenticity detection task to be performed is the subtitle matching detection task, subtitle text and audio text are extracted from the audio and video data; The subtitle text and the audio text are respectively transformed into feature vectors to obtain the subtitle text feature vector and the audio text feature vector; Based on the cosine similarity between the feature vectors of the subtitle text and the feature vectors of the audio text, the subtitle text and the audio text are matched and detected to obtain the detection result corresponding to the subtitle matching and detection task.

[0025] As a preferred embodiment, the target image frame acquisition module is used to acquire at least two target image frames of the video to be detected, specifically including: At least two key image frames are extracted from the video to be detected according to a preset extraction time interval. Based on the intensity value distribution difference between two adjacent key image frames, at least two target image frames are selected from one or more groups of adjacent key image frames according to the filtering order of the intensity value distribution difference from large to small.

[0026] As a preferred embodiment, the forgery detection result determination module is used to determine the forgery detection result of the video to be detected based on each of the detection results, specifically including: Based on each of the detection results and the preset weight values, the weighted confidence score corresponding to each of the detection results is obtained; Based on the weighted confidence scores and the preset total confidence score threshold, the forgery detection result of the video to be detected is determined.

[0027] A third aspect of the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the fake video detection method described in any of the first aspects.

[0028] A fourth aspect of the present invention provides a computer-readable storage medium comprising a stored computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to perform the fake video detection method described in any of the first aspects.

[0029] A fifth aspect of the present invention provides a computer program product, including a computer program / instructions, wherein when the computer program / instructions are executed by a processor, they implement the steps of the fake video detection method described in any of the first aspects.

[0030] Compared with the prior art, the beneficial effect of the embodiments of the present invention is that by filtering at least two target image frames from the video to be detected based on the differences between different image frames, and performing forgery detection of the video to be detected based on the image description information of the target image frames, it is possible to identify the authenticity of the content of the video to be detected, avoid the limitation of training set coverage caused by recognition based on the visual attributes of faces and images, have better generalization, and thus improve the accuracy of detecting various forged videos. Attached Figure Description

[0031] Figure 1 This is a flowchart illustrating the fake video detection method in an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of the fake video detection device in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of the electronic device in an embodiment of the present invention. Detailed Implementation

[0032] 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.

[0033] Please see Figure 1 The first aspect of this invention provides a method for detecting fake videos, comprising the following steps S1 to S4: Step S1: Obtain at least two target image frames from the video to be detected; wherein the target image frames are obtained by filtering based on the differences between different image frames in the video to be detected.

[0034] Specifically, based on considerations of efficiency and computing power, this embodiment does not detect all image frames in the video to be detected. Instead, it extracts at least two target image frames with significant differences for detection. The extracted target image frames have many changes in the picture, and the picture content will undergo substantial changes, such as sudden changes in the expression of the person after face swapping, abnormal connection of facial lighting and shadow, and discontinuity in the matching of background and foreground. Therefore, it is easier to capture traces of forgery by detecting these target image frames.

[0035] Step S2: Obtain the image description information of the target image frame.

[0036] Furthermore, considering that current AI face-swapping and deepfake technologies can produce highly realistic visual effects, and that video forgery usually alters the content of the video itself, this embodiment obtains image description information of the target image frame to identify the authenticity of the video based on the content of the video itself. This avoids the limitations of the training set coverage caused by recognition based on visual attributes of faces and images such as lighting, contrast, and saturation, and ensures high detection accuracy when faced with forged videos created using different forgery technologies.

[0037] Step S3: Perform preset authenticity detection tasks on the target detection data of the video to be detected, and obtain the detection results corresponding to each authenticity detection task; wherein, the target detection data includes at least the image description information.

[0038] Furthermore, preset authenticity detection tasks are performed on the target detection data of the video to be detected. The target detection data includes at least the image description information of the acquired target image frames, so as to combine different factors to determine the authenticity of the video to be detected and ensure the accuracy of the forgery detection of the video to be detected.

[0039] Step S4: Based on each of the detection results, determine the forgery detection result of the video to be detected.

[0040] Furthermore, after obtaining the detection results corresponding to each authenticity detection task, the possibility of the current video being detected being forged can be reflected, thus clarifying the forgery detection result of the video being detected.

[0041] The fake video detection method provided in this invention filters out at least two target image frames from the video to be detected based on the differences between different image frames, and performs fake detection on the video to be detected based on the image description information of the target image frames. This enables the identification of the authenticity of the content of the video to be detected, avoids the limitation of training set coverage caused by recognition based on the visual attributes of faces and images, has better generalization, and thus improves the accuracy of detecting various fake videos.

[0042] As a preferred embodiment, obtaining the image description information of the target image frame specifically includes: Based on a preset face database, the target image frame is subjected to person recognition to obtain person recognition results; Based on the person recognition result and the target image frame, generate an image description instruction for the target image frame; The image description command is input into a multimodal large model for analysis to obtain the image description information.

