False news intelligent detection system based on multi-dimensional fusion

Abstract

The invention belongs to the technical field of artificial intelligence information processing, and particularly relates to a false news intelligent detection system based on multi-dimensional fusion. The system comprises a data acquisition and processing module used for acquiring and preprocessing news data; the rapid detection module is used for carrying out true and false binary classification rapid detection on the news by adopting an ERNIE 3.0 model; the multi-dimensional slow auditing module is used for carrying out deep analysis from six dimensions of news source confidence evaluation, text and question consistency detection, AI generation content detection, multi-modal image-text consistency analysis, user propagation confidence evaluation and logic judgment based on a large language model; and the multi-agent collaborative analysis module constructs seven specialized agents to perform parallel analysis and collaborative decision, and outputs a four-level classification conclusion. According to the method, a rapid detection and slow auditing double-layer detection framework is constructed, and a multi-agent cooperation mechanism is combined, so that all-around and high-precision detection of false news is realized, the threat of AI generation content is effectively dealt with, and the media quality of a user is improved.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence information processing technology, specifically relating to a fake news intelligent detection system based on multi-dimensional fusion. Background Technology

[0002] In recent years, the explosive development of large-scale model technology has brought revolutionary breakthroughs to the field of artificial intelligence, but it has also provided fertile ground for the generation and spread of fake news. The powerful text generation capabilities of these large models make it easier to create high-quality, indistinguishable fake content, providing unprecedented technological support for the generation and spread of fake news. Therefore, the innovation and exploration of fake news detection systems cannot be separated from the application of artificial intelligence technology.

[0003] Among existing fake news detection technologies, natural language processing (NLP) encompasses multiple dimensions such as text segmentation, word vector representation, semantic analysis, and sentiment recognition. It can deeply understand the internal structure and semantic features of fake news texts, extracting key language patterns and expressive characteristics that distinguish between genuine and fake news. Machine learning technologies, including supervised learning, unsupervised learning, and deep learning, can automatically learn the deep features of fake news texts. Multimodal technologies break through the limitations of single-text analysis, enabling comprehensive analysis of various information carriers such as text, images, audio, and video, thereby capturing the multi-dimensional feature differences of fake news.

[0004] However, current fake news detection systems have the following technical problems: First, the platform lacks multi-dimensional automated analysis capabilities. Most existing platforms rely on single dimensions or manual review, lacking the ability to comprehensively assess the authenticity of news from multiple perspectives and effectively handle complex forms of misinformation.

[0005] Second, there is a lack of AI-generated content detection capabilities. With the development of large-scale model technology, AI-generated fake news is becoming increasingly difficult to identify, but most platforms have not yet developed dedicated detection mechanisms for this type of content.

[0006] Third, the multimodal analysis capabilities are weak. Existing platforms mainly focus on text analysis and lack effective consistency detection for multimodal content that combines text and images, which has a greater impact on dissemination.

[0007] Fourth, its intelligent interaction and decision-making support capabilities are limited. There is a lack of application of advanced AI technologies, particularly the insufficient use of large models and multimodal analysis in fake news identification.

[0008] Fifth, user engagement and educational functions are insufficient. Existing platforms mostly provide information in a one-way manner, lacking community interaction and user education functions, thus failing to improve users' media literacy and judgment. Summary of the Invention

[0009] The purpose of this invention is to address the shortcomings of the aforementioned background technology and provide a multi-dimensional fusion-based intelligent fake news detection system. By constructing a dynamic multimodal joint detection system, it achieves comprehensive and high-precision detection of fake news.

[0010] The technical solution adopted in this invention is: a fake news intelligent detection system based on multi-dimensional fusion, including... The data acquisition and processing module is used to collect news data from multiple news platforms and perform preprocessing. The news data includes news titles, news text, news images, news videos, news audio, publishing platform information, dissemination time records, dissemination geographical location data, and dissemination user profile data. The fast detection module uses the ERNIE 3.0 model, which is fine-tuned based on a three-level architecture of word segmentation-enhancement-pre-training, to perform fast binary classification detection of true and false news data and output the true and false binary classification results. The true and false binary classification results include the classification label of true or false and the corresponding confidence score. The multi-dimensional slow review module performs in-depth analysis on news data after rapid detection of true and false information from six evaluation dimensions. The six evaluation dimensions include news source confidence assessment, title consistency detection, AI-generated content detection, multimodal image and text consistency analysis, user dissemination confidence assessment, and logical judgment based on a large language model. A comprehensive trust score is calculated through a weighted fusion algorithm. The multi-agent collaborative analysis module constructs seven agents, including a title analysis expert, a syntax verification expert, a common sense verification expert, a credibility assessment expert, a multi-source retrieval expert, a user attribute analysis expert, and a consistency detection expert. These seven agents perform parallel analysis and collaborative decision-making on the output results of the fast detection module and the multi-dimensional slow review module, and output the final detection conclusion.

[0011] Preferably, the preprocessing includes generating a summary using the TextRank algorithm and extracting keywords using the TF-IDF algorithm.

[0012] Preferably, the news source confidence assessment uses the publishing platform information collected by the data acquisition and processing module as input data. The publishing platform information includes the platform URL and platform identifier, including: Construct a news platform reputation database, and classify platforms into five levels based on the platform identifier: high authority, relatively authoritative, neutral, relatively niche, and niche, and output the platform authority level identifier; The basic reputation score is calculated and output based on the official background, review mechanism, historical accuracy and user base of the platform corresponding to the published platform information. The credibility score is dynamically adjusted based on the platform's recent rate of fake news publication, correction rate, and complaint rate, and a news source confidence score is output.

[0013] Preferably, the title consistency detection includes: The pre-trained text2vec-base-chinese model is used to convert the news headlines and news text into high-dimensional vector representations, calculate the cosine similarity between the two vectors, and output the semantic consistency score. The TF-IDF algorithm is used to extract keywords from the news headlines, calculate the keyword coverage and weight matching degree in the news body, and output the keyword coverage and weight relevance scores. The final title-text consistency score is calculated based on semantic consistency score, keyword coverage, and weighted relevance score.

[0014] Preferably, the AI-generated content detection includes: Using the news text as input, a Transformer-based bidirectional encoder structure is adopted. By analyzing the text's perplexity, suddenness, and lexical diversity indicators, the AI-generated text content is identified, and the text AI generation probability score is output. Using the news image as input, a convolutional neural network is used to extract the texture features, color distribution features, and edge detail features of the image. A support vector machine classifier is used to determine whether the image is generated by AI, and the AI ​​generation probability score of the image is output. Using the news video as input, the inter-frame difference features, motion vector features and temporal consistency features of the video are extracted. A recurrent neural network is used to perform AI-generated video recognition and output the video AI generation probability score. Using the news audio as input, the Mel frequency cepstral coefficients, fundamental frequency profile, and formant features of the audio are analyzed. A deep neural network is used to identify AI-synthesized audio and output an audio AI generation probability score. The AI-generated content detection system outputs a comprehensive detection score based on the detection results of news-style text, images, videos, and audio.

