Cross-domain fake information detection method and system based on explicit and implicit common feature enhancement learning

Abstract

The present application relates to the technical field of cyberspace security, in particular to a cross-domain false information detection method and system based on explicit and implicit common feature enhanced learning, which trains a false information detection model by using multi-source domain sample data, the model comprising: a first feature extraction unit for extracting explicit domain common features, a second feature extraction unit for extracting domain implicit semantic features, a feature fusion unit for interactive fusion of the explicit domain common features and the domain implicit semantic features, and a false information detection unit for outputting false information detection based on the fused features, wherein the fused features are simultaneously input into a domain adversarial training branch, a supervised contrast learning branch and a false information detection branch for joint enhanced learning during model training; and the false information detection model is used to judge the authenticity of target domain information data. The present application can improve the detection performance of the model on target domain false information under zero sample conditions.

Description

Technical Field

[0001] This invention relates to the field of cyberspace security technology, and in particular to a cross-domain method and system for detecting misinformation based on reinforcement learning of explicit and implicit common features. Background Technology

[0002] With the widespread use of the internet, people are increasingly inclined to share information and obtain the latest news on social networks. However, the lowered barriers to information dissemination, the low cost of anonymous posting, coupled with the unethical behavior of some media outlets or individuals seeking to gain traffic and manipulate public opinion, have led to an exponential surge in misinformation, posing a significant challenge to public safety and social trust. Faced with this increasingly severe challenge, researching how to automatically and accurately detect misinformation is of great practical significance for timely identifying various public opinion events, maintaining cyberspace security, and ultimately ensuring social stability.

[0003] Early misinformation detection methods typically extracted features from content, social context, and external evidence to achieve misinformation detection. These methods often performed well in single-domain detection tasks, but their effectiveness significantly decreased when applied to different domains. A "domain" can be understood as a collection of information surrounding a specific theme or event; information from different domains inherently differs in data distribution. For example, rumors in the health domain often contain professional medical terminology, while news in the entertainment domain uses more colloquial expressions. Furthermore, some studies, given the inherent multi-domain nature of information, have proposed multi-domain misinformation detection methods that model domain differences; however, these methods still lack sufficient detection capability for unseen domains. However, the need for misinformation detection in unseen domains such as breaking news and emerging topics is more urgent in real-world scenarios, and neither single-domain nor multi-domain methods can significantly improve their detection performance. To address this issue, cross-domain misinformation detection methods effectively identify misinformation in the target domain by transferring knowledge from the source domain to the target domain. These methods mainly include two categories: Domain Adaptation (DA) and Domain Generalization (DG). Domain adaptation methods either reduce inter-domain differences by aligning the feature distribution of unlabeled target domain samples or fine-tuning the transfer using a small number of labeled target domain samples. While these methods can improve performance, they rely on target domain data, reducing detection timeliness. To address this, domain generalization methods, in scenarios where target domain data is unavailable, learn common features from multiple source domains to construct models that can be directly adapted to any unseen target domain, achieving zero-shot cross-domain detection.

[0004] While existing domain generalization methods have made some progress in cross-domain misinformation detection, they still have certain limitations. Many existing methods rely on domain adversarial training, forcing feature extractors to generate implicit semantic features that are difficult for domain discriminators to distinguish through gradient inversion. However, their optimization goals focus on domain commonalities rather than the optimality of the detection task, resulting in the extraction of primarily domain-independent semantic features, failing to accurately extract common features for identifying misinformation. Specifically, this manifests in two ways: first, while suppressing domain-specific features, domain adversarial training retains domain commonalities irrelevant to misinformation identification; second, existing methods struggle to fully extract explicit common features of misinformation from implicit semantic representations, leading to the loss of crucial common features for identifying misinformation. These problems collectively limit the ability of existing methods to detect misinformation in the target domain under zero-shot conditions. Summary of the Invention

[0005] To address the shortcomings of existing methods in extracting common features of cross-domain misinformation, this invention provides a cross-domain misinformation detection method and system based on explicit and implicit common feature reinforcement learning. It learns cross-domain common features from implicit semantic representation through domain adversarial training and extracts multi-dimensional explicit common features of misinformation. Supervised contrastive learning is introduced into the domain adversarial training to improve the model's detection performance of misinformation in the target domain under zero-sample conditions.

[0006] According to the design scheme provided by this invention, on the one hand, a cross-domain fake information detection method based on explicit and implicit common feature reinforcement learning is provided, comprising:

[0007] Acquire multi-source domain sample data, where each source domain sample corresponds to an information topic or event;

[0008] A fake information detection model is constructed and trained using multi-source domain sample data. The fake information detection model includes: a first feature extraction unit for extracting explicit domain common features, a second feature extraction unit for extracting domain implicit semantic features, a feature fusion unit for interactively fusing explicit domain common features and domain implicit semantic features, and a fake information detection unit for outputting fake information detection based on the fused features. During model training, the fused features are simultaneously input into the domain adversarial training branch, the supervised contrastive learning branch, and the fake information detection branch for joint reinforcement learning.

[0009] The target domain information data is input into the trained fake information detection model, and the fake information detection model is used to determine the authenticity of the target domain information data.

