False news detection method based on multi-modal feature adaptive fusion

Abstract

The invention discloses a false news detection method (MAFAN) based on multi-modal feature adaptive fusion, and the method comprises the steps: firstly, employing BERT to extract text features, employing ViT to extract image visual features, and designing a multi-branch feature extraction network based on wavelet feature enhanced convolution to extract the frequency domain information of an image; a double-stage fusion structure is adopted, firstly, text and frequency domain features are fused, and then the text and the frequency domain features are further fused with image features. Each stage is provided with a leading mode and an auxiliary mode to realize cross-mode interaction. A global gating mechanism is introduced to realize self-adaptive distribution of weights among modals, fusion weights are dynamically calculated according to semantic contribution of each modal feature, finally fusion features are self-adaptively generated through the gating mechanism, and dynamic adjustment of importance weights among the modals is realized. According to the mechanism, the adaptive capacity and the discrimination capacity of the model to multi-modal information are enhanced, and the accuracy and the robustness of false news detection are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of Internet information security and information detection technology, and in particular to a fake news detection method based on adaptive fusion of multimodal features. Background Art

[0002] In 2024, the China Internet Joint Rumor Refutation Platform received over 14,700 reports of online rumors, collected over 7,000 rumor samples and debunking data, and compiled over 21,000 debunking articles covering a wide range of hot topics both domestically and internationally. This phenomenon reflects the speed, breadth, and impact of disinformation, and its harmful effects on politics, public health, the economy, and other areas. In recent years, most news on social media has been accompanied by images. Compared to news with text alone, images not only convey more visual information but also enhance the emotional tone and credibility of the information. Fake news exploits this characteristic to attract and mislead readers. Such news contains false text accompanied by manipulated, misleading, or out-of-context images, resulting in semantic inconsistencies between the modalities. Analyzing solely from the perspective of images or text can easily lead to model classification errors. Multimodal approaches combine data from multiple modalities to provide a more comprehensive perspective. This approach can effectively address the limitations of single-modality approaches and more effectively identify key features of fake news, thereby improving the accuracy and reliability of fake news detection.

[0003] Reference 1 (CHEN J, WU Z, YANG Z, et al. Multimodal fusion network with latenttopic memory for rumor detection [C] / / 2021 IEEE International Conference on Multimedia and Expo (ICME). IEEE, 2021: 1-6.) uses cross-attention to assign corresponding weights to complementary modalities to achieve feature-level fusion, and introduces latent topic memory to store semantic information about rumor and non-rumor events, which is conducive to identifying upcoming posts. However, it is still challenging to fuse multiple modal features with complex heterogeneous relationships. Reference 2 (Wang Zhenyu, Zhu Xuefang. Research on fake news detection based on multimodal Transformer [J]. Journal of the China Society for Information Science, 2023, 42 (12): 1477-1486.) simplifies the processing of image input to a convolution-free method consistent with processing text input, unifies the feature extraction process of different modalities, and uses a cross-attention mechanism to achieve multimodal fusion. Reference 3 (Xie Xintong, Hu Yueyang, Liu Xinzhe, et al. Rumor detection method based on learning representative features of propagating users [J]. Computer Science and Technology, 2022, 16(6): 1334-1342.) conducted a statistical analysis on the distribution of user features in the dataset and selected representative user features, and then proposed an early rumor detection method based on learning representative features of propagating users.

[0004] Patent 1 (Sun Jiande, Shao Yi, Li Jing, et al. Multimodal Fake News Detection Method Based on Multi-Level Fusion and Attention Mechanism [P]. Shandong Province: CN202311702012.4, March 12, 2024) uses a pre-trained model to extract descriptive text and news text features from news images. These features are then fed into a hybrid heterogeneous expert network to obtain text fusion features and description fusion features, respectively. Finally, the fused features are concatenated with the image features and fed into a classifier for fake news classification. However, this simple concatenation approach to feature fusion fails to fully account for the complex correlations and importance differences between modalities. Patent 2 (Huang Qi, Li Bifan, Fu Yirui, et al. Fine-Grained Fake News Detection Method Based on Sentiment Distribution [P]. Jiangxi Province: CN202410832534.4, July 30, 2024) proposes a fake news detection method based on sentiment distribution, improving the accuracy of fake news detection by introducing a sentiment wheel model. While introducing new features is simple and effective, the model fails to fully learn the complementarity and consistency between modalities. Patent 3 (Dai Lican, Liu Xin, Sun Wen, et al. A multimodal fake news detection system and method based on a self-attention mechanism [P]. Sichuan Province: CN202410781447.0, 2024-09-10.) proposes a multimodal fake news detection method based on a self-attention mechanism. This method uses a multimodal feature extraction module to extract text, image, and video features, and uses self-attention to perform multimodal feature fusion. However, before performing cross-modal attention calculations on multimodal features, each modal feature must first undergo self-attention calculations, resulting in a high computational load during modal fusion.