[0043] Specifically, to avoid information deviations when describing the target image frame, this embodiment first performs person recognition on the target image frame based on a preset face database. It is understood that the face database in this embodiment can be an open face database such as FACELIB, which is developed based on face search technology and can quickly identify the identity, gender, and age of public figures. Alternatively, the face database in this embodiment can be a self-trained face database. However, considering the difficulty of collecting facial information from public figures, a self-trained face database is more suitable for small-scale business scenarios, primarily collecting facial information from various users within the business. If no specific person is identified, the person's gender is identified; a male can be recorded as a male, and a female as a female. For example, the name variable is recorded as {name}, and the gender variable is recorded as {s}.

[0044] Furthermore, according to a preset image description instruction template, and based on the target image frame and its person recognition results, an image description instruction for the target image frame can be formed by filling in the image description instruction template. For example, the image description instruction template is shown below: Please observe the following image, understand its content, and describe it in detail: {image} Notice: The person in the image is {name}, and is of {s} gender; The description includes, but is not limited to: the relationship between the observed environment, objects, locations, time, and people, as well as the people's actions and expressions, etc. Then, the image description instruction is input into the multimodal large model for analysis. It is worth noting that the multimodal large model in this embodiment is a large model that can complete image understanding and image interpretation tasks. After receiving the image description instruction, it can provide in-depth analysis and interpretation of the image in the image description instruction, thereby obtaining image description information.

[0045] In one optional embodiment, the image description instruction is a prompt instruction. The image description process can be executed by constructing a workflow, where a workflow is a method of breaking down a complex task into a series of steps or stages. For example, the image description workflow can be: [Start] → [Input Image] → [Open Face Database Face Recognition] → [Person Matching] → [Record Person Name / Gender] → [Construct Prompt Instruction] → [Call Multimodal Large Model] → [End].

[0046] This invention identifies people in target image frames by performing human recognition, thereby avoiding information bias during image description. Furthermore, by utilizing a multimodal large model to describe the target image frames, the accuracy of image description can be improved.

[0047] As a preferred embodiment, the authenticity detection task includes at least an event authenticity detection task and an image description information authenticity detection task.

[0048] Specifically, this embodiment considers event authenticity detection and image description information authenticity detection as minimum required authentication tasks. It is understood that this embodiment takes into account that current methods for detecting fake videos neglect the assessment of the reasonableness of the video content. Reasonableness here refers to whether the combination of factors such as the actual content of the video, geographical location, physical relationships within the image, people, current events, culture, and history is authentic. By performing the event authenticity detection task, it is possible to determine whether the event in the target image frame actually occurred. Similarly, by performing the image description information authenticity detection task, it is possible to determine whether the content in the target image frame is reasonable. This allows for a comprehensive and accurate determination of the authenticity of the content in the video being detected, while avoiding biased detection results due to a single aspect of the assessment.

[0049] As a preferred embodiment, the step of performing preset authenticity detection tasks on the target detection data of the video to be detected, and obtaining the detection results corresponding to each authenticity detection task, specifically includes: When the authenticity detection task to be executed is the authenticity detection task of the event, the large language model is used to extract event information from the image description information to obtain the target event information. Based on a preset event database, the authenticity of the target event information is detected to obtain the detection results corresponding to the event authenticity detection task.

[0050] Specifically, when performing the event authenticity detection task, considering that the large language model is a deep learning model trained with a large amount of text data, enabling the model to understand the meaning of language text, this embodiment utilizes the large language model to extract event information from the image description information. This event information consists of entities related to the event, such as time, location, people, and event type, etc., which are not specifically limited in this embodiment. After obtaining the target event information from the image description information, event matching is performed based on a preset event database to determine whether the event composed of these target event information actually occurred. The event database refers to a database that collects and records various events occurring globally. The most well-known example is the GDELT project (Global Database of Events, Language, and Tone). The GDELT project monitors global broadcast, print, and online news to identify people, locations, organizations, topics, sources, sentiments, counts, citations, images, and events, providing a real-time open platform for the global community. The specific usage process of the GDELT project is as follows: Obtain the GDELT API Key (GDELT Application Programming Interface Key) from the official website, construct query parameters based on the target event information, then obtain the response information via an HTTP GET request, parse the JSON data, and extract relevant information as needed. Based on the information of each event in the event database, verify whether the combination of target event information is reasonable, such as whether the time, location, and participants are reasonable.

[0051] This invention utilizes an event database to perform authenticity detection on target event information in image description information, effectively verifying whether the target event information truly exists and ensuring the accuracy of the detection results corresponding to the event authenticity detection task.

[0052] As a preferred embodiment, the step of extracting event information from the image description information using a large language model to obtain target event information specifically includes: Generate an event information extraction instruction according to at least one preset event information type to be acquired and the image description information; The event information extraction command is input into the large language model to extract event information and obtain initial extracted event information. If the initial extracted event information includes event information of each of the event information types to be acquired, then the initial extracted event information is taken as the target event information; If any event information of the type of event information to be acquired is missing in the initial extracted event information, the missing event information in the initial extracted event information is supplemented using the video metadata of the video to be detected, thereby obtaining the target event information.