[0010] As a cross-domain misinformation detection method based on explicit and implicit common feature reinforcement learning according to the present invention, further, the explicit domain common features are extracted using the first feature extraction unit, including:

[0011] Cross-domain common features were extracted from four dimensions: word level, sentence level, document level, and sentiment. The types of cross-domain common features for word level, sentence level, and document level were determined by experimental screening, while the type of cross-domain common feature for sentiment dimension was sentiment intensity.

[0012] The extracted cross-domain common features are screened, spliced, and preprocessed to obtain explicit domain common features. The preprocessing includes normalization and nonlinear transformation.

[0013] As a cross-domain misinformation detection method based on explicit and implicit common feature reinforcement learning, this invention further extracts cross-domain common features from four dimensions: word level, sentence level, document level, and sentiment, including:

[0014] By statistically analyzing the proportion of each part of speech and punctuation mark in the text, word-level features are obtained to describe the distribution pattern of the text at the part of speech and punctuation level.

[0015] Sentence-level features describing the syntactic complexity and structure of a text are obtained by extracting the average dependency tree depth, average dependency distance, and average theater.

[0016] Document-level features describing text size and readability are obtained by statistically analyzing the total number of words, total number of characters, and number of sentences in the text.

[0017] We use a large-scale sentiment model to extract the sentiment intensity of text.

[0018] As a cross-domain fake information detection method based on explicit and implicit common feature enhancement learning according to the present invention, further, the method utilizes a second feature extraction unit to extract domain implicit semantic features, including:

[0019] Each text is segmented and embedded to obtain a text embedding vector;

[0020] The text embedding vector is input into a pre-trained language model for encoding to obtain a sequence of hidden text states.

[0021] The hidden vector corresponding to the specified tag is extracted based on the hidden state sequence of the text, and the extracted hidden vector is used as the global semantic representation of the text.

[0022] As a cross-domain fake information detection method based on explicit and implicit common feature enhancement learning, this invention further utilizes a feature fusion unit to interactively fuse explicit domain common features and domain implicit semantic features, including:

[0023] Explicit domain common features are used as queries, and implicit domain semantic features are used as keys. The features are mapped to the attention space through linear projection. A multi-head attention mechanism is used to evenly divide the projection vector according to the number of heads. The scaling dot product attention of each attention head is calculated, and the multi-head outputs are concatenated and integrated through linear projection to obtain the enhanced implicit domain semantic features.

[0024] Using implicit semantic features of the domain as queries and explicit domain-shared features as keys, the features are mapped to the attention space through linear projection and information related to the implicit semantic features of the domain is mined through a multi-head attention mechanism to obtain enhanced explicit domain-shared features.

[0025] The enhanced domain implicit semantic features and explicit domain shared features are concatenated, and the fusion weights are calculated through a gating network. The final fused features are obtained based on the weights. The gating network is used to dynamically balance the fusion ratio of the domain implicit semantic features and explicit domain shared features.

[0026] As a cross-domain fake information detection method based on explicit and implicit common feature reinforcement learning according to the present invention, the model training process further includes:

[0027] The task of detecting false information is combined with domain adversarial training and supervised contrastive learning, and a joint loss function is constructed. The joint loss function is obtained by weighted fusion of false information detection training loss, domain adversarial training loss and supervised contrastive learning loss. In the domain adversarial training, domain-independent features are automatically learned through adversarial game between feature extractor and domain discriminator. In the supervised contrastive learning, an in-batch supervised sampling strategy is adopted and positive and negative sample pairs are constructed based on real label information to enhance the ability to distinguish between true and false information of common domain features through supervised contrastive learning.

[0028] The model is trained and optimized based on the joint loss function and by utilizing the fusion features of multi-source domain sample data.

[0029] As a cross-domain fake information detection method based on explicit and implicit common feature enhancement learning, this invention further includes automatically learning domain-independent features through adversarial game between a feature extractor and a domain discriminator, comprising:

[0030] In the forward propagation, the gradient inversion layer is used to pass the fused features to the neighborhood discriminator, so that the neighborhood discriminator can predict the source domain of the sample based on the fused features;

[0031] In backpropagation, a gradient inversion layer is used to invert the gradient sign so that the feature extractor generates domain-independent features.

[0032] Furthermore, this invention also provides a cross-domain fake information detection system based on explicit and implicit common feature reinforcement learning, comprising: a sample acquisition module, a model training module, and a target detection module, wherein,

[0033] The sample acquisition module is used to acquire multi-source domain sample data, where each source domain sample corresponds to an information topic or event;

[0034] The model training module is used to construct a fake information detection model and train the model using multi-source domain sample data. The fake information detection model includes: a first feature extraction unit for extracting explicit domain common features, a second feature extraction unit for extracting domain implicit semantic features, a feature fusion unit for interactively fusing explicit domain common features and domain implicit semantic features, and a fake information detection unit for outputting fake information detection based on the fused features. In the model training, the fused features are simultaneously input into the domain adversarial training branch, the supervised contrastive learning branch, and the fake information detection branch for joint reinforcement learning.

[0035] The target detection module is used to input target domain information data into the trained fake information detection model, and use the fake information detection model to determine the authenticity of the target domain information data.