[0005] Although the above methods perform well in the fake news detection task, they still have several shortcomings. On the one hand, although the introduction of new features can improve the detection accuracy of the model, this improvement is mainly limited to the single modality itself, and its contribution to the complementarity and consistency learning between multimodal features is limited. On the other hand, most existing models only use the visual information of the image, ignoring the multi-level frequency domain features contained in the image, resulting in a large amount of important task-related information being discarded. In addition, when processing multimodal information, current methods mostly use static fusion strategies, which makes it difficult to fully capture the complex correlation relationships between modalities. Cross-modal information interaction is limited, which in turn affects the overall detection performance. In contrast, this method introduces a multimodal adaptive fusion mechanism, which effectively enhances the collaborative modeling capabilities between modalities and improves the detection effect. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this paper proposes a fake news detection method based on multimodal feature adaptive fusion (MAFAN) to improve inter-modal interaction modeling capabilities and overall detection performance. This method first uses BERT to extract text features and ViT to extract image visual features. A multi-branch feature extraction network based on wavelet feature-enhanced convolution is designed to obtain frequency domain information from the image, achieving a multi-level representation of semantic, visual, and detail information. Subsequently, a two-stage multimodal feature adaptive fusion architecture (MAFA) is introduced to achieve cross-modal fusion in stages: the first stage fuses text and frequency domain features, and the second stage further fuses the preliminary fused features with image features. Each stage uses different modalities as both leading and supporting, enabling cross-modal information interaction. A global gating mechanism is introduced to adaptively allocate inter-modal weights. Fusion weights are dynamically calculated based on the semantic contribution of each modal feature. The final fused feature is adaptively generated through the gating mechanism, enabling dynamic adjustment of inter-modal importance weights. This mechanism enhances the model's adaptability and discriminative capabilities for multimodal information, significantly improving the accuracy and robustness of fake news detection.

[0007] 1. A method for detecting fake news based on adaptive fusion of multimodal features, comprising the following steps:

[0008] S1. Construct a multimodal fake news detection model framework. The model framework includes three processes: feature extraction, feature fusion, and fake news detection. Three sub-models are used to extract visual features, frequency domain features, and text features, respectively. The fused features are obtained through a multimodal feature adaptive fusion module, and finally fake news classification is performed.

[0009] S2. Use the text feature extractor to extract text semantic features and generate a fixed-length text embedding representation R T , use the image feature extractor to extract the visual features R in the image I , use the multi-scale feature extraction network to extract frequency domain features R F ;

[0010] S3. Integrate multimodal features through a two-stage fusion mechanism. First, use the multimodal feature adaptive fusion attention module to fuse text features and frequency domain features to generate a preliminary multimodal representation. It is then further integrated with visual features to generate a global multimodal feature representation

[0011] S4, the global multimodal features Input the news into the classifier for classification and use the softmax function to get the classification result.

[0012] 2. The method for detecting fake news based on adaptive fusion of multimodal features according to claim 1, wherein step S2 comprises the following steps:

[0013] S21. Global semantic feature representation of text T bert It can be directly generated by the BERT model, and then the dimension of the text feature is adjusted to d×1 (d is the text feature embedding dimension) through a fully connected layer with a ReLU activation function, expressed as R T :

[0014] R T =FC(W t T bert +b t )

[0015] Among them, FC is the fully connected layer, W t is the weight parameter of the fully connected layer, b t is the offset;