[0053] Specifically, this embodiment generates an event information extraction instruction based on at least one preset type of event information to be acquired, such as time, location, event type, participants, etc., combined with image description information, as shown in the following example: Please help extract the time, location, people, and event type from the following image description information: {content} The extracted results are returned in the following format: {date:"", location:"", persons:"","type":""}.

[0054] The event information extraction command is then input into a large language model to extract event information, obtaining initial extracted event information. Since the large language model may not be able to completely extract event information from the image description information, the initial extracted event information may not cover all event types to be acquired. Therefore, if the initial extracted event information includes event information for each event type to be acquired, it is directly used as the target event information; if any event type is missing from the initial extracted event information, the corresponding missing event information is extracted from the video metadata of the video to be detected and supplemented to form complete target event information. In this embodiment, video metadata includes, but is not limited to, video time and shooting location.

[0055] This invention, in its embodiments, supplements the missing event information in the initial extracted event information output by the large language model using the video metadata of the video to be detected. This ensures the completeness of the target event information in the image description information and guarantees the accuracy of the authenticity detection of the target event information.

[0056] As a preferred embodiment, the step of performing preset authenticity detection tasks on the target detection data of the video to be detected, and obtaining the detection results corresponding to each authenticity detection task, specifically further includes: When the authenticity detection task to be executed is the authenticity detection task of the image description information, an image description information authenticity discrimination instruction is generated based on the image description information. The image description information authenticity judgment instruction is input into the large language model to perform image description information authenticity judgment, and the detection result corresponding to the image description information authenticity detection task is obtained.

[0057] Specifically, the large language model itself possesses information and knowledge from the entire internet, and can determine the authenticity of input information by combining its own information and knowledge. Therefore, in this embodiment, when performing the task of detecting the authenticity of image description information, it first generates an image description information authenticity judgment instruction based on the image description information, as shown in the following example: "You are a professional image authentication expert. Here is an image description as follows:" {content} Please determine the authenticity of this image based on the description.

[0058] If the scene is found to be fake or unrealistic, return 1; if it is determined to be a real scene or is uncertain, return 0. The image description information authenticity judgment instruction is input into the large language model to judge the authenticity of the image description information, thereby directly obtaining the detection result corresponding to the currently executed image description information authenticity detection task.

[0059] This invention utilizes a large language model to determine the authenticity of image description information, thereby effectively and accurately judging the rationality of image description information and ensuring the accuracy of the authenticity detection of image description information.

[0060] As a preferred embodiment, the target detection data also includes audio and video data between two adjacent target image frames.

[0061] Specifically, in order to further improve the accuracy of forgery detection, the target detection data of the video to be detected in this embodiment also includes audio and video data between two adjacent target image frames. This allows for further combination of video elements and audio data between target image frames to perform authenticity detection, so as to distinguish the authenticity of the video to be detected from more perspectives and ensure the accuracy of forgery detection of the video to be detected.

[0062] As a preferred embodiment, the authenticity detection task also includes a lip shape authenticity detection task; The step of performing preset authenticity detection tasks on the target detection data of the video to be detected, and obtaining the detection results corresponding to each authenticity detection task, specifically includes: When the authenticity detection task to be performed is the lip shape authenticity detection task, each image frame to be tested and its corresponding audio data are obtained from the audio and video data. For each of the image frames to be tested and its corresponding audio data, feature extraction is performed to obtain at least one feature pair; wherein, the feature pair includes the lip shape feature vector corresponding to the image frame to be tested and the audio feature vector corresponding to the audio data; Based on the cosine similarity of each set of feature pairs, lip-sync detection is performed on each image frame to be tested and its corresponding audio data to obtain the detection result corresponding to the lip-sync authenticity detection task.

[0063] Specifically, the authenticity detection task in this embodiment also includes a lip-sync detection task, which is performed based on audio and video data between two adjacent target image frames. When performing the lip-sync detection task, this embodiment first determines the audio and video segments between target image frames from the audio and video data. Then, it pairs each image frame to be tested in the video segment with the audio data at the corresponding time point in the audio segment, forming an audio-image data set. Further, the Wav2Lip model is used to extract features from each image frame to be tested and its corresponding audio data, obtaining the lip-sync feature vector corresponding to the image frame and the audio feature vector corresponding to the audio data, such as Mel-spectral features. It is worth noting that the training of the Wav2Lip model is divided into two stages: a pre-training stage and a GAN (Generative Adversarial Networks) training stage. In the pre-training stage, the expert lip-sync discriminator in the Wav2Lip model learns how to accurately judge the synchronization between lip-sync and audio through a large number of audio-image pairs; while the GAN network in the Wav2Lip model is responsible for generating realistic lip movements. In this embodiment, each image frame to be tested and its corresponding audio data are input into the expert lip-sync discriminator in the Wav2Lip model for discrimination. SyncNet (Synchronization Network) serves as the basic model of the expert lip-sync discriminator, which evaluates synchronization by calculating the cosine similarity between the lip-sync feature vector and the audio feature vector. The loss function of the expert lip-sync discriminator can be expressed as: ; in, D ( V , S ) is the lip shape feature vector V and audio feature vectors S The cosine similarity between them is expressed as follows: ; V · S Represents lip shape feature vector V and audio feature vectors S dot product, and These are lip shape feature vectors. V The modulus or norm, and the audio feature vector. S The modulus or norm.