[0036] The beneficial effects of this invention are:

[0037] This invention explicitly extracts cross-domain common features of false information from multiple dimensions. It identifies explicit common features of false information at the word, sentence, and document levels, including language style and sentence / text structure, and combines these with sentiment characteristics to enrich the representational dimensions of these cross-domain common features. A bidirectional cross-attention mechanism and gating units are used to achieve deep bidirectional interactive fusion of explicit common features and implicit semantic features. The gating units adaptively adjust the fusion ratio of the two types of features to fully explore the complementary information between them. Furthermore, supervised contrastive learning is introduced based on domain adversarial training. By constructing positive and negative samples guided by true and false labels to enhance the discriminative ability of features, the extraction capability of cross-domain common features of false information is strengthened, improving the model's detection performance of false information in the target domain under zero-sample conditions. Experiments were conducted on the Chinese dataset Weibo21 and the English datasets PHEME and AMTCele. The results show that the F1 scores of the proposed scheme reached 86.87%, 73.34%, and 74.37%, respectively, which are 2.26%, 1.71%, and 4.35% higher than the best domain generalization baseline, further verifying the effectiveness of the proposed scheme. Attached Figure Description

[0038] Figure 1 This is a schematic diagram of the cross-domain misinformation detection process based on explicit and implicit common feature reinforcement learning in the embodiment.

[0039] Figure 2 This is a schematic diagram of the cross-domain misinformation detection model architecture in the embodiment. Detailed Implementation

[0040] To make the objectives, technical solutions, and advantages of this invention clearer and more understandable, the invention will be further described in detail below with reference to the accompanying drawings and technical solutions.

[0041] Given Source domain Each source domain corresponds to a specific topic or event. Source domain Include A labeled information sample, among which Indicates the first The text content of the message, For the corresponding category labels (0 for true, 1 for false), each source domain shares the label space but there is a distribution offset, i.e. , Using only source domain data Learning a fake information detection model This allows it to generalize directly to target domains not seen during training. To detect false information in the target domain, including , .

[0042] Cross-domain misinformation detection aims to identify misinformation in uncharted domains, such as emerging events and topics, online. Existing detection methods based on domain generalization learn cross-domain common features from implicit semantic representations through domain adversarial training. However, these methods primarily extract domain-independent semantic features, failing to accurately extract common features for identifying misinformation, thus limiting the model's detection capabilities. Therefore, this invention provides embodiments, see [link to embodiments]. Figure 1 As shown, a cross-domain fake information detection method based on explicit and implicit common feature reinforcement learning is provided, comprising:

[0043] S101. Obtain multi-source domain sample data, wherein each source domain sample corresponds to an information topic or event;

[0044] S102. Construct a fake information detection model and train the model using multi-source domain sample data. The fake information detection model includes: a first feature extraction unit for extracting explicit domain common features, a second feature extraction unit for extracting domain implicit semantic features, a feature fusion unit for interactively fusing explicit domain common features and domain implicit semantic features, and a fake information detection unit for outputting fake information detection based on the fused features. In the model training, the fused features are simultaneously input into the domain adversarial training branch, the supervised contrastive learning branch, and the fake information detection branch for joint reinforcement learning.

[0045] S103. Input the target domain information data into the trained fake information detection model, and use the fake information detection model to determine the authenticity of the target domain information data.

[0046] like Figure 2 As shown, for a given multi-source domain input sample, the model first extracts explicit common features from four dimensions: word level, sentence level, document level, and sentiment. Simultaneously, it acquires implicit semantic features through a pre-trained language model, RoBERTa. Then, a bidirectional cross-attention mechanism is used to achieve deep interaction between the two types of features, and a gating unit adaptively adjusts the fusion ratio. During the training phase, the fused features are simultaneously input into three branches for joint reinforcement learning. Domain adversarial training suppresses domain-specific features through a gradient reversal layer, enhancing the domain invariance of the features. Supervised contrastive learning constructs positive and negative sample pairs guided by true and false labels, bringing similar samples closer together and pushing away dissimilar samples, strengthening the feature's ability to distinguish between true and false information. The fake information detector maps the fused features to a classification space, outputting the probability distribution of true and false information. Finally, the model is optimized through a joint loss function to detect fake information in unseen target domains.

[0047] Specifically, the first feature extraction unit extracts explicit domain common features, which can be designed to include:

[0048] Cross-domain common features were extracted from four dimensions: word level, sentence level, document level, and sentiment. The types of cross-domain common features for word level, sentence level, and document level were determined by experimental screening, while the type of cross-domain common feature for sentiment dimension was sentiment intensity.

[0049] The extracted cross-domain common features are screened, spliced, and preprocessed to obtain explicit domain common features. The preprocessing includes normalization and nonlinear transformation.

[0050] A hierarchical explicit feature extraction framework is used to extract cross-domain common features from four dimensions: word level, sentence level, document level, and sentiment. Word-level, sentence-level, and document-level common features are obtained through feature analysis experiments, while sentiment features use sentiment intensity (Vreg) as a common feature. Given input text... Explicit feature extractors are defined as mappings ,in, This represents the total dimension of explicit features.