[0016] S22. Use the Vision Transformer (ViT) model to extract image visual features and design a classifier structure consisting of two layers of fully connected networks to perform dimensionality transformation to generate a d×1-dimensional visual feature representation R of the input image. I :

[0017] R I =FC(W i (ViT(image)+b i )

[0018] Among them, FC is the fully connected layer, W i is the weight parameter of the fully connected layer, b i is the offset;

[0019] S23. Use a multi-branch feature extraction network based on wavelet-enhanced Feature Extraction Convolution (WFEConv) to extract frequency domain features. First, build an initial feature extraction module consisting of a frequency domain convolution module consisting of a wavelet-enhanced feature convolution, a two-dimensional batch normalization layer (BatchNorm2d) and a ReLU activation function layer, and then downsample through maximum pooling (MaxPool2d):

[0020] x'=MaxPool(F ReLU (F BN (F WFEConv (x))))

[0021] Where x is the input image, F WFEConvrepresents wavelet feature enhanced convolution, F BN is batch normalization, F ReLU is the activation function, and the output x' is the feature map;

[0022] After obtaining the initial feature x', the multi-scale frequency domain feature response is obtained through the four feature extraction branches of Branch1, Branch2, Branch3, and Branch4. Branch1 uses 1×1 convolution; Branch2 uses 1×1 convolution followed by a 3×3 convolution; Branch3 uses 1×1 convolution followed by two 3×3 convolutions in series; Branch4 uses 3×3 maximum pooling followed by 1×1 convolution. All use the WFEConv module for convolution operations. The outputs of the four branches are spliced ​​in the channel dimension, expressed as:

[0023] x"=Concat(F B1 (x'),F B2 (x'),F B3 (x'),F B4 (x'))

[0024] Among them, F B1 (x'),F B2 (x'),F B3 (x'),F B4 (x') represents the output of the four branches respectively. Finally, the features are compressed by 1×1 convolution, and the frequency domain feature representation R is adjusted using the fully connected layer. F The spatial dimension of is d×1:

[0025] R F =FC(W f x”+b f )

[0026] Among them, FC is the fully connected layer, W f is the weight parameter of the fully connected layer, b f is the offset.

[0027] 3. The method for detecting fake news based on adaptive fusion of multimodal features according to claim 1, wherein step S3 comprises the following steps:

[0028] S31. For the multimodal features obtained in S2, the multimodal feature fusion attention module (MFA) is used to efficiently integrate text, image and frequency domain features. The multimodal feature fusion attention consists of a single-head attention and a feedforward network, both of which contain a residual connection. Single-head attention is used for cross-modal attention calculation to achieve feature fusion. The calculation formula is as follows:

[0029]

[0030] b rel is the relative position encoding, τ(R) is the input sequence length, and the attention function f is ReLU 2 ;

[0031] S32, the present invention adopts dual-path MFA to cross-fuse the modalities, and the first stage realizes text R T and frequency domain R F Cross-fusion, stage two realizes the preliminary fusion feature R TF With image R I In each stage, different modal features are used as the source of Q, K, and V to achieve cross-modal information interaction. The fusion process is as follows:

[0032] R TF =MFA(R T ,R F )

[0033] R TFI =MFA(R TF ,R I )

[0034] In order to achieve adaptive modal feature fusion, two global gating modules are introduced to calculate the weight coefficients of inter-modal fusion respectively:

[0035] α=Softmax(MLP([R T ; R F ]))

[0036] β=Softmax(MLP([R TF ; R I ]))

[0037] Final fusion features It is obtained by the following formula, thus realizing the adaptive fusion of multimodal features:

[0038]

[0039] Among them, α1, α2, β1, β2∈[0,1], and α1+α2=1, β1+β2=1, respectively represent the importance weight of each modality in the gating mechanism. This mechanism enables the model to have the ability to dynamically adjust between modalities during the fusion process, reflecting its sensitivity and adaptability to the contribution of different modalities.

[0040] 4. The method for detecting fake news based on adaptive fusion of multimodal features according to claim 1, wherein step S4 comprises the following steps:

[0041] S41. According to the feature fusion module, multimodal fusion feature representation of text, image and frequency domain can be obtained. Then it will Input into the classifier to get the classification result:

[0042]

[0043] Among them, MLP represents the classifier, is a probability value between 0 and 1, and the model uses a binary cross entropy loss function:

[0044]

[0045] Among them, y∈{0,1} is the true label, is the model's predicted probability for label 1.