[0015] Preferably, the multimodal image-text consistency analysis includes: Using the news article as input, the BERT model is used to extract the semantic features of the text and output a 768-dimensional text feature vector. Using the news image as input, the ResNet-50 model is used to extract the visual features of the image and output a 2048-dimensional image feature vector. Calculate the correlation matrix between the text feature vector and the image feature vector, find consistent features in high correlation regions, discover inconsistencies in low correlation regions, and output consistency scores and inconsistency scores. Cross-modal attention features are concatenated with the original features, and a multimodal image-text consistency score is output through a fully connected layer.

[0016] Preferably, the user propagation confidence assessment includes: Using the aforementioned dissemination time record as input, calculate the time span and dissemination speed of the news dissemination, and output a timeline dimension score; Using the aforementioned geographical location data as input, the dissemination coverage and geographical diversity of news in different provinces are evaluated, and a geographical distribution dimension score is output. Using the aforementioned user profile data as input, the age distribution and geographical diversity of the users participating in the dissemination are analyzed, and a user diversity score is output. The user propagation confidence assessment outputs a user propagation confidence score by weightedly fusing the timeline dimension score, the geographical distribution dimension score, and the user diversity score.

[0017] Preferably, the logical judgment based on the large language model uses the DeepSeek model for three-dimensional logical evaluation: The news article is analyzed to detect whether there are self-contradictory factual descriptions, timeline conflicts, and inconsistencies in details, and a content logic consistency score is output. The events described in the news article are analyzed to assess whether the causal chain of the events conforms to common sense and basic logic, to determine whether the strength of the causal relationship is consistent with the actual impact, and to output a causal relationship rationality score. The narrative style of the news article is analyzed, words with emotional bias are identified in the text, the balance of different viewpoints and whether leading language is used are evaluated, and an objectivity score of the narrative technique is output. The three-dimensional logical evaluation outputs a logical quality score by averaging the scores for logical consistency of content, rationality of causal relationships, and objectivity of narrative techniques.

[0018] Preferably, the multi-agent collaborative analysis is implemented using the Swarm framework, including: The seven agents simultaneously and independently analyze the true / false binary classification results and the six-dimensional evaluation results, with each agent outputting the evaluation result and confidence level for a specific dimension. The analysis results of each agent are integrated through a weighted voting mechanism, with the weights dynamically adjusted based on historical accuracy, and a weighted comprehensive score is output. When the disagreement among the agents exceeds a threshold, a negotiation mechanism is initiated to reach a consensus through evidence exchange and reasoning verification, and output a four-level classification conclusion, including fake news, suspicious content, possibly true and credible content. During the knowledge update phase, the verified new knowledge is structured and stored in the knowledge base to optimize subsequent analysis capabilities and output updated knowledge base data.

[0019] Preferably, it also includes a user interaction and education module, which takes detection process data, supplementary evidence submitted by users, user error judgment records, and user contribution records as input data: The evidence chain visualization unit takes the detection process data as input, and intuitively displays the detection process and judgment basis through semantic traceability diagram and logic heat map, and outputs a visualization display interface; The community collaborative verification unit takes the supplementary evidence submitted by the user as input, allows the user to submit supplementary evidence, and after the administrator reviews it, it is included in the evidence library and outputs the verified supplementary evidence to the evidence chain dashboard. The media literacy education unit takes the user's incorrect judgment records as input, generates personalized learning paths and test questions based on the user's incorrect judgment records, and outputs personalized learning paths and test questions. The contribution incentive unit uses the user contribution records as input to establish a user contribution ranking list, rewards high-contribution users with points and upgrades, and outputs user points and level data.

[0020] Compared with the prior art, the beneficial effects of the present invention are: This invention constructs a two-layer detection architecture that combines rapid detection with slow review, which ensures both detection efficiency and detection accuracy. Rapid detection can achieve a detection accuracy of 95% within 2 seconds and clearly outputs true and false binary classification results and confidence scores.

[0021] This invention adopts a six-dimensional comprehensive evaluation system, which conducts in-depth analysis from multiple perspectives such as news source, text-title consistency, AI-generated content, image-text consistency, user dissemination characteristics, and logical consistency. Each dimension module clearly defines the input data and output results, making the detection results more comprehensive, reliable, and traceable.

[0022] This invention enables comprehensive detection of all major AI-generated content types, including text, images, audio, and video. Each detector outputs an AI generation probability score for its corresponding media type, and the final output is a comprehensive detection score for AI-generated content, effectively addressing the new threat of fake content brought about by large model technology.

[0023] This invention employs a cross-modal feature fusion network (CFFN) for image-text consistency analysis. The text encoder outputs a 768-dimensional text feature vector, the image encoder outputs a 2048-dimensional image feature vector, and the cross-modal attention mechanism outputs consistency and inconsistency scores. This effectively identifies common forgery techniques that combine real images with fake text.

[0024] This invention implements a seven-agent collaborative detection mechanism through the Swarm framework. Each agent outputs evaluation results and confidence levels in specific dimensions. A weighted comprehensive score is output through a weighted voting mechanism, and a four-level classification conclusion is output during the negotiation and decision-making stage. This not only improves the detection accuracy but also enhances the interpretability of the results.

[0025] This invention constructs a collaborative governance ecosystem of detection, interpretation, and education. The evidence chain visualization unit outputs a visual display interface, the community collaborative verification unit outputs supplementary evidence after verification, the media literacy education unit outputs personalized learning paths and test questions, and the contribution incentive unit outputs user points and level data, thus realizing a closed-loop governance from technological barriers to cognitive improvement. Attached Figure Description

[0026] Figure 1 A block diagram of a multi-dimensional fusion-based intelligent fake news detection system provided in an embodiment of this application; Figure 2 This is a flowchart of a data acquisition crawler in one embodiment of this application; Figure 3 This is a flowchart of the TextRank summary generation algorithm in one embodiment of this application; Figure 4 This is a flowchart of the TF-IDF keyword extraction algorithm in one embodiment of this application; Figure 5 This is a diagram of the fast detection module architecture in one embodiment of this application; Figure 6 This is a weight distribution diagram of the six dimensions of evaluation in one embodiment of this application; Figure 7 This is a flowchart of a news source confidence assessment algorithm in one embodiment of this application; Figure 8 This is a flowchart of a Chinese title consistency detection algorithm in one embodiment of this application; Figure 9 This is a flowchart of AI-generated text detection in one embodiment of this application; Figure 10 This is a diagram of a user propagation confidence evaluation architecture in one embodiment of this application; Figure 11 This is a flowchart of a three-dimensional logic evaluation algorithm in one embodiment of this application; Figure 12This is a flowchart of the Swarm framework for multi-agent collaborative analysis in one embodiment of this application; Figure 13 The flowchart of the intelligent fake news detection system based on multi-dimensional fusion provided in one embodiment of this application is shown. Detailed Implementation

[0027] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings. It should be noted that these descriptions are for the purpose of aiding understanding the present invention, but do not constitute a limitation thereof. Furthermore, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0028] like Figure 1 , Figure 13 As shown in the figure, this application provides a fake news intelligent detection system based on multi-dimensional fusion, which adopts a four-layer architecture design of data acquisition and processing, fast detection, slow review and multi-agent analysis.