[0051] Specifically, cross-domain common features are extracted from four dimensions: word level, sentence level, document level, and sentiment, including:

[0052] By statistically analyzing the proportion of each part of speech and punctuation mark in the text, word-level features are obtained to describe the distribution pattern of the text at the part of speech and punctuation level.

[0053] Sentence-level features describing the syntactic complexity and structure of a text are obtained by extracting the average dependency tree depth, average dependency distance, and average theater.

[0054] Document-level features describing text size and readability are obtained by statistically analyzing the total number of words, total number of characters, and number of sentences in the text.

[0055] We use a large-scale sentiment model to extract the sentiment intensity of text.

[0056] Word-level features characterize the statistical distribution pattern of misinformation text at the part-of-speech and punctuation levels, represented by the proportion of each category in the text. Let the text... Include There are 1 word elements, and the set of part-of-speech categories is 1. The part-of-speech distribution feature is defined as follows: ,in Indicates part of speech The number of times a character appears in the text; let the set of punctuation marks be... The characteristics of punctuation marks are defined as follows: Word-level features are ultimately represented as .

[0057] Sentence-level features characterize the syntactic complexity and structure of misinformation texts by extracting the average dependency tree depth, average dependency distance, and average sentence length. The average dependency tree depth is the arithmetic mean of the depths of the entire text's lexical dependency trees, where lexical depth is the number of dependency edges traversed upwards to the root node. Where T is the set of text lexical units, and depth(t) is the depth of lexical unit t; the average dependency distance is the arithmetic mean of the absolute values ​​of the differences between the positions of lexical units in the whole text and their direct headwords, i.e. Where pos(⋅) represents the lexical position index, and h(t) is the direct headword of lexical t; the average sentence length is the arithmetic mean of the number of lexical units in all sentences of the text, i.e. Where S is the set of sentences in the text, and |s| is the number of lexical units in sentence s. The sentence-level features are ultimately represented as... .

[0058] Document-level features characterize the size and readability of misinformation texts. Text size features include the total number of words. Total number of characters and number of sentences Text readability features are quantitative indicators that measure the ease with which readers can understand and absorb text content. These include the Flesch-Kincaid Grade Level, Gunning Fog Index, and SMOG Index. Document-level features are ultimately represented as... .

[0059] False information often exhibits a stronger emotional bias to attract readers' attention and remains stable across different domains. In this embodiment, the EmoLLaMA-chat-7B emotional big language model can be used to extract the emotional intensity feature Vreg of the text to obtain an emotional feature representation. ,in, For the emotional characteristics dimension.

[0060] The above features are filtered and concatenated according to different languages ​​to obtain the original explicit domain feature vector:

[0061]

[0062] Since different features differ significantly in units and numerical ranges, the global mean is first calculated based on data from all source domains. and standard deviation Z-Score standardization was used to obtain To map explicit features to a representation space consistent with implicit features, a multilayer perceptron (MLP) is introduced for nonlinear transformation.

[0063]

[0064] in For learnable parameters, For activation function, The target feature dimension.

[0065] Specifically, the second feature extraction unit, which extracts implicit semantic features of the domain, can be designed to include:

[0066] Each text is segmented and embedded to obtain a text embedding vector;

[0067] The text embedding vector is input into a pre-trained language model for encoding to obtain a sequence of hidden text states.

[0068] The hidden vector corresponding to the specified tag is extracted based on the hidden state sequence of the text, and the extracted hidden vector is used as the global semantic representation of the text.

[0069] The pre-trained language model RoBERTa can be used as an implicit semantic feature encoder. RoBERTa demonstrates superior semantic representation capabilities across multiple natural language understanding tasks through dynamic masking strategies, larger-scale pre-training corpora, and longer training cycles. Given input text... After word segmentation and embedding, the sequence is input into the RoBERTa encoder to obtain the hidden state sequence:

[0070]

[0071] in The sequence length is given. The hidden vector corresponding to the special marker [CLS] is extracted as the global semantic representation of the text, i.e. Considering that the RoBERTa encoder keeps its parameters frozen during training to preserve pre-trained knowledge, a task adaptation layer is introduced to enhance the downstream task adaptability of features:

[0072]

[0073] in and For learnable parameters, This is the activation function.

[0074] Simple feature concatenation is insufficient to fully uncover the intrinsic relationship between explicit and implicit features. In this embodiment, a feature fusion unit is used to interactively fuse explicit domain-common features and domain-implicit semantic features, which can be designed to include:

[0075] Explicit domain common features are used as queries, and implicit domain semantic features are used as keys. The features are mapped to the attention space through linear projection. A multi-head attention mechanism is used to evenly divide the projection vector according to the number of heads. The scaling dot product attention of each attention head is calculated, and the multi-head outputs are concatenated and integrated through linear projection to obtain the enhanced implicit domain semantic features.

[0076] Using implicit semantic features of the domain as queries and explicit domain-shared features as keys, the features are mapped to the attention space through linear projection and information related to the implicit semantic features of the domain is mined through a multi-head attention mechanism to obtain enhanced explicit domain-shared features.

[0077] The enhanced domain implicit semantic features and explicit domain shared features are concatenated, and the fusion weights are calculated through a gating network. The final fused features are obtained based on the weights. The gating network is used to dynamically balance the fusion ratio of the domain implicit semantic features and explicit domain shared features.

