CN-116863366-B - Method and system for detecting cross-sample false news video

Abstract

The invention provides a detection method of cross-sample false news videos, which comprises the steps of obtaining news videos and refute a rumour videos of target events, extracting single sample characteristics of the news videos as first characteristics, extracting single sample characteristics of the refute a rumour videos as second characteristics, constructing an event diagram taking the first characteristics as nodes and importance between the nodes as edges, obtaining node characteristics of the nodes through information aggregation, selecting news videos corresponding to the node characteristics classified as true as the videos to be detected through true and false classification of the node characteristics, and selecting second false videos in the videos to be detected through detection of fact information conflict between the second characteristics and the first characteristics of the videos to be detected, and taking the first false videos and the second false videos as the news videos. The invention also provides a detection system of the cross-sample false news video and a data processing device for realizing the detection of the cross-sample false news video.

Inventors

• CAO JUAN
• QI PENG
• TANG SHENG

Assignees

• 中国科学院计算技术研究所

Dates

Publication Date

20260505

Application Date

20230605

Claims (8)

1. 1. The method for detecting the cross-sample false news video is characterized by comprising the following steps of: Acquiring a news video and refute a rumour video of a target event, extracting single sample characteristics of the news video as first characteristics, and extracting single sample characteristics of the refute a rumour video as second characteristics; Acquiring node characteristics of the node by information aggregation, and selecting news videos corresponding to the node characteristics classified as true as videos to be detected and news videos corresponding to the node characteristics classified as false as first false videos by carrying out true and false classification on the node characteristics; selecting a second false video in the video under test by detecting a factual information conflict between the second feature and a first feature of the video under test, including comparing text features in the second feature And text features in the first feature Generating feature pairs Modeling is carried out through a BERT model, and text conflict characteristics are obtained Adding temporal position coding in the key frame feature set in the second feature Classification marking Obtaining the key frame feature set after refute a rumour video processing Adding temporal position coding in the key frame feature set in the first feature Classification marking Obtaining the key frame feature set after the video processing to be detected Enhancement with stacked self-attention and cross-attention modules And And will And (3) with Vector stitching to obtain vision consistency characteristics By self-focusing layer pairs And Dynamic fusion is carried out, and the fusion characteristics are subjected to two classification to obtain the probability that the video to be detected is the second false video Predicting probability of the news video as false news video , And taking the first false video and the second false video as false news videos of the target event.
2. 2. The method of detecting cross-sample false news video of claim 1, wherein the event map is organized in the form of a graph-annotating force neural network Edges (V) Obtained through the attention mechanism, representing the node Is a first feature pair node of (a) Is of importance to the first feature of (a).
3. 3. The method for detecting cross-sample false news video according to claim 2, wherein the node Node characteristics of (2) : Wherein, the , , Is that Is set of the neighbor node set of (c), , Is that And (3) with The weight of the two-way valve is equal to the weight of the two-way valve, 、 In order for the parameters to be trainable, The operation of the splice is indicated and, Is a nonlinear operation.
4. 4. The method for detecting cross-sample false news video according to claim 3, wherein the method is characterized by node pairs Loss function for true and false classification Wherein Representing the actual tag of the news video.
5. 5. The method of claim 1, wherein the first feature and the second feature are extracted using an SV-fed single sample detector.
6. 6. A system for detecting cross-sample false news video, comprising: the feature extraction module is used for acquiring a news video and refute a rumour video of a target event, extracting single sample features of the news video as first features, and extracting single sample features of the refute a rumour video as second features; The image aggregation module is used for constructing an event graph taking the first feature as a node and taking importance among the nodes as an edge, acquiring the node feature of the node by information aggregation, and selecting a news video corresponding to the node feature classified as true as a video to be detected and a news video corresponding to the node feature classified as false as a first false video by carrying out true and false classification on the node feature; The sample correction module is used for selecting a second false video in the video to be detected by detecting the fact information conflict between the second feature and the first feature of the video to be detected, and taking the first false video and the second false video as false news videos of the target event, and comprises a text conflict detection module, an attention fusion and classification module and an attention fusion and classification module, wherein, The text conflict detection module is used for acquiring text conflict characteristics and comparing the text characteristics in the second characteristics And text features in the first feature Generating feature pairs Modeling is carried out through a BERT model, and text conflict characteristics are obtained ; The visual consistency evaluation module is used for acquiring visual consistency characteristics, and adding time position codes in the key frame characteristic set in the second characteristics Classification marking Obtaining the key frame feature set after refute a rumour video processing Adding temporal position coding in the key frame feature set in the first feature Classification marking Obtaining the key frame feature set after the video processing to be detected Enhancement with stacked self-attention and cross-attention modules And And will And (3) with Vector stitching to obtain vision consistency characteristics ; The attention fusion and classification module is used for passing the self-attention layer pair And Dynamic fusion is carried out, and the fusion characteristics are subjected to two classification to obtain the probability that the video to be detected is the second false video Predicting probability of the news video as false news video , To predict the probability that the news video is the first false video.
7. 7. A computer readable storage medium storing computer executable instructions which, when executed, implement the cross-sample false news video detection of any one of claims 1 to 5.
8. 8. A data processing apparatus comprising the computer readable storage medium of claim 7, which when fetched and executed by a processor of the data processing apparatus, performs detection of cross-sample false news videos.