[0029] The data acquisition and processing module is located at the bottom layer of the system and is responsible for collecting news data from multiple news platforms and preprocessing it. The news data includes news titles, news text, news images, news videos, news audio, publishing platform information (including platform URLs and platform identifiers), dissemination time records, dissemination geographic location data, and dissemination user profile data.

[0030] The fast detection module is located in the second layer. It adopts the ERNIE3.0 model, which is finely tuned based on a three-level architecture of word segmentation-enhancement-pre-training. It performs fast binary classification detection of real and fake news data, achieving a performance index of 95% detection accuracy within 2 seconds, and outputs the binary classification results of real or fake, including real or fake classification labels and corresponding confidence scores.

[0031] The multi-dimensional slow review module is located in the third layer. It conducts in-depth analysis of news data from six evaluation dimensions, including news source confidence assessment (38% weight), multimodal image and text consistency analysis (37% weight), logical judgment (15% weight), text and title consistency detection (5% weight), user dissemination confidence assessment (5% weight), and AI-generated content detection (5% weight). It calculates the comprehensive trust score through a weighted fusion algorithm and outputs the six-dimensional evaluation results and the comprehensive trust score.

[0032] The multi-agent collaborative analysis module is located at the top level. It uses the Swarm framework to build seven specialized agents. It takes the true / false binary classification results output by the fast detection module and the six-dimensional evaluation results output by the multi-dimensional slow review module as input data, performs parallel analysis and collaborative decision-making, and outputs the final detection conclusion, which includes a four-level classification conclusion (fake news, suspicious content, possibly true, credible content) and a comprehensive confidence score.

[0033] In one embodiment, the data acquisition and processing module is the basic support module of the system in this embodiment, which mainly completes two major tasks: data acquisition and data preprocessing.

[0034] The data sources consist of three parts: First, platform collection, establishing a news source database covering different levels of authority from high-authority platforms such as Platform A, Platform B, and Platform C to local small news websites; second, open-source dataset collection, integrating public datasets such as Chinese Rumor Dataset, DoubleCheck, CHECKED, CrossFake, and CHEF to construct a large-scale fake news training database; third, platform data crawling, using web crawling technology to crawl the latest news data from platforms such as Platform D, Platform E, Platform F, Platform G, Platform H, Platform I, and Platform K in real time. In this embodiment, "rumors" refers to "fake news".

[0035] In this application, the platform collects information on the source of rumors, which is a crucial dimension in determining the veracity of rumors. In today's information-saturated society, the reliability of the source of a rumor directly impacts the veracity of its content. Some official platforms have their own review mechanisms, which can greatly verify the authenticity of rumors; however, smaller platforms have weak review mechanisms, or may even lack them altogether, significantly increasing the probability of rumors appearing. To design a rumor source score, we first collected information from major rumor-publishing platforms across the entire internet and ranked them according to their frequency and authority, as shown in Table 1 below.

[0036] Table 1

[0037] Open source data collection: In order to build a large-scale rumor database, a large number of publicly available Chinese rumor detection datasets were collected, as shown in Table 2.

[0038] Table 2

[0039] like Figure 2As shown, data collection is achieved through web crawling technology. The collected news data includes news titles, news articles, news images, news videos, news audio, publishing platform information (including platform URLs and platform identifiers), dissemination time records, dissemination geographic location data, and dissemination user profile data. Taking a certain news platform as an example, the crawling process is divided into three stages: The first stage is the task initiation and page loading stage, where the system first initializes the browser and constructs the search URL, and then loads the target page; the second stage is the data crawling and content parsing stage, where the system extracts the news title and original link of the current page, accesses the original address through a new tab, and crawls the news article content and news image resources; the third stage is the loop control and task termination stage, where all search results are processed through a loop crawling mechanism.

[0040] Specifically, data crawling from rumor platforms: In addition to rumor data provided by enterprises, common platforms (Platform D, Platform E, Platform F, Platform G, Platform H, Platform I, Platform K, Platform L, Platform M) also provide real-time rumor data. This embodiment uses web crawling technology to crawl the latest rumor data from these platforms. The crawling process for Platform G is divided into three stages.

[0041] Phase 1: Task Initiation and Page Loading The process is triggered by "Start Request". The system first initializes the browser and constructs the search URL, then loads the target page. After the page opens, it automatically checks if there are any search results. If there are no results, the data is saved and the task ends. If there are results, it proceeds to the next stage.

[0042] Phase Two: Data Scraping and Content Analysis The system extracts the title and original link of the current page, accesses the original address through a new tab, and retrieves the main content and image resources. The text content is cleaned and saved as a TXT file, while image information is stored separately (including URLs or local downloads). After processing a single data entry, the system returns to the search list page to prepare for processing the next page.

[0043] Phase 3: Loop Control and Task Termination The crawling process continues in a loop, starting with the "Continue processing next page" step, until all search results have been processed. Finally, the browser instance is closed, the task is marked "Task complete," and all crawled data is output, terminating the process.

[0044] like Figure 3As shown, TextRank summary generation is based on a graph ranking algorithm. Using raw news text consisting of collected news titles and body text as input, it automatically extracts the most representative sentences from the news text to form a summary. The algorithm steps are as follows: Step 1: Clean the raw news text, removing HTML tags, special characters, and extra spaces, and output the cleaned text; Step 2: Segment the text into sentence units using punctuation marks, and output a list of sentence units; Step 3: Calculate the semantic similarity between each pair of sentences, construct a sentence similarity graph, and output a sentence similarity matrix; Step 4: Apply the PageRank algorithm iteratively to calculate the importance score of each sentence, and output a list of sentence importance scores; Step 5: Select the top N sentences with the highest scores, rearrange them in the original text order to generate a summary, and output the news summary text.

[0045] like Figure 4 As shown, TF-IDF keyword extraction uses the news text as input data to identify and extract the most representative and distinctive keywords from the news text. The algorithm steps are as follows: Step 1, text preprocessing: Clean the news text, removing HTML tags, special characters, numbers, and other interfering elements, and then use the jieba word segmentation tool to segment the text into words or phrases; Step 2, stop word filtering: Remove common stop words that do not contribute much to the text's theme; Step 3, term frequency (TF) calculation; Step 4, inverse document frequency (IDF) calculation; Step 5, TF-IDF value calculation; Step 6, keyword extraction: Sort the words in descending order according to their TF-IDF values, select the top N words with the highest scores, and output a list of text keywords.

[0046] In one embodiment, such as Figure 5As shown, the fast detection module employs a finely tuned ERNIE 3.0 model based on a three-level architecture of word segmentation, augmentation, and pre-training to achieve rapid binary classification detection of news data for true and false information. The first level is the word segmentation stage, which achieves cross-platform compatibility through a standardized file format (.csv dataset input, safetensors model output), and performs special symbol cleaning and stop word filtering on Chinese news text to reduce data noise. The second level is the data augmentation stage, which uses adversarial data augmentation techniques to expand the training data through synonym replacement, random insertion, and random deletion, thereby breaking down the superficial plausibility of false information and improving the model's generalization ability. The third level is the pre-training and fine-tuning stage, which, based on the Baidu Wenxin Yiyan ERNIE 3.0 pre-trained model, leverages its deep semantic understanding capabilities to fine-tune the fake news detection task, forming a detection system that combines feature purification and generalization capabilities. This three-tier architecture effectively addresses the complexity of Chinese news, achieving a 95% detection accuracy rate within 2 seconds and outputting true / false binary classification results, including true or false classification labels and corresponding confidence scores, providing a rapid screening foundation for subsequent in-depth analysis.