[0064] Based on the loss function described above, the expert lip-sync discriminator outputs the result of whether the lip-sync audio is synchronized, thereby obtaining the detection result corresponding to the lip-sync authenticity detection task.

[0065] This invention employs the lip-sync feature vector corresponding to each image frame under test in the audio and video data and the audio feature vector corresponding to the audio data to perform lip-sync audio synchronization detection. This effectively detects lip-sync audio asynchrony in the video under test, thereby enabling more comprehensive forgery detection of the video under test.

[0066] As a preferred embodiment, the authenticity detection task also includes a subtitle matching detection task; The step of performing preset authenticity detection tasks on the target detection data of the video to be detected, and obtaining the detection results corresponding to each authenticity detection task, specifically includes: When the authenticity detection task to be performed is the subtitle matching detection task, subtitle text and audio text are extracted from the audio and video data; The subtitle text and the audio text are respectively transformed into feature vectors to obtain the subtitle text feature vector and the audio text feature vector; Based on the cosine similarity between the feature vectors of the subtitle text and the feature vectors of the audio text, the subtitle text and the audio text are matched and detected to obtain the detection result corresponding to the subtitle matching and detection task.

[0067] Specifically, the authenticity detection task in this embodiment also includes a subtitle matching detection task, which is performed based on the audio and video data between two adjacent target image frames. When performing the subtitle matching detection task, this embodiment uses the subtitle text within the video segment of the audio and video data as reference information and compares it with the audio text. It is understood that the audio text can be directly extracted through speech recognition.

[0068] Furthermore, in order to compare the subtitle text and the audio text, feature vector conversion needs to be performed on the subtitle text and the audio text separately. In this embodiment, the subtitle text and the audio text are uniformly converted into TF-IDF feature vectors, as shown in the following expression: ; Among them, TF( t , d IDF (Index Frequency) represents the frequency of word t in document d. For subtitle text, the subtitle text of each image frame can be considered as a document d; for audio text, the audio text corresponding to each image frame can be considered as a document d. t ) is the inverse document frequency of word t, calculated using the following formula: ; N This represents the total number of documents. N t This represents the number of documents containing the word t.

[0069] Finally, the cosine similarity between the subtitle text feature vector and the audio text feature vector is used as the detection result for the subtitle matching and detection task.

[0070] The embodiments of the present invention can effectively detect mismatches between subtitles and audio in the video to be tested, thereby enabling more comprehensive forgery detection of the video to be tested.

[0071] As a preferred embodiment, acquiring at least two target image frames of the video to be detected specifically includes: At least two key image frames are extracted from the video to be detected according to a preset extraction time interval. Based on the intensity value distribution difference between two adjacent key image frames, at least two target image frames are selected from one or more groups of adjacent key image frames according to the filtering order of the intensity value distribution difference from large to small.

[0072] Specifically, in order to avoid low efficiency and high computing power requirements due to detecting all image frames of the video to be detected, this embodiment extracts at least two key image frames from the video to be detected according to a preset extraction time interval. The extraction can start from the first image frame in the video to be detected or from other image frames. This embodiment does not make a specific limitation.

[0073] Furthermore, considering that image frames with many changes in image quality may have large differences in intensity value distribution, this embodiment selects target image frames based on the difference in intensity value distribution between two adjacent key image frames, in descending order of intensity value distribution difference. For example, when there are two target image frames, the two adjacent key image frames with the largest difference in intensity value distribution are selected as target image frames.

[0074] In an optional embodiment, a difference histogram is used to statistically analyze the intensity value distribution differences between two adjacent key image frames. Assuming I1 and I2 are the intensity values ​​of two adjacent key image frames, and their histograms are H(I1) and H(I2), respectively, the histogram H(I) is a function that counts the frequency of each intensity value in the key image frame. The difference histogram B can be calculated using the following formula: ; in, L The number of intensity values, for example, in an 8-bit image. L=256; H ( I 1)( k ) is the number of times the k-th intensity value appears in the intensity value I1 of one of the key image frames; H ( I 2)( k ) is the number of times the k-th intensity value appears in the intensity value I2 of another key image frame.

[0075] This invention filters target image frames based on the intensity value distribution difference between two adjacent key image frames. This effectively filters out image frames with more changes in the image, making it easier to capture traces of forgery. At the same time, it ensures high efficiency in forgery detection and low computational power requirements.

[0076] As a preferred embodiment, determining the forgery detection result of the video to be detected based on each of the detection results specifically includes: Based on each of the detection results and the preset weight values, the weighted confidence score corresponding to each of the detection results is obtained; Based on the weighted confidence scores and the preset total confidence score threshold, the forgery detection result of the video to be detected is determined.