[0078] Implicit feature enhancement branch guided by explicit features As a query, implicit features For key / value pairs, semantic information related to explicit common features is highlighted. First, features are mapped to the attention space via linear projection:

[0079]

[0080] in This is the projection matrix. To enhance the model's ability to capture multi-angle feature interactions, a multi-head attention mechanism is adopted, which distributes the projection vectors according to the number of heads. Uniformly cut, that is ,in The feature dimension of each head is Calculate scaled dot product attention for each attention head:

[0081]

[0082] Multi-head outputs are spliced ​​together and integrated using linear projection:

[0083]

[0084] By introducing residual connections and layer normalization to alleviate gradient vanishing, and by enhancing nonlinear expressive power through a feedforward network, we finally obtain the enhanced implicit semantic feature representation:

[0085]

[0086]

[0087] The implicit feature-guided explicit feature enhancement branch is symmetrical to the above process, using implicit features... For Query, explicit features For Key / Value pairs, explicit features related to semantic features are mined. Through the same multi-head attention, residual connections, and feedforward network operations, an enhanced representation of explicit common features is obtained. .

[0088] The contributions of explicit and implicit features to the false information detection task vary across different domains and samples. To achieve adaptive adjustment of feature contributions, a gating unit is introduced to dynamically balance the fusion ratio. After concatenating the bidirectional attention outputs, the fusion weights are calculated through a gating network:

[0089]

[0090] in This is the weight matrix. For the Sigmoid function. Gate value. Reflecting the relative importance of the two types of features, The closer to 1, the more explicit the semantic features. The more important it is, the closer it is to 0, indicating an explicit common characteristic of implicit guidance. The more important. Final fusion features. .

[0091] Specifically, the model training process can be designed to include:

[0092] The task of detecting false information is combined with domain adversarial training and supervised contrastive learning, and a joint loss function is constructed. The joint loss function is obtained by weighted fusion of false information detection training loss, domain adversarial training loss and supervised contrastive learning loss. In the domain adversarial training, domain-independent features are automatically learned through adversarial game between feature extractor and domain discriminator. In the supervised contrastive learning, an in-batch supervised sampling strategy is adopted and positive and negative sample pairs are constructed based on real label information to enhance the ability to distinguish between true and false information of common domain features through supervised contrastive learning.

[0093] The model is trained and optimized based on the joint loss function and by utilizing the fusion features of multi-source domain sample data.

[0094] Specifically, domain-independent features are automatically learned through adversarial game between the feature extractor and the domain discriminator. In the forward propagation, the gradient inversion layer is used to pass the fused features to the domain discriminator so that the domain discriminator can predict the source domain of the sample based on the fused features. In the backpropagation, the gradient inversion layer is used to invert the gradient sign so that the feature extractor can generate domain-independent features.

[0095] Domain adversarial training aims to constrain the domain invariance of fused features, enabling the common features learned by the model to be stably transferred to unseen target domains. The core idea originates from domain adversarial neural networks, which achieve automatic learning of domain-independent features through an adversarial game between a feature extractor and a domain discriminator.

[0096] Domain discriminator based on fused features The source domain of the predicted samples is determined using a two-layer fully connected network as the underlying architecture.

[0097]

[0098]

[0099] in For hidden layer representation, For the sample to belong to The predicted probability distribution of each source domain satisfies ; These are learnable parameters. The Gradient Reversal Layer (GRL) is a key component for implementing adversarial training, and its behavior is defined as... , ,in The gradient reversal coefficient, The identity matrix is ​​used. During forward propagation, the GRL acts as an identity mapping, ensuring that the fused features are properly transmitted to the neighborhood discriminator. During backward propagation, the GRL reverses the gradient sign, forcing the feature extractor to generate features that are difficult for the neighborhood discriminator to distinguish.

[0100] Multi-class cross-entropy loss is used to quantify the discriminative power of the domain discriminator. Let... For the one-hot encoding of the real domain labels of the sample, where Indicates that the sample comes from the first Each source domain. Domain adversarial loss can be defined as:

[0101]

[0102] During training, the domain discriminator and the feature extractor engage in a minimax game: the domain discriminator minimizes... To improve the accuracy of domain discrimination; the feature extractor maximizes the reception through gradient inversion. The gradient signal prompts it to generate domain-independent features. This adversarial mechanism can be formalized as... ,in and These are the parameters for the feature extractor and the neighborhood discriminator, respectively.

[0103] In this embodiment, based on domain adversarial training and the introduction of supervised contrastive learning, positive and negative sample pairs guided by real and fake labels are constructed to enhance the ability to distinguish between real and fake common features in the domain.

[0104] The effectiveness of contrastive learning largely depends on the quality of the constructed sample pairs. By employing an in-batch supervised sampling strategy, positive and negative sample pairs are constructed based on the true label information. Given a training batch... For anchor point samples Define the set of positive samples This includes all other samples within the batch that have the same label as the anchor sample; the negative sample set. This includes all samples within the batch whose labels differ from the anchor sample's. Supervised contrastive loss is used as the optimization objective for the anchor sample. Its loss is defined as:

[0105]

[0106] in The cosine similarity function is used. Temperature is a hyperparameter used to control the smoothness of the similarity distribution. The contrast loss for the entire batch is:

[0107]

[0108] During the optimization process, the feature distance between the anchor point and the positive sample is reduced, while the feature distance between the anchor point and the negative sample is increased.