[0047] In one embodiment, such as Figure 6 As shown, the multi-dimensional slow review module performs in-depth analysis of news data from six evaluation dimensions and calculates a comprehensive trust score through a weighted fusion algorithm. The final comprehensive trust score calculation formula is as follows: Score total =Score source ×0.38+Score image-text ×0.37+Score logic ×0.15+Score consistency ×0.05+Score propagation ×0.05+Score AI ×0.05. Where, Score source Score the confidence level of the news source (input data is the publishing platform information, output is the confidence level of the news source). image-text To score the consistency between text and images (input data consists of news text and accompanying images; output is a multimodal text-image consistency score), Score logic Score for logical reasoning (input data is news text, output is logical reasoning quality score). consistency Score the consistency between title and text (input data is the news headline and the news body, output is the consistency score). propagation Score the user's confidence score in spreading the message (input data includes spread time records, spread geolocation data, and spread user profile data; output is the user's confidence score in spreading the message). AIGenerate content detection scores for AI (input data includes news text, news images, news videos, and news audio; output is a comprehensive content detection score generated by AI).

[0048] like Figure 7 As shown, the news source confidence assessment uses the publishing platform information collected by the data collection and processing module as input data. The publishing platform information includes the platform URL and platform identifier, which is achieved by constructing a news platform reputation database. The specific implementation includes: First, platform classification, dividing news platforms into five levels based on the platform identifier: high-authority platforms (such as national media such as Platform A, Platform B, Platform C, Platform N, Platform F), relatively authoritative platforms (such as comprehensive or vertical authoritative platforms such as Platform D, Platform H, Platform O), neutral platforms (such as large portal websites such as Platform P, Platform Q, Platform R), relatively niche platforms (such as vertical platforms such as Platform S, Platform T, Platform U), and niche platforms (such as local small news websites and unregistered self-media platforms), outputting the platform authority level identifier; Second, basic reputation score calculation, calculating a basic reputation score based on the official background, review mechanism, historical accuracy rate, and user base of the platform corresponding to the published platform information, outputting the basic reputation score; Third, dynamic adjustment, dynamically adjusting the reputation score based on the recent false news publication rate, correction rate, and complaint rate of the platform corresponding to the published platform information, ensuring that the evaluation results reflect the real-time status of the platform, outputting a news source confidence score, the score being normalized and ranging from 0 to 1.

[0049] Specifically, this includes: Let N be the total number of media outlets reporting a certain news story, and i represent the i-th media source (i=1,2,...,N). For the final confidence level, The authority level is weighted into a score to assess the credibility of "who is reporting this news." If a news story is reported by both Platform A (high authority) and several self-media platforms (niche), according to this formula, the overall authority score will be significantly boosted by Platform A's participation, reflecting the reality that "one authoritative media outlet is better than multiple ordinary media outlets." The source consistency score is used to assess whether the content reported from different sources is consistent. The coverage breadth coefficient is used to assess "how many different types of media are reporting on this news," reflecting the breadth and diversity of information dissemination.

[0050] The final confidence level calculation formula is as follows: ; Define the set of media authority levels This corresponds to five levels of authority: high authority, relatively authoritative, neutral, less mainstream, and niche. For media outlet i, its authority level is... The corresponding weight is The weight value corresponding to each level is shown in the formula: ; The weighted score based on authority level is shown in the formula: ; Where α is the adjustment coefficient, as shown in the formula: ; This formula indicates that when at least one equal When α=1.2, when there is no equal But there are at least two equal When α = 1.1, α = 1.0 in all other cases.

[0051] The exponential function I represents a value of 1 when the condition is true, and a value of 0 otherwise.

[0052] in, l i This indicates the authority level of the i-th media outlet. l k This represents the k-th level of authority (k=1,2,3,4,5 correspond to high authority, relatively high authority, neutral, less popular, and less popular, respectively).

[0053] Let K be the total number of key information points extracted from all reports. This represents the number of conflicting information points. The degree of inconsistency is defined. As shown in the formula: ; Where β is the adjustment coefficient for highly authoritative media, as shown in the formula: ; Source consistency score As shown in the formula: ; Let C be the total number of media categories (in this system, C=5). This represents the number of different media categories mentioned in the report. Define the report breadth coefficient. As shown in the formula: ; Final confidence level after considering boundary conditions As shown in the formula: ; condition This indicates that the conditions are met. The number of elements in the sequence is exactly equal to N, meaning that all elements are in the fourth and fifth authority ranks. Pick The smaller of 0.5 and 0.

[0054] To ensure the final score is within the [0,1] interval, the final score is further normalized, as shown in the formula: .

[0055] like Figure 8 As shown, the title-text consistency detection uses the news headlines and news text collected by the data acquisition and processing module as input data. It employs the text2vec-base-chinese model to evaluate the consistency between the headline and the text text from two dimensions: semantic similarity and key information matching. The semantic similarity calculation unit uses the pre-trained text2vec-base-chinese model to convert the news headlines and news text into high-dimensional vector representations, calculates the cosine similarity between the two vectors, and outputs a semantic consistency score. The key information matching unit uses the TF-IDF algorithm to extract keywords from the news headlines, calculates the keyword coverage and weight matching degree in the news text, and outputs keyword coverage and weight relevance scores. The comprehensive scoring unit calculates the final title-text consistency score using the following formula.

[0056] In application, after obtaining the vectors of the title and body text, the consistency between the title and body text is evaluated from two dimensions: semantic similarity and consistency of key information. Let... For title word sequence, If the text vector is the same as the title vector, then the vector representations of the title and text are calculated using the following formula: ; Semantic similarity is obtained by calculating the cosine similarity between the title and body text vectors, as shown in the formula: ; set up For the title keyword set, This is the set of keywords for the main text. The formula for calculating the semantic similarity between keywords in the title and words in the main text is as follows: ; in, , This metric assesses the presence of key concepts in the title within the body text by calculating the semantic similarity between each title keyword and the most similar word in the body text.

[0057] Key information consistency is determined by calculating the overlap ratio between title keywords and body text keywords, as shown in the formula: ; The scores from these two dimensions are weighted and combined to obtain the final consistency score, as shown in the formula: ; Where α is the semantic similarity weight coefficient and β is the keyword matching weight coefficient, satisfying α + β = 1. In this embodiment, α = 0.6 and β = 0.4, meaning that semantic similarity accounts for 60% of the final score, and keyword matching accounts for 40%. This weight allocation reflects that semantic consistency is a better indicator of the true relevance between the title and the body text than superficial keyword matching.

[0058] The title-content consistency score ranges from 0 to 1. A higher score indicates better consistency between the title and content; conversely, a lower score indicates that the title may be exaggerated, misleading, or inconsistent with the content.