[0077] Specifically, each detection result in this embodiment is a confidence score. Since each detection result corresponds to a different aspect of authenticity detection, this embodiment pre-sets corresponding weight values ​​to adjust the weight of the detection results corresponding to each authenticity detection task in the total confidence score, forming weighted confidence scores. Then, the weighted confidence scores are summed to obtain the total confidence score. When the total confidence score is greater than a preset total confidence score threshold, it indicates a high probability that the video to be detected is forged, and the forgery detection result is determined to be a forged video. Taking the authenticity detection tasks including event authenticity detection, image description information authenticity detection, lip-sync authenticity detection, and subtitle matching detection as an example, to ensure the accuracy of forgery detection, the event authenticity detection and image description information authenticity detection tasks are mandatory, while the lip-sync authenticity detection and subtitle matching detection tasks can be selected by the user according to actual needs. The total confidence score is expressed as follows: S=w1·T1+w2·T2+w3·T3+w4·T4; Where S is the total confidence score; T1 is the detection result corresponding to the image description information authenticity detection task, with a value of 0 or 1, and w1 is the weight value corresponding to the detection result T1; T2 is the detection result corresponding to the event authenticity detection task, with a value of 0 or 1, and w2 is the weight value corresponding to the detection result T2; T3 is the detection result corresponding to the lip movement authenticity detection task, with a value of 0 to 1, and w3 is the weight value corresponding to the detection result T3; T4 is the detection result corresponding to the subtitle matching detection task, with a value of 0 to 1, and w4 is the weight value corresponding to the detection result T4. The values ​​of w1 and w2 are positive numbers, and the sum of w1, w2, w3, and w4 is 1. When the lip movement authenticity detection task and the subtitle matching detection task are not selected for execution, w3 and w4 are 0.

[0078] This invention determines the forgery detection result of the video to be detected by weighting the various detection results, thereby integrating the detection results of various authenticity detection tasks, realizing forgery detection of the video to be detected from different aspects, comprehensively covering the detection of forged videos using different forgery techniques, having good generalization, and improving the accuracy of detecting various forged videos.

[0079] Please see Figure 2 A second aspect of the present invention provides a fake video detection device 100, comprising: The target image frame acquisition module 11 is used to acquire at least two target image frames of the video to be detected; wherein, the target image frames are obtained by filtering based on the differences between different image frames in the video to be detected; Image description information acquisition module 12 is used to acquire image description information of the target image frame; The authenticity detection module 13 is used to perform preset authenticity detection tasks on the target detection data of the video to be detected, and obtain the detection results corresponding to each authenticity detection task; wherein, the target detection data includes at least the image description information; The forgery detection result determination module 14 is used to determine the forgery detection result of the video to be detected based on each of the detection results.

[0080] As a preferred embodiment, the image description information acquisition module 12 is used to acquire the image description information of the target image frame, specifically including: Based on a preset face database, the target image frame is subjected to person recognition to obtain person recognition results; Based on the person recognition result and the target image frame, generate an image description instruction for the target image frame; The image description command is input into a multimodal large model for analysis to obtain the image description information.

[0081] As a preferred embodiment, the authenticity detection task includes at least an event authenticity detection task and an image description information authenticity detection task.

[0082] As a preferred embodiment, the authenticity detection module 13 is used to perform preset authenticity detection tasks on the target detection data of the video to be detected, and obtain the detection results corresponding to each authenticity detection task, specifically including: When the authenticity detection task to be executed is the authenticity detection task of the event, the large language model is used to extract event information from the image description information to obtain the target event information. Based on a preset event database, the authenticity of the target event information is detected to obtain the detection results corresponding to the event authenticity detection task.

[0083] As a preferred embodiment, the authenticity detection module 13 is used to extract event information from the image description information using a large language model to obtain target event information, specifically including: Generate an event information extraction instruction according to at least one preset event information type to be acquired and the image description information; The event information extraction command is input into the large language model to extract event information and obtain initial extracted event information. If the initial extracted event information includes event information of each of the event information types to be acquired, then the initial extracted event information is taken as the target event information; If any event information of the type of event information to be acquired is missing in the initial extracted event information, the missing event information in the initial extracted event information is supplemented using the video metadata of the video to be detected, thereby obtaining the target event information.

[0084] As a preferred embodiment, the authenticity detection module 13 is used to perform preset authenticity detection tasks on the target detection data of the video to be detected, and obtain the detection results corresponding to each authenticity detection task, specifically including: When the authenticity detection task to be executed is the authenticity detection task of the image description information, an image description information authenticity discrimination instruction is generated based on the image description information. The image description information authenticity judgment instruction is input into the large language model to perform image description information authenticity judgment, and the detection result corresponding to the image description information authenticity detection task is obtained.

[0085] As a preferred embodiment, the target detection data also includes audio and video data between two adjacent target image frames.