[0109] The fake information detection layer, as the final task layer of the model, is responsible for binary classification of true and false information. A two-layer fully connected network can be used to map the fused features to the discrimination space. The forward propagation process is defined as follows:

[0110]

[0111]

[0112] in For hidden layer representation, For the hidden layer dimension; To predict the probability distribution, satisfying , representing the probability that the sample is judged as real information and false information, respectively; These are learnable parameters.

[0113] Cross-entropy loss is used as the optimization objective for the classification task. Let... The sample's true label is converted into a one-hot encoding. ,in The classification loss is defined as:

[0114]

[0115] The classification task is combined with domain adversarial training and supervised contrastive learning, and optimized through a joint loss function:

[0116]

[0117] in, and This is the loss weighting coefficient.

[0118] The model is optimized through a joint loss function to detect false information in unseen target domains.

[0119] Furthermore, based on the above method, this embodiment of the invention also provides a cross-domain fake information detection system based on explicit and implicit common feature reinforcement learning, comprising: a sample acquisition module, a model training module, and a target detection module, wherein,

[0120] The sample acquisition module is used to acquire multi-source domain sample data, where each source domain sample corresponds to an information topic or event;

[0121] The model training module is used to construct a fake information detection model and train the model using multi-source domain sample data. The fake information detection model includes: a first feature extraction unit for extracting explicit domain common features, a second feature extraction unit for extracting domain implicit semantic features, a feature fusion unit for interactively fusing explicit domain common features and domain implicit semantic features, and a fake information detection unit for outputting fake information detection based on the fused features. In the model training, the fused features are simultaneously input into the domain adversarial training branch, the supervised contrastive learning branch, and the fake information detection branch for joint reinforcement learning.

[0122] The target detection module is used to input target domain information data into the trained fake information detection model, and use the fake information detection model to determine the authenticity of the target domain information data.

[0123] To verify the effectiveness of this solution, the following explanation is based on experimental data:

[0124] We selected three commonly used public datasets for multi-domain fake information detection for experiments, including the Chinese dataset Weibo21 and the English datasets PHEME and AMTCele:

[0125] Weibo21 is one of the largest multi-domain Chinese misinformation detection benchmark datasets currently available. It was collected from public posts on Sina Weibo from December 2014 to March 2021, covering nine subject areas.

[0126] PHEME is an English-language rumor detection dataset originating from Twitter, containing both rumors and non-rumors related to nine breaking news events.

[0127] AMTCele consists of two subset datasets: FakeNewsAMT, built using the Amazon Mechanical Turk crowdsourcing platform, covers news in six fields, with fake news generated by crowdsourcing workers by rewriting real news in the style of real news; the Celebrity dataset focuses on celebrity news, collecting real and fake reports from websites.

[0128] Table 1 shows the detailed sample distribution for each domain in the three datasets.

[0129] Table 1 Statistical information for the three datasets

[0130]

[0131] Feature analysis was performed on the three datasets to screen for explicit common features that showed statistically significant differences between true and false information. A hierarchical feature system covering three granularities—word level, sentence level, and document level—was constructed, as shown in Table 2.

[0132] Table 2 Explicit Feature System

[0133]

[0134] Feature selection employed hypothesis testing: First, each feature was grouped into true and false categories and subjected to the Shapiro-Wilk normality test. If both groups followed a normal distribution, a t-test was used; otherwise, the Mann-Whitney U rank-sum test was used. Feature significance was determined with p ≤ 0.05. Dozens of features, including part-of-speech distribution, punctuation, syntactic structure, text size, and readability metrics, were systematically analyzed. Due to space limitations, only the selected explicit common features are listed.

[0135] English features were selected following the principle of cross-dataset consistency, taking the intersection of features that showed statistically significant differences between the PHEME and AMTCele datasets; Chinese features were selected based on the analysis results of the Weibo21 dataset. The final selected explicit common features are shown in Table 3, and the sentiment feature uses the sentiment intensity Vreg.

[0136] Table 3 Comparison of Explicit Common Characteristics in Chinese and English

[0137]

[0138] A leave-one-domain cross-validation strategy is employed to evaluate cross-domain generalization ability. Specifically, in each round of experiments, one domain is selected from all domains in the dataset as the target test domain, and the remaining domains are merged to form the source training domain. The model is trained using only the source domain data and then tested on the target domain. This process is repeated until each domain has been evaluated once as the target test domain. The final performance metric is the average of the results from all rounds.

[0139] Explicit feature extraction primarily utilizes the spaCy toolkit (Chinese zh_core_web_trf, English en_core_web_trf), readability metrics are calculated based on cntext 2.2.0, and sentiment common features are extracted using EmoLLaMA-chat-7B. Implicit semantic features employ the RoBERTa encoder (Chinese chinese-roberta-wwm-ext, English robberta-base), which remains frozen during training. The Adam optimizer is used with a learning rate of 2×10⁻⁶. −5 Batch size 32, maximum training epochs 50, early stopping based on validation set F1 score. Weight coefficients in the loss function. (Domain adversarial loss weight) (Contrastive learning loss weights) and contrastive learning temperature parameters All values ​​were determined through grid search, and the optimal values ​​were selected for experimentation. ; ; ).