Description

Method and system for detecting cross-sample false news video Technical Field The invention relates to the technical field of news credibility authentication, in particular to a cross-sample false news video detection method and system. Background In recent years, short video platforms such as tremble, fast-handedness and the like have spawned a large number of false news videos. Compared with traditional false news based on text and graphics, the false news in the video form is more attractive and convincing, which makes false news video detection an emerging research point in multi-modal false news detection tasks. False news videos are typically composed of news headlines and videos together. The goal of the false news video detection task is to give a true or false classification decision to the incoming news video. Most of the existing work focuses on how to make full use of multi-modal information in a single sample for classification. As one of the most representative works, qi et al propose the detection model SV-FEND (short video false news detection model) with the best detection performance at present, as shown in fig. 1. The model extracts multi-mode features such as titles, subtitles, audios, key frames, video clips, comments, user images and the like, uses two cross-mode transformers to model the association between text, audios and key frame features, and finally uses the transformers to fuse all the features for classification. Although the existing methods have been more fully utilized for multimodal information in a single sample, detection cues presented in a single video tend not to be obvious due to careful tampering by counterfeiters, which limits the performance of existing single sample-based detection methods. For example, partial false news videos only modify news elements such as time, place, etc. in the headline of real news videos, and it is difficult for a single-sample multi-modal content-based detector to successfully detect such false news videos. Disclosure of Invention To achieve effective automatic detection of false news videos. Aiming at the technical problem of limited single sample clues in the prior art, the invention provides a cross-sample false news video detection method which comprises the steps of obtaining news videos and refute a rumour videos of a target event, extracting single sample characteristics of the news videos as first characteristics, extracting single sample characteristics of the refute a rumour videos as second characteristics, constructing an event diagram taking the first characteristics as nodes and importance between the nodes as edges, obtaining node characteristics of the nodes through information aggregation, selecting news videos corresponding to the node characteristics classified as true as the videos to be detected by carrying out true and false classification on the node characteristics, and selecting news videos corresponding to the node characteristics classified as false as the first false videos by detecting fact information conflict between the second characteristics and the first characteristics of the videos to be detected, and selecting second false videos in the videos to be detected and taking the first false videos and the second false videos as the false news videos of the target event. The invention discloses a method for detecting cross-sample false news video, wherein the step of detecting fact information conflict between a second feature and a first feature of the video to be detected comprises the steps of generating a feature pair < S D,SC > by text feature S D in the second feature and text feature S C in the first feature, modeling by a BERT model to obtain text conflict feature x t＝BERT([CLS]SD [SEP] SC [ SEP ]), and adding time position codes f tem and classification marks in a key frame feature set in the second featureObtaining the key frame feature set after refute a rumour video processingAdding temporal position coding ftem and classification labels in a keyframe feature set in the first featureObtaining the key frame feature set after the video processing to be detectedEnhancement with stacked self-attention and cross-attention modulesAndAnd will beAnd (3) withVector stitching to obtain vision consistency characteristicsDynamically fusing x t and x v through a self-attention layer, and performing two classification on fusion characteristics to obtain probability that the video to be detected is a second false videoPredicting the probability that the news video is a false news videoTo predict the probability that the news video is the first false video. The invention relates to a detection method of cross-sample false news video, wherein the event map is organized in the form of a graph-annotating force neural networkEdge e ij is derived by the attention mechanism and represents the importance of the first feature of node v j to the first feature of node v i. The invention relates to a method for detectin