[0046] Beneficial effects:

[0047] This paper proposes an end-to-end fake news detection model that leverages three types of information: text, visual images, and frequency domains. It can identify fake news on social media solely based on text content and accompanying images, without the need for additional information or auxiliary tasks. By integrating multimodal information, it effectively improves detection accuracy and overcomes the limited expressive power of a single modality when dealing with complex false information.

[0048] 2. To fully utilize the multi-level frequency domain information contained in images, this paper constructs a frequency domain feature extraction network based on wavelet feature enhanced convolution (WFEConv). This network combines wavelet transforms with a multi-branch feature extraction mechanism to capture image details and structural information at different scales and frequencies, enhancing the model's ability to recognize textures, edges, and abnormal regions. This method effectively improves the quality of frequency domain feature expression, providing richer and more detailed feature support for fake news detection.

[0049] 3. In terms of multimodal fusion, the present invention proposes a feature adaptive fusion method based on a global gating mechanism. By designing a two-stage modal cross-fusion structure (MAFA), the cross-fusion of text and frequency domain features, as well as the re-fusion of preliminary fusion features and visual information are achieved respectively. A global gating mechanism is introduced in the fusion process to dynamically assign the importance weights of each modality, achieve adaptive collaboration and information complementarity between modalities, and significantly improve the expressive power of the fusion and the robustness of fake news detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 This is an overall structural diagram of a fake news detection method based on adaptive fusion of multimodal features used in an embodiment of the present invention;

[0051] Figure 2 A structural diagram of the multi-branch frequency domain feature extraction network of a fake news detection method based on multimodal feature adaptive fusion used in an embodiment of the present invention;

[0052] Figure 3 The multimodal fusion attention module and single-head attention mechanism structure diagram of a fake news detection method based on multimodal feature adaptive fusion used in an embodiment of the present invention;

[0053] Figure 4 The training method for a fake news detection method based on adaptive fusion of multimodal features used in an embodiment of the present invention is implemented;

[0054] Figure 5 This is a visualization effect diagram of the fusion features of a fake news detection method based on adaptive fusion of multimodal features used in an embodiment of the present invention. DETAILED DESCRIPTION

[0055] To provide a clearer understanding of the technical features, objectives, and beneficial effects of the present invention, an embodiment of the present invention is further described with reference to the accompanying drawings. The embodiment is intended only to further illustrate the present invention and is not to be construed as limiting the scope of protection of the present invention. Non-essential improvements and adjustments made by those skilled in the art based on the contents of the present invention also fall within the scope of protection of the present invention.

[0056] 1. A method for detecting fake news based on adaptive fusion of multimodal features, comprising the following steps:

[0057] S1. Construct a multimodal fake news detection model framework, whose structure is as follows Figure 1 As shown in the figure, the model framework includes three processes: feature extraction, feature fusion, and fake news detection. Three sub-models are used to extract visual features, frequency domain features, and text features respectively. The fused features are obtained through the multimodal feature adaptive fusion module, and finally fake news classification is performed.

[0058] S2. Use the text feature extractor to extract text semantic features and generate a fixed-length text embedding representation R T , use the image feature extractor to extract the visual features R in the image I , use the multi-scale feature extraction network to extract frequency domain features R F ;

[0059] S3. Integrate multimodal features through a two-stage fusion mechanism. First, use the multimodal feature adaptive fusion attention module to fuse text features and frequency domain features to generate a preliminary multimodal representation. It is then further integrated with visual features to generate a global multimodal feature representation

[0060] S4, the global multimodal features Input the news into the classifier for classification and use the softmax function to get the classification result.

[0061] 2. The method for detecting fake news based on adaptive fusion of multimodal features according to claim 1, wherein step S2 comprises the following steps:

[0062] S21. Global semantic feature representation of text T bert It can be directly generated by the BERT model, and then the dimension of the text feature is adjusted to d×1 (d is the text feature embedding dimension) through a fully connected layer with a ReLU activation function, expressed as R T :

[0063] R T =FC(W t T bert +b t )

[0064] Among them, FC is the