[0140] Accuracy and F1 score are used as evaluation metrics for model performance. Both are standard metrics widely used in the field of fake information detection. The final result is the average value of all target domains.

[0141] Four categories of ten methods were selected as baselines: (1) Single-domain methods: BERT and RoBERTa are fine-tuned directly on the mixed source domain without considering domain differences; (2) Multi-domain methods: MDFEND uses a hybrid expert network, and M3FEND introduces a domain memory mechanism. Both are designed for multi-domain scenarios but are not for generalization to unseen domains; (3) Domain adaptation methods: CANMD is based on unsupervised adaptation of contrastive learning, and MetaAdapt uses a meta-learning strategy. Both require target domain data to participate in training; (4) Domain generalization methods: EANN and BDANN use domain adversarial training to learn domain-invariant features, and FADED further introduces dual-granularity adversarial training to alleviate entity bias. Both do not depend on target domain data.

[0142] To verify the effectiveness of the EICL method proposed in this case, Table 4 compares its performance with that of each baseline method on three datasets, with the best results marked in bold.

[0143] Table 4. Results of EICL Comparison Experiment

[0144]

[0145] Experimental results show that EICL achieves state-of-the-art performance on all datasets: F1 scores of 86.87%, 73.34%, and 74.37% on Weibo21, PHEME, and AMTCele, respectively, representing improvements of 2.26%, 1.71%, and 4.35% over the best domain generalization baseline FADED. Compared to single-domain methods, EICL significantly outperforms BERT and RoBERTa, highlighting the importance of explicit common features for cross-domain generalization. Compared to multi-domain methods, M3FEND and others lag significantly behind on PHEME and AMTCele due to their lack of design for unseen domains. Compared to domain-adaptive methods, EICL still outperforms CANMD and MetaAdapt, which require target domain data, even with zero-sample settings, demonstrating its stronger generalization ability. Compared to existing domain generalization methods such as EANN and BDANN, which rely solely on implicit features, EICL enhances the model from both feature representation and optimization objectives by introducing explicit common features and supervised contrastive learning, achieving continuous performance improvements. Experimental results validate the effectiveness of EICL in reinforcement learning through the fusion of explicit and implicit features.

[0146] To verify the effectiveness of each core component of EICL in this scheme, ablation experiments were conducted to remove explicit common features (w / o ECF), explicit and implicit interaction fusion (w / o EIF), domain adversarial training (w / o DAT), and supervised contrastive learning (w / o SCL) in sequence. The results are shown in Table 5.

[0147] Table 5 Results of EICL Ablation Analysis

[0148]

[0149] Experimental results show that the complete model performs optimally, and the absence of any component leads to performance degradation. Specifically, the removal of domain adversarial training causes the most significant performance decline, demonstrating its fundamental role in learning domain-invariant features. The lack of explicit common features also leads to a significant decline, indicating that the explicit common features it provides in lexical, syntactic, and sentiment aspects are key cross-domain discriminative cues. The removal of supervised contrastive learning results in similar losses, validating the necessity of constructing positive and negative sample pairs to enhance feature discriminativity. Furthermore, the absence of the interactive fusion mechanism also impairs performance, indicating that bidirectional attention and gating units achieve deep complementary fusion of features.

[0150] The experimental data above demonstrates that the proposed solution has superior performance in detecting misinformation and can meet the needs of detecting misinformation in unseen areas such as sudden events and emerging topics in real-world scenarios.

[0151] Unless otherwise specifically stated, the relative steps, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of the invention.

[0152] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0153] The units and method steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of each example have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations are not considered to be beyond the scope of this invention.

[0154] Those skilled in the art will understand that all or part of the steps in the above methods can be implemented by a program instructing related hardware, and the program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk. Optionally, all or part of the steps in the above embodiments can also be implemented using one or more integrated circuits. Accordingly, each module / unit in the above embodiments can be implemented in hardware or as a software functional module. This invention is not limited to any particular combination of hardware and software.

[0155] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A cross-domain fake information detection method based on explicit and implicit common feature reinforcement learning, characterized in that, Include: Acquire multi-source domain sample data, where each source domain sample corresponds to an information topic or event; A fake information detection model is constructed and trained using multi-source domain sample data. The fake information detection model includes: a first feature extraction unit for extracting multi-dimensional explicit domain common features, a second feature extraction unit for extracting domain implicit semantic features, a feature fusion unit for interactively fusing explicit domain common features and domain implicit semantic features, and a fake information detection unit for outputting fake information detection based on the fused features. In the model training, the fused features are simultaneously input into the domain adversarial training branch, the supervised contrastive learning branch, and the fake information detection branch for joint reinforcement learning. The target domain information data is input into the trained fake information detection model, and the fake information detection model is used to determine the authenticity of the target domain information data.