[0059] like Figure 9 As shown, the AI-generated content detection uses the news text, news images, news videos, and news audio collected by the data acquisition and processing module as input data to achieve full coverage detection of four media types: text, images, videos, and audio. The text detector takes the news text as input, employs a Transformer-based bidirectional encoder structure, and identifies AI-generated text by analyzing three core indicators: perplexity, suddenness, and lexical diversity, outputting a text AI generation probability score. The image detector takes the news image as input, uses a convolutional neural network to extract texture features, color distribution features, and edge detail features of the image, and uses a support vector machine classifier to determine whether the image is AI-generated, outputting an image AI generation probability score. The video detector takes the news video as input, extracts inter-frame difference features, motion vector features, and temporal consistency features of the video, uses a recurrent neural network to identify AI-generated video, outputting a video AI generation probability score. The audio detector takes the news audio as input, analyzes the Mel-frequency cepstral coefficients (MFCC), fundamental frequency contour, and formant features of the audio, uses a deep neural network to identify AI-synthesized audio, outputting an audio AI generation probability score. The AI-generated content detection integrates the detection results of text, images, videos, and audio based on the news format, outputting a comprehensive AI-generated content detection score.

[0060] Multimodal image-text consistency analysis uses the news text and accompanying images collected by the data acquisition and processing module as input data. It employs a Cross-modal Feature Fusion Network (CFFN) to achieve deep consistency detection of the image-text content. The text encoder uses the news text as input and employs the BERT model to extract semantic features, outputting a 768-dimensional text feature vector. The image encoder uses the accompanying images as input and employs the ResNet-50 model to extract visual features, outputting a 2048-dimensional image feature vector. A cross-modal attention mechanism calculates the correlation matrix between the text feature vector and the image feature vector. By analyzing the correlation matrix, the system searches for consistent features in highly correlated regions and discovers inconsistencies in low-correlation regions, outputting consistency and inconsistency scores to effectively identify common forgery techniques involving real images paired with fake text. The fusion classifier concatenates the cross-modal attention features with the original features and outputs a multimodal image-text consistency score through a fully connected layer. This score is used to determine the degree of matching between the image and text content.

[0061] Specifically, it includes: Feature extraction: Feature extraction includes text feature extraction and image feature extraction. Text encoding uses a pre-trained BERT model, which can explore the bidirectional context of each word and extract its representation. Image feature extraction uses the Swin Transformer model, which has demonstrated excellent performance in various vision tasks.

[0062] Text encoding represents text as a sequence of words. The word embedding matrix is ​​obtained by inputting it into the BERT model. As shown in the formula below: ; Where N is the word sequence length of the text, and d is the dimension of the feature vector. In this embodiment, the feature dimension output by the BERT model is d=768. After projection through the fully connected layer, the word feature representation... .

[0063] Image feature extraction directly utilizes Swin-T to process the image and obtain the region feature matrix. The definition is as shown in the formula: E v =Swin-T(I img ); in, Let M represent the input image, and M represent the number of regions the image is divided into. In this embodiment, SwinTransformer divides the image into M=49 regions (i.e., a 7×7 grid). Finally, a fully connected layer projects the region features onto the same space as the word features to obtain the region representation. .

[0064] Existing multimodal rumor detection methods often utilize cross-modal attention mechanisms to select complementary information; however, these methods only focus on highly relevant information, neglecting the importance of low-relevance components in rumor detection. The fusion module of the CFFN model processes and fuses text and image features in a fine-grained manner, and explores the consistency of highly relevant components and the inconsistency of low-relevance components separately.

[0065] Correlation separation is a method that uses a threshold to classify the correlation matrix into high-correlation and low-correlation categories. This module first uses cosine similarity to calculate the correlation scores between different modal segments, as shown in the formula: ; in, Indicates the first Feature vectors of each word Indicates the first Feature vectors of image regions Representing vectors The transpose of . and Representing vectors respectively and The L2 norm (Euclidean norm) is calculated using the following formula: ,in is the dimension of the feature vector. The length of the word sequence in the text. The number of regions to divide the image into. In this embodiment, , (i.e., dividing the image into...) The formula essentially calculates the cosine similarity between word features and image region features, with values ​​ranging from [value range missing]. .

[0066] The correlation matrix S is divided into two parts by using a threshold λ∈[0,1), as shown in the formula: ; in, This is the correlation threshold, used to adjust the correlation matrix. It is divided into a highly correlated part and a low correlated part. In this embodiment, . The matching part represents all word-region pairs with relevance scores greater than a threshold. These pairs are considered to be regions where the text and image are semantically consistent. This section represents the inconsistency candidate portion, containing all word-region pairs with relevance scores less than or equal to the threshold. These pairings may indicate a mismatch between the text and the image. (Subscript) Represents "matching", subscript It represents "candidate".

[0067] For word-region pairs in the consistent part, aggregation is performed using a reference attention mechanism, as shown in the formula: ; Where σ is the softmax activation function, used to normalize the relevance score into attention weights. It is supplementary information in the visual modality, using words as queries to select background information in an image that is not related to the text.

[0068] For word-region pairs in the inconsistent candidate region, element-wise addition is used to obtain the representation, as shown in the formula: ; in the formula This represents element-wise addition. Finally, and Consistency scores are obtained by passing through the MLP layer respectively. Inconsistency score .

[0069] The MLP layer is a multilayer perceptron, consisting of two fully connected layers and a ReLU activation function. Its calculation process is as follows: ; ; in, For a trainable weight matrix, σ is the bias term, and σ is the Sigmoid activation function, which is used to normalize the output to the [0,1] interval. This represents the consistency score between the i-th word and the j-th image region. Let represent the inconsistency score corresponding to the i-th word.

[0070] The final aggregation process is shown in the formula: ; in, and These represent the consistent part and the candidate inconsistent part, respectively.

[0071] The selection module is the core of the CFFN model's decision-making process. It uses intelligent evaluation to determine whether consistent or inconsistent features are more crucial for judging the veracity of a rumor. The selection module receives the consistent portion obtained from the fusion module. and inconsistent candidate parts For each representation, an importance weight is calculated as shown in the formula: ; In this formula and These are trainable parameters. Then, the two weights are concatenated and normalized, as shown in the formula: ; in Indicates a connection operation. It is the softmax activation function. Finally, the two representations are aggregated based on the normalized weights to obtain the final multimodal content representation, as shown in the formula.

[0072] ; The total loss function is a linear combination of the binary classification cross-entropy loss and the segmentation loss, as shown in the formula below: ; In this formula It's a tag. The weights used to amplify the consensus portion of non-rumor amplification, while The weights used to amplify inconsistent candidate parts for rumor propagation are then applied. The total loss function is shown in the formula: ; in, It is the crossover loss function. It is a linear parameter.