[0086] As a preferred embodiment, the authenticity detection task also includes a lip shape authenticity detection task; The authenticity detection module 13 is used to perform preset authenticity detection tasks on the target detection data of the video to be detected, and obtain the detection results corresponding to each authenticity detection task, specifically including: When the authenticity detection task to be performed is the lip shape authenticity detection task, each image frame to be tested and its corresponding audio data are obtained from the audio and video data. For each of the image frames to be tested and its corresponding audio data, feature extraction is performed to obtain at least one feature pair; wherein, the feature pair includes the lip shape feature vector corresponding to the image frame to be tested and the audio feature vector corresponding to the audio data; Based on the cosine similarity of each set of feature pairs, lip-sync detection is performed on each image frame to be tested and its corresponding audio data to obtain the detection result corresponding to the lip-sync authenticity detection task.

[0087] As a preferred embodiment, the authenticity detection task also includes a subtitle matching detection task; The authenticity detection module 13 is used to perform preset authenticity detection tasks on the target detection data of the video to be detected, and obtain the detection results corresponding to each authenticity detection task, specifically including: When the authenticity detection task to be performed is the subtitle matching detection task, subtitle text and audio text are extracted from the audio and video data; The subtitle text and the audio text are respectively transformed into feature vectors to obtain the subtitle text feature vector and the audio text feature vector; Based on the cosine similarity between the feature vectors of the subtitle text and the feature vectors of the audio text, the subtitle text and the audio text are matched and detected to obtain the detection result corresponding to the subtitle matching and detection task.

[0088] As a preferred embodiment, the target image frame acquisition module 11 is used to acquire at least two target image frames of the video to be detected, specifically including: At least two key image frames are extracted from the video to be detected according to a preset extraction time interval. Based on the intensity value distribution difference between two adjacent key image frames, at least two target image frames are selected from one or more groups of adjacent key image frames according to the filtering order of the intensity value distribution difference from large to small.

[0089] As a preferred embodiment, the forgery detection result determination module 14 is used to determine the forgery detection result of the video to be detected based on each of the detection results, specifically including: Based on each of the detection results and the preset weight values, the weighted confidence score corresponding to each of the detection results is obtained; Based on the weighted confidence scores and the preset total confidence score threshold, the forgery detection result of the video to be detected is determined.

[0090] The forged video detection device 100 provided in this embodiment of the invention filters out at least two target image frames from the video to be detected based on the differences between different image frames, and performs forgery detection on the video to be detected based on the image description information of the target image frames. This enables the identification of the authenticity of the content of the video to be detected, avoids the limitation of training set coverage caused by recognition based on the visual attributes of faces and images, has better generalization, and thus improves the accuracy of detecting various forged videos.

[0091] Please see Figure 3 The third aspect of the present invention provides an electronic device 200, including a memory 22, a processor 21, and a computer program stored in the memory 22 and executable on the processor 21. When the processor 21 executes the computer program, it implements the fake video detection method described in any embodiment of the first aspect.

[0092] For example, the computer program may be divided into one or more modules / units, which are stored in the memory 22 and executed by the processor 21 to complete the present invention. The one or more modules / units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in the electronic device 200.

[0093] The electronic device 200 may include, but is not limited to, a processor 21 and a memory 22. Those skilled in the art will understand that the schematic diagram is merely an example of the electronic device 200 and does not constitute a limitation on the electronic device 200. It may include more or fewer components than illustrated, or combine certain components, or different components. For example, the electronic device 200 may also include input / output devices, network access devices, buses, etc.

[0094] The processor 21 can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor, or the processor 21 can be any conventional processor 21. The processor 21 is the control center of the electronic device 200, connecting various parts of the electronic device 200 via various interfaces and lines.

[0095] The memory 22 can be used to store the computer programs and / or modules. The processor 21 implements various functions of the electronic device 200 by running or executing the computer programs and / or modules stored in the memory 22 and calling the data stored in the memory 22. The memory 22 may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the mobile phone (such as audio data, phonebook, etc.). In addition, the memory 22 may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0096] A fourth aspect of the present invention provides a computer-readable storage medium comprising a stored computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to perform the fake video detection method described in any embodiment of the first aspect.

[0097] A fifth aspect of the present invention provides a computer program product, including a computer program / instructions, which, when executed by a processor, implement the steps of the fake video detection method described in any embodiment of the first aspect.

[0098] Wherein, if the modules / units integrated in the electronic device are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc.

[0099] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A method for detecting forged videos, characterized in that, include: At least two target image frames of the video to be detected are obtained; wherein the target image frames are obtained by filtering based on the differences between different image frames in the video to be detected; Obtain the image description information of the target image frame; The target detection data of the video to be detected is subjected to preset authenticity detection tasks to obtain the detection results corresponding to each authenticity detection task; wherein, the target detection data includes at least the image description information; Based on each of the detection results, the forgery detection result of the video to be detected is determined.