2. The cross-domain fake information detection method based on explicit and implicit common feature reinforcement learning according to claim 1, characterized in that, The first feature extraction unit is used to extract multi-dimensional explicit domain common features, including: Cross-domain common features were extracted from four dimensions: word level, sentence level, document level, and sentiment. The types of cross-domain common features for word level, sentence level, and document level were determined by experimental screening, while the type of cross-domain common feature for sentiment dimension was sentiment intensity. The extracted cross-domain common features are screened, spliced, and preprocessed to obtain explicit domain common features. The preprocessing includes normalization and nonlinear transformation.

3. The cross-domain fake information detection method based on explicit and implicit common feature reinforcement learning according to claim 2, characterized in that, Cross-domain common features are extracted from four dimensions: word level, sentence level, document level, and sentiment, including: By statistically analyzing the proportion of each part of speech and punctuation mark in the text, word-level features are obtained to describe the distribution pattern of the text at the part of speech and punctuation level. Sentence-level features describing the syntactic complexity and structure of a text are obtained by extracting the average dependency tree depth, average dependency distance, and average theater. Document-level features describing text size and readability are obtained by statistically analyzing the total number of words, total number of characters, and number of sentences in the text. We use a large-scale sentiment model to extract the sentiment intensity of a text.

4. The cross-domain fake information detection method based on explicit and implicit common feature reinforcement learning according to claim 1, characterized in that, The second feature extraction unit extracts implicit semantic features of the domain, including: Each text is segmented and embedded to obtain a text embedding vector; The text embedding vector is input into a pre-trained language model for encoding to obtain a sequence of hidden text states. The hidden vector corresponding to the specified marker is extracted based on the hidden state sequence of the text, and the extracted hidden vector is used as the global semantic representation of the text.

5. The cross-domain fake information detection method based on explicit and implicit common feature reinforcement learning according to claim 1, characterized in that, The feature fusion unit is used to interactively fuse explicit domain common features and domain implicit semantic features, including: Explicit domain common features are used as queries, and implicit domain semantic features are used as keys. The features are mapped to the attention space through linear projection. A multi-head attention mechanism is used to evenly divide the projection vector according to the number of heads. The scaling dot product attention of each attention head is calculated, and the multi-head outputs are concatenated and integrated through linear projection to obtain the enhanced implicit domain semantic features. Using implicit semantic features of the domain as queries and explicit domain-shared features as keys, the features are mapped to the attention space through linear projection and information related to the implicit semantic features of the domain is mined through a multi-head attention mechanism to obtain enhanced explicit domain-shared features. The enhanced domain implicit semantic features and explicit domain shared features are concatenated, and the fusion weights are calculated through a gating network. The final fused features are obtained based on the weights. The gating network is used to dynamically balance the fusion ratio of the domain implicit semantic features and explicit domain shared features.

6. The cross-domain fake information detection method based on explicit and implicit common feature reinforcement learning according to claim 1, characterized in that, The model training process includes: The task of detecting false information is combined with domain adversarial training and supervised contrastive learning, and a joint loss function is constructed. The joint loss function is obtained by weighted fusion of false information detection training loss, domain adversarial training loss and supervised contrastive learning loss. In the domain adversarial training, domain-independent features are automatically learned through adversarial game between feature extractor and domain discriminator. In the supervised contrastive learning, an in-batch supervised sampling strategy is adopted and positive and negative sample pairs are constructed based on real label information to enhance the ability to distinguish between true and false information of common domain features through supervised contrastive learning. The model is trained and optimized based on the joint loss function and by utilizing the fusion features of multi-source domain sample data.

7. The cross-domain fake information detection method based on explicit and implicit common feature reinforcement learning according to claim 6, characterized in that, Automatically learn domain-independent features through an adversarial game between a feature extractor and a domain discriminator, including: In the forward propagation, the gradient inversion layer is used to pass the fused features to the neighborhood discriminator, so that the neighborhood discriminator can predict the source domain of the sample based on the fused features; In backpropagation, a gradient inversion layer is used to invert the gradient sign so that the feature extractor generates domain-independent features.

8. A cross-domain fake information detection system based on reinforcement learning of explicit and implicit common features, characterized in that, It includes: a sample acquisition module, a model training module, and an object detection module, among which, The sample acquisition module is used to acquire multi-source domain sample data, where each source domain sample corresponds to an information topic or event; The model training module is used to construct a fake information detection model and train the model using multi-source domain sample data. The fake information detection model includes: a first feature extraction unit for extracting explicit domain common features, a second feature extraction unit for extracting domain implicit semantic features, a feature fusion unit for interactively fusing explicit domain common features and domain implicit semantic features, and a fake information detection unit for outputting fake information detection based on the fused features. In the model training, the fused features are simultaneously input into the domain adversarial training branch, the supervised contrastive learning branch, and the fake information detection branch for joint reinforcement learning. The target detection module is used to input target domain information data into the trained fake information detection model, and use the fake information detection model to determine the authenticity of the target domain information data.

9. An electronic device, characterized in that, include: At least one processor, and a memory coupled to said at least one processor; The memory stores a computer program that can be executed by the at least one processor to implement the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed, enables the implementation of the method as described in any one of claims 1 to 7.

Images