[0073] like Figure 10As shown, the user dissemination confidence assessment uses the dissemination time records, dissemination geographic location data, and dissemination user profile data collected by the data acquisition and processing module as input data. It analyzes the dissemination characteristics of news from three perspectives: time dimension, geographic distribution, and user profile. The time dimension analysis (40% weight) uses the dissemination time records as input to calculate the time span and speed of news dissemination, outputting a timeline dimension score. The geographic distribution analysis (30% weight) uses the dissemination geographic location data as input to assess the news's coverage and geographic diversity across different provinces, outputting a geographic distribution dimension score. The user profile analysis (30% weight) uses the dissemination user profile data as input to analyze the age distribution and geographic diversity of participating users, outputting a user diversity score. The user dissemination confidence assessment outputs a user dissemination confidence score by weightedly fusing the timeline dimension score, geographic distribution dimension score, and user diversity score.

[0074] Specifically, it includes: definition Scoring based on the timeline dimension Scoring for the geographical distribution dimension User diversity is scored, and the final propagation confidence score is calculated as shown in the formula: ; The confidence calculation for each branch will be explained in detail below.

[0075] ① Timeline analysis (40% weight) Timeline analysis focuses on the time span and dissemination patterns of rumors, including two core indicators: time density and platform diversity. Time density is used to assess the time credibility of a rumor, comprehensively considering both the year the rumor was made and its dissemination density, as shown in the formula: ; in, It is the current year. It is the year the rumor was published. It is a time decay function (the earlier the rumor, the lower its credibility), where N is the total number of propagation events. It is the observation time window, for time difference α is the time decay weighting coefficient, used to control the influence of the year of news release on time density; β is the dissemination density weighting coefficient, used to control the influence of the number of dissemination events per unit time on time density. In this embodiment, α=0.5, β=0.5, that is, the time decay factor and the dissemination density factor each account for 50% of the weight. When it is necessary to emphasize the timeliness of news, the value of α can be appropriately increased; when it is necessary to pay more attention to the dissemination activity, the value of β can be appropriately increased, but the constraint condition α+β=1 must be met. The definition of the decay function is shown in the formula: ; Platform diversity is used to assess the breadth of a rumor's spread across different platforms, as shown in the formula: ; In this formula, It refers to the number of platforms with recorded dissemination data. This refers to the total number of reference platforms. This represents the number of disseminations on the platform with the largest reach, where N is the total number of dissemination events. According to... and The final timeline dimension score can be obtained, as shown in the formula: ; ② Geographical distribution analysis (30% weight) Geographic distribution analysis examines the spatial characteristics of rumor spread, including two core indicators: geographic coverage and geographic diversity. Geographic coverage is used to assess the number of provinces affected by the spread. If the number of provinces involved in the spread of rumors is the number of provinces, then the geographical coverage is calculated as shown in the formula: ; in, It represents the total number of times province i appears. It is a reference benchmark value, representing the total number of provinces a rumor should spread under ideal or standard conditions. The ideal conditions differ for different types of rumors.

[0076] Geographic diversity is used to assess the evenness of transmission across different provinces, excluding overlapping provinces, as shown in the formula: ; in, It refers to the number of different provinces with records of transmission. This represents the total number of provinces in China. This represents the number of times the virus spread in the province with the highest transmission volume. The final geographic distribution dimension score can be calculated as shown in the formula: ; ③ User profile analysis (30% weight) Suppose that the users participating in the dissemination are divided into K age groups, u l This represents the number of users in the l-th age group (l=1,2,...,K). In this embodiment, user ages are divided into K=6 age groups: under 18, 18-25, 26-35, 36-45, 46-55, and over 55. The age distribution degree A is defined. d The formula used to assess the breadth of age distribution among users participating in the dissemination is shown below: ; In this formula, K represents the number of age groups. Indicates the first Whether users in each age group participated, when The value is 1 when the value is greater than 0, and 0 otherwise. This is used to count the number of age groups with user participation. The number of users in the largest age group; This represents the total number of users participating in the dissemination. The design principle of this formula is: the wider the age range coverage and the more evenly the distribution of users across age groups, the higher the age distribution score.

[0077] User geographic diversity comprehensively assesses the correlation between users and geographic location from two dimensions: regional user dispersion and cross-regional propagation rate. Regional user dispersion is used to assess the evenness of user distribution across regions, as shown in the formula: ; Where R represents the total number of regions included in the statistics. This refers to the number of users in the region with the largest user base. This represents the total number of users across all regions. A higher value indicates a more balanced distribution of users across regions.

[0078] The cross-regional dissemination ratio measures the extent to which information spreads between different regions, as shown in the formula: ; in This refers to the number of cross-regional transmission events (i.e., the source and recipient of the transmission are located in different regions). This represents the total number of dissemination events. This metric assesses the geographical mobility of information; non-rumored information typically has a higher cross-regional dissemination rate. Ultimately, a user diversity score can be obtained, as shown in the formula: .

[0079] R dispersion This refers to the previously mentioned regional user dispersion, C ratio It is the cross-regional transmission rate.

[0080] like Figure 11As shown, the logical judgment based on the large language model uses the news text collected by the data acquisition and processing module as input data, and adopts the DeepSeek model for three-dimensional logical evaluation. Through deep semantic understanding and reasoning capabilities, it mines the characteristics of false information from the content itself. Content logical consistency evaluation analyzes the news text, detecting the degree of consistency in factual statements, timelines, and detailed descriptions, and outputs a content logical consistency score. Causal relationship rationality evaluation analyzes the events described in the news text, assessing whether the causal chain of events conforms to common sense and basic logic, and outputs a causal relationship rationality score. Narrative objectivity evaluation analyzes the narrative style of the news text, assessing the objectivity and neutrality of the news narrative style, and outputs a narrative objectivity score. The three-dimensional logical evaluation outputs a logical quality score by averaging the content logical consistency score, causal relationship rationality score, and narrative objectivity score.

[0081] Based on the analysis results from the three dimensions, the system calculates the final logic quality score, as shown in the formula: .

[0082] in, Score for content logical consistency. To score the reasonableness of the causal relationship, The objectivity of narrative techniques is scored. The weighting of the three dimensions is as follows: logical consistency and causal rationality each account for 35% of the weight, because these two dimensions directly reflect the factual accuracy of the news content; narrative objectivity accounts for 30% of the weight, serving as an auxiliary criterion for judging writing style. All three scores are normalized to the [0,1] interval, resulting in the final logical quality score. Similarly, the value range is [0,1].

[0083] In one embodiment, such as Figure 12 As shown, the multi-agent collaborative analysis module uses the true / false binary classification results output by the fast detection module and the six-dimensional evaluation results output by the multi-dimensional slow review module as input data. It is implemented using the Swarm framework and constructs seven specialized agents for parallel analysis and collaborative decision-making.

[0084] Title Analysis Expert (Role 1): Evaluate the emotional intensity and objective expression of the text, identify exaggeration / supplementary expression forms, and confirm the integrity of the title's grammatical structure. Syntax Verification Expert (Role 2): Verify the consistency of cross-linguistic elements, exclude disordered arrangements and abnormal symbols, and correct abnormal expressions. Common Sense Verification Expert (Role 3): Compare with authoritative knowledge bases to verify the compliance of common sense of core claims. Credibility Assessment Expert (Role 4): Conduct in-depth analysis of the information source URL, and evaluate the credibility of the domain name registration years and historical data through WHOIS queries. Multi-source Retrieval Expert (Role 5): Perform cross-engine data scraping, and verify the consistency of information dissemination through similarity comparison of reports from multiple sources. User Attribute Analysis Expert (Role 6): Confirm the matching degree of the query purpose in combination with user portrait data, and set the weight influence values of different attributes. Consistency Detection Expert (Role 7): Construct a logical verification matrix to eliminate expression contradictions between multi-source data.