2. The method for detecting fake videos as described in claim 1, characterized in that, The acquisition of the image description information of the target image frame specifically includes: Based on a preset face database, the target image frame is subjected to person recognition to obtain person recognition results; Based on the person recognition result and the target image frame, generate an image description instruction for the target image frame; The image description command is input into a multimodal large model for analysis to obtain the image description information.

3. The method for detecting fake videos as described in claim 1, characterized in that, The authenticity detection task includes at least the event authenticity detection task and the image description information authenticity detection task.

4. The method for detecting fake videos as described in claim 3, characterized in that, The step of performing preset authenticity detection tasks on the target detection data of the video to be detected, and obtaining the detection results corresponding to each authenticity detection task, specifically includes: When the authenticity detection task to be executed is the authenticity detection task of the event, the large language model is used to extract event information from the image description information to obtain the target event information. Based on a preset event database, the authenticity of the target event information is detected to obtain the detection results corresponding to the event authenticity detection task.

5. The method for detecting fake videos as described in claim 4, characterized in that, The step of extracting event information from the image description information using a large language model to obtain target event information specifically includes: Generate an event information extraction instruction according to at least one preset event information type to be acquired and the image description information; The event information extraction command is input into the large language model to extract event information and obtain initial extracted event information. If the initial extracted event information includes event information of each of the event information types to be acquired, then the initial extracted event information is taken as the target event information; If any event information of the type of event information to be acquired is missing in the initial extracted event information, the missing event information in the initial extracted event information is supplemented using the video metadata of the video to be detected, thereby obtaining the target event information.

6. The method for detecting fake videos as described in claim 3, characterized in that, The step of performing preset authenticity detection tasks on the target detection data of the video to be detected, and obtaining the detection results corresponding to each authenticity detection task, specifically includes: When the authenticity detection task to be executed is the authenticity detection task of the image description information, an image description information authenticity discrimination instruction is generated based on the image description information. The image description information authenticity judgment instruction is input into the large language model to perform image description information authenticity judgment, and the detection result corresponding to the image description information authenticity detection task is obtained.

7. The method for detecting fake videos as described in claim 3, characterized in that, The target detection data also includes audio and video data between two adjacent target image frames.

8. The method for detecting fake videos as described in claim 7, characterized in that, The authenticity detection task also includes the lip shape authenticity detection task; The step of performing preset authenticity detection tasks on the target detection data of the video to be detected, and obtaining the detection results corresponding to each authenticity detection task, specifically includes: When the authenticity detection task to be performed is the lip shape authenticity detection task, each image frame to be tested and its corresponding audio data are obtained from the audio and video data. For each of the image frames to be tested and its corresponding audio data, feature extraction is performed to obtain at least one feature pair; wherein, the feature pair includes the lip shape feature vector corresponding to the image frame to be tested and the audio feature vector corresponding to the audio data; Based on the cosine similarity of each set of feature pairs, lip-sync detection is performed on each image frame to be tested and its corresponding audio data to obtain the detection result corresponding to the lip-sync authenticity detection task.

9. The method for detecting fake videos as described in claim 7, characterized in that, The authenticity detection task also includes a subtitle matching detection task; The step of performing preset authenticity detection tasks on the target detection data of the video to be detected, and obtaining the detection results corresponding to each authenticity detection task, specifically includes: When the authenticity detection task to be performed is the subtitle matching detection task, subtitle text and audio text are extracted from the audio and video data; The subtitle text and the audio text are respectively transformed into feature vectors to obtain the subtitle text feature vector and the audio text feature vector; Based on the cosine similarity between the feature vectors of the subtitle text and the feature vectors of the audio text, the subtitle text and the audio text are matched and detected to obtain the detection result corresponding to the subtitle matching and detection task.

10. The method for detecting fake videos as described in claim 1, characterized in that, The acquisition of at least two target image frames from the video to be detected specifically includes: At least two key image frames are extracted from the video to be detected according to a preset extraction time interval. Based on the intensity value distribution difference between two adjacent key image frames, at least two target image frames are selected from one or more groups of adjacent key image frames according to the filtering order of the intensity value distribution difference from large to small.

11. The method for detecting fake videos as described in claim 1, characterized in that, The step of determining the forgery detection result of the video to be detected based on each of the detection results specifically includes: Based on each of the detection results and the preset weight values, the weighted confidence score corresponding to each of the detection results is obtained; Based on the weighted confidence scores and the preset total confidence score threshold, the forgery detection result of the video to be detected is determined.

12. A device for detecting forged videos, characterized in that, include: The target image frame acquisition module is used to acquire at least two target image frames of the video to be detected; wherein, the target image frames are obtained by filtering based on the differences between different image frames in the video to be detected; Image description information acquisition module, used to acquire image description information of the target image frame; The authenticity detection module is used to perform preset authenticity detection tasks on the target detection data of the video to be detected, and obtain the detection results corresponding to each authenticity detection task; wherein, the target detection data includes at least the image description information; The forgery detection result determination module is used to determine the forgery detection result of the video to be detected based on each of the detection results.