[0085] Specifically include: ① Title Analysis Expert (Role 1): Evaluate the emotional intensity of the text (the detection result is neutral) and the objective expression, identify exaggeration / supplementary expression forms (the detected quantization value is 0.2), and confirm the integrity of the title's grammatical structure.

[0086] ② Syntax Verification Expert (Role 2): Verify the consistency of cross-linguistic elements (such as the correct embedding of Chinese and English terms "NBA"), exclude disordered arrangements and abnormal symbols (detect the abnormal expression "the 14th* are all" in the example and correct it).

[0087] ③ Common Sense Verification Expert (Role 3): Compare with authoritative knowledge bases in the sports field (such as ESPN, etc.), and verify the compliance of common sense of core claims such as "Kobe's historical status in the NBA" (the top 10 rankings are verified).

[0088] ④ Credibility Assessment Expert (Role 4): Conduct in-depth analysis of the information source URL (find that http: / / example.com is a placeholder domain name in the example), and evaluate the domain name registration years (detect the release date 2017 / 8 / 6) and the credibility of historical data through WHOIS queries.

[0089] ⑤ Multi-source Retrieval Expert (Role 5): Perform cross-engine data scraping (Google / Baidu / Academic Library), verify the consistency of information dissemination through similarity comparison of reports from more than 20 sources, and confirm the multi-party verification of the statement "historical superstar".

[0090] ⑥ User Attribute Analysis Expert (Role 6): Confirm the matching degree of the query purpose in combination with portrait data (the suitability of the sports theme and user preferences), and set the weight influence values of different attributes.

[0091] ⑦ Consistency Detection Expert (Role 7): Construct a logical verification matrix to eliminate inconsistencies in the representation of multi-source data (such as correcting the geographical association error of "Heilongjiang Province").

[0092] Parallel Analysis Phase: Seven agents simultaneously and independently analyze the true / false binary classification results and the six-dimensional evaluation results, with each agent outputting the evaluation result and confidence level for a specific dimension. Information Aggregation Phase: The analysis results of each agent are integrated through a weighted voting mechanism. The weights are dynamically adjusted based on historical accuracy. The weight calculation model includes core dimensions such as information source traceability quality (40%), content logical consistency (25%), multi-source verification coverage (20%), and user attribute suitability (15%), outputting a weighted comprehensive score. Negotiation and Decision-Making Phase: When the disagreement among agents exceeds a threshold, a negotiation mechanism is initiated. Consensus is reached through evidence exchange and reasoning verification. Based on the weighted score, the information is divided into four levels of conclusions: fake news (total score less than 1), suspicious content (1 to 2 points), possibly true (2 to 3 points), and credible content (3 points or more), outputting the four-level classification conclusions. Knowledge Update Phase: The verified new knowledge is structured and stored in the knowledge base to optimize subsequent analysis capabilities. Personalized knowledge graphs are iterated in real time based on user query characteristics, outputting updated knowledge base data.

[0093] The user interaction and education module constructs a collaborative governance ecosystem of detection, explanation, and education. It uses detection process data, user-submitted supplementary evidence, user error records, and user contribution records as input data, and includes four functional units: The evidence chain visualization unit uses the detection process data as input, visually displaying the detection process and judgment criteria through semantic traceability diagrams and logical heatmaps, and outputting a visual display interface; the community collaborative verification unit uses the user-submitted supplementary evidence as input, allowing users to submit supplementary evidence, which is then included in the evidence library after administrator review, and outputting the verified supplementary evidence to the evidence chain dashboard; the media literacy education unit uses the user error records as input, generating personalized learning paths and test questions based on these records, and outputting personalized learning paths and test questions; and the contribution incentive unit uses the user contribution records as input, establishing a user contribution ranking list, rewarding high-contribution users with points and level upgrades, and outputting user point and level data.

[0094] The system of this invention has achieved the following technical effects after actual testing: In terms of detection efficiency, the fast detection algorithm adopts a three-level processing architecture for model fine-tuning, achieving a detection accuracy of 95% within 2 seconds. It can detect simple news quickly and accurately, and clearly output true and false binary classification results and confidence scores.

[0095] In terms of detection depth, the slow review algorithm adopts a six-dimensional evaluation architecture design, which comprehensively judges the authenticity of news content from six aspects: confidence of news source, consistency between text and title, AI generation detection, consistency of multimodal text and image, confidence of user dissemination, and logical judgment based on DeepSeek. Each dimension module clearly outputs the corresponding evaluation score, making the analysis results reasonable and well-founded.

[0096] In terms of AI content detection, while focusing primarily on text generation, the system also comprehensively detects video, audio, and images. Each detector outputs an AI generation probability score for its corresponding media type, and the final output is a comprehensive AI-generated content detection score. This module is not only used for scoring in multi-dimensional analysis, but can also be used as a standalone module.

[0097] In terms of intelligent interaction, the system integrates the DeepSeek large model, allowing users to directly ask questions to the model and obtain accurate and valuable information, greatly improving the efficiency and accuracy of information retrieval.

[0098] In terms of multi-agent analysis, the multi-agent search function combines advanced search engine technology and the multi-agent framework Swarm to provide comprehensive and in-depth evaluation and analysis from multiple dimensions and perspectives. Each agent outputs evaluation results and confidence levels for specific dimensions, and finally outputs a four-level classification conclusion, ensuring the accuracy and comprehensiveness of the analysis results.

[0099] In terms of user education, the system is designed with user participation functions, including administrator review, evidence chain dashboard, user contribution leaderboard and community discussion, etc. Each unit clearly outputs a visual display interface, supplementary evidence after verification, personalized learning path and user point level data, which solves the problems of low user participation and lack of evidence sources.

[0100] In summary, the system of this invention, through its multi-dimensional fusion detection architecture, comprehensive AI-generated content detection, innovative multi-agent collaborative mechanism, and comprehensive user interaction and education functions, with each module clearly defining input data and output results, effectively solves the shortcomings of existing fake news detection systems in terms of multi-dimensional analysis capabilities, AI-generated content detection, multimodal analysis, and user participation, providing strong technical support for building a clean cyberspace.

[0101] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Contents not described in detail in this specification belong to prior art known to those skilled in the art.