13. The forged video detection device as described in claim 12, characterized in that, The image description information acquisition module is used to acquire the image description information of the target image frame, specifically including: Based on a preset face database, the target image frame is subjected to person recognition to obtain person recognition results; Based on the person recognition result and the target image frame, generate an image description instruction for the target image frame; The image description command is input into a multimodal large model for analysis to obtain the image description information.

14. The forged video detection device as described in claim 12, characterized in that, The authenticity detection task includes at least the event authenticity detection task and the image description information authenticity detection task.

15. The forged video detection device as described in claim 14, characterized in that, The authenticity detection module is used to perform preset authenticity detection tasks on the target detection data of the video to be detected, and obtain the detection results corresponding to each authenticity detection task, specifically including: When the authenticity detection task to be executed is the authenticity detection task of the event, the large language model is used to extract event information from the image description information to obtain the target event information. Based on a preset event database, the authenticity of the target event information is detected to obtain the detection results corresponding to the event authenticity detection task.

16. The forged video detection device as described in claim 15, characterized in that, The authenticity detection module is used to extract event information from the image description information using a large language model to obtain target event information, specifically including: Generate an event information extraction instruction according to at least one preset event information type to be acquired and the image description information; The event information extraction command is input into the large language model to extract event information and obtain initial extracted event information. If the initial extracted event information includes event information of each of the event information types to be acquired, then the initial extracted event information is taken as the target event information; If any event information of the type of event information to be acquired is missing in the initial extracted event information, the missing event information in the initial extracted event information is supplemented using the video metadata of the video to be detected, thereby obtaining the target event information.

17. The forged video detection device as described in claim 14, characterized in that, The authenticity detection module is used to perform preset authenticity detection tasks on the target detection data of the video to be detected, and obtain the detection results corresponding to each authenticity detection task, specifically including: When the authenticity detection task to be executed is the authenticity detection task of the image description information, an image description information authenticity discrimination instruction is generated based on the image description information. The image description information authenticity judgment instruction is input into the large language model to perform image description information authenticity judgment, and the detection result corresponding to the image description information authenticity detection task is obtained.

18. The forged video detection device as described in claim 14, characterized in that, The target detection data also includes audio and video data between two adjacent target image frames.

19. The forged video detection device as described in claim 18, characterized in that, The authenticity detection task also includes the lip shape authenticity detection task; The authenticity detection module is used to perform preset authenticity detection tasks on the target detection data of the video to be detected, and obtain the detection results corresponding to each authenticity detection task. Specifically, it also includes: When the authenticity detection task to be performed is the lip shape authenticity detection task, each image frame to be tested and its corresponding audio data are obtained from the audio and video data. For each of the image frames to be tested and its corresponding audio data, feature extraction is performed to obtain at least one feature pair; wherein, the feature pair includes the lip shape feature vector corresponding to the image frame to be tested and the audio feature vector corresponding to the audio data; Based on the cosine similarity of each set of feature pairs, lip-sync detection is performed on each image frame to be tested and its corresponding audio data to obtain the detection result corresponding to the lip-sync authenticity detection task.

20. The forged video detection device as described in claim 18, characterized in that, The authenticity detection task also includes a subtitle matching detection task; The authenticity detection module is used to perform preset authenticity detection tasks on the target detection data of the video to be detected, and obtain the detection results corresponding to each authenticity detection task. Specifically, it also includes: When the authenticity detection task to be performed is the subtitle matching detection task, subtitle text and audio text are extracted from the audio and video data; The subtitle text and the audio text are respectively transformed into feature vectors to obtain the subtitle text feature vector and the audio text feature vector; Based on the cosine similarity between the feature vectors of the subtitle text and the feature vectors of the audio text, the subtitle text and the audio text are matched and detected to obtain the detection result corresponding to the subtitle matching and detection task.

21. The forged video detection device as described in claim 12, characterized in that, The target image frame acquisition module is used to acquire at least two target image frames of the video to be detected, specifically including: At least two key image frames are extracted from the video to be detected according to a preset extraction time interval. Based on the intensity value distribution difference between two adjacent key image frames, at least two target image frames are selected from one or more groups of adjacent key image frames according to the filtering order of the intensity value distribution difference from large to small.

22. The forged video detection device as described in claim 12, characterized in that, The forgery detection result determination module is used to determine the forgery detection result of the video to be detected based on each of the detection results, specifically including: Based on each of the detection results and the preset weight values, the weighted confidence score corresponding to each of the detection results is obtained; Based on the weighted confidence scores and the preset total confidence score threshold, the forgery detection result of the video to be detected is determined.

23. An electronic device, characterized in that, The device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the fake video detection method according to any one of claims 1 to 11.

24. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored computer program, wherein, when the computer program is executed, it controls the device on which the computer-readable storage medium is located to perform the fake video detection method according to any one of claims 1 to 11.

25. A computer program product, characterized in that, Includes a computer program / instruction that, when executed by a processor, implements the steps of the fake video detection method according to any one of claims 1 to 11.

Images