Claims

1. A fake news intelligent detection system based on multi-dimensional fusion, characterized in that: include The data acquisition and processing module is used to collect news data from multiple news platforms and perform preprocessing. The news data includes news titles, news text, news images, news videos, news audio, publishing platform information, dissemination time records, dissemination geographical location data, and dissemination user profile data. The fast detection module uses the ERNIE 3.0 model, which is finely tuned based on a three-level architecture of word segmentation-enhancement-pre-training, to perform fast binary classification detection of true and false news data and output the true and false binary classification results. The multi-dimensional slow review module performs in-depth analysis on news data after rapid detection of true and false information from six evaluation dimensions. The six evaluation dimensions include news source confidence assessment, text-title consistency detection, AI-generated content detection, multimodal image-text consistency analysis, user dissemination confidence assessment, and logical judgment based on a large language model. A comprehensive trust score is calculated through a weighted fusion algorithm. The multi-agent collaborative analysis module constructs seven agents, including a title analysis expert, a syntax verification expert, a common sense verification expert, a credibility assessment expert, a multi-source retrieval expert, a user attribute analysis expert, and a consistency detection expert. These seven agents perform parallel analysis and collaborative decision-making on the output results of the fast detection module and the multi-dimensional slow review module, and output the final detection conclusion.

2. The intelligent fake news detection system based on multi-dimensional fusion according to claim 1, characterized in that, The preprocessing includes generating a summary using the TextRank algorithm and extracting keywords using the TF-IDF algorithm.

3. The fake news intelligent detection system based on multi-dimensional fusion according to claim 1, characterized in that, The news source confidence assessment uses the publishing platform information collected by the data acquisition and processing module as input data. The publishing platform information includes the platform URL and platform identifier, including: Construct a news platform reputation database, and classify platforms into five levels based on the platform identifier: high authority, relatively authoritative, neutral, relatively niche, and niche, and output the platform authority level identifier; The basic reputation score is calculated and output based on the official background, review mechanism, historical accuracy and user base of the platform corresponding to the published platform information. The credibility score is dynamically adjusted based on the platform's recent rate of fake news publication, correction rate, and complaint rate, and a news source confidence score is output.

4. The intelligent fake news detection system based on multi-dimensional fusion according to claim 1, characterized in that, The title consistency check includes: The pre-trained text2vec-base-chinese model is used to convert the news headlines and news text into high-dimensional vector representations, calculate the cosine similarity between the two vectors, and output the semantic consistency score. The TF-IDF algorithm is used to extract keywords from the news headlines, calculate the keyword coverage and weight matching degree in the news body, and output the keyword coverage and weight relevance scores. The final title-text consistency score is calculated based on semantic consistency score, keyword coverage, and weighted relevance score.

5. The intelligent fake news detection system based on multi-dimensional fusion according to claim 1, characterized in that, The AI-generated content detection includes: Using the news text as input, a Transformer-based bidirectional encoder structure is adopted. By analyzing the text's perplexity, suddenness, and lexical diversity indicators, the AI-generated text content is identified, and the text AI generation probability score is output. Using the news image as input, a convolutional neural network is used to extract the texture features, color distribution features, and edge detail features of the image. A support vector machine classifier is used to determine whether the image is generated by AI, and the AI ​​generation probability score of the image is output. Using the news video as input, the inter-frame difference features, motion vector features and temporal consistency features of the video are extracted. A recurrent neural network is used to perform AI-generated video recognition and output the video AI generation probability score. Using the news audio as input, the Mel frequency cepstral coefficients, fundamental frequency profile, and formant features of the audio are analyzed. A deep neural network is used to identify AI-synthesized audio and output an audio AI generation probability score. The AI-generated content detection system outputs a comprehensive detection score based on the detection results of news-style text, images, videos, and audio.

6. The intelligent fake news detection system based on multi-dimensional fusion according to claim 1, characterized in that, The multimodal image-text consistency analysis includes: Using the news article as input, the BERT model is used to extract the semantic features of the text and output a 768-dimensional text feature vector. Using the news image as input, the ResNet-50 model is used to extract the visual features of the image and output a 2048-dimensional image feature vector. Calculate the correlation matrix between the text feature vector and the image feature vector, find consistent features in high correlation regions, discover inconsistencies in low correlation regions, and output consistency scores and inconsistency scores. Cross-modal attention features are concatenated with the original features, and a multimodal image-text consistency score is output through a fully connected layer.

7. The fake news intelligent detection system based on multi-dimensional fusion according to claim 1, characterized in that, The user propagation confidence assessment includes: Using the aforementioned dissemination time record as input, calculate the time span and dissemination speed of the news dissemination, and output a timeline dimension score; Using the aforementioned geographical location data as input, the dissemination coverage and geographical diversity of news in different provinces are evaluated, and a geographical distribution dimension score is output. Using the aforementioned user profile data as input, the age distribution and geographical diversity of users participating in the dissemination are analyzed, and a user diversity score is output. The user propagation confidence assessment outputs a user propagation confidence score by weightedly fusing the timeline dimension score, the geographical distribution dimension score, and the user diversity score.

8. The intelligent fake news detection system based on multi-dimensional fusion according to claim 1, characterized in that, The logical judgment based on the large language model uses the DeepSeek model for three-dimensional logical evaluation: The news article is analyzed to detect whether there are self-contradictory factual descriptions, timeline conflicts, and inconsistencies in details, and the content logic consistency score is output. The events described in the news article are analyzed to assess whether the causal chain of the events conforms to common sense and basic logic, to determine whether the strength of the causal relationship is consistent with the actual impact, and to output a causal relationship rationality score. The narrative style of the news article is analyzed, words with emotional bias are identified in the text, the balance of different viewpoints and whether leading language is used are evaluated, and an objectivity score of the narrative technique is output. The three-dimensional logical evaluation outputs a logical quality score by averaging the scores for logical consistency of content, rationality of causal relationships, and objectivity of narrative techniques.

9. The intelligent fake news detection system based on multi-dimensional fusion according to claim 1, characterized in that, The multi-agent collaborative analysis is implemented using the Swarm framework, including: The seven agents simultaneously and independently analyze the true / false binary classification results and the six-dimensional evaluation results, with each agent outputting the evaluation result and confidence level for a specific dimension. The analysis results of each agent are integrated through a weighted voting mechanism, with the weights dynamically adjusted based on historical accuracy, and a weighted comprehensive score is output. When the disagreement among the agents exceeds a threshold, a negotiation mechanism is initiated to reach a consensus through evidence exchange and reasoning verification, and output a four-level classification conclusion, including fake news, suspicious content, possibly true and credible content. During the knowledge update phase, the verified new knowledge is structured and stored in the knowledge base to optimize subsequent analysis capabilities and output updated knowledge base data.

10. The intelligent fake news detection system based on multi-dimensional fusion according to claim 1, characterized in that, It also includes a user interaction and education module, which takes detection process data, supplementary evidence submitted by users, user error records, and user contribution records as input data. The evidence chain visualization unit takes the detection process data as input, and intuitively displays the detection process and judgment basis through semantic traceability diagram and logic heat map, and outputs a visualization display interface; The community collaborative verification unit takes the supplementary evidence submitted by the user as input, allows the user to submit supplementary evidence, and after the administrator reviews it, it is included in the evidence library and outputs the verified supplementary evidence to the evidence chain dashboard. The media literacy education unit takes the user's incorrect judgment records as input, generates personalized learning paths and test questions based on the user's incorrect judgment records, and outputs personalized learning paths and test questions. The contribution incentive unit uses the user contribution records as input to establish a user contribution ranking list, rewards high-contribution users with points and upgrades, and outputs user points and level data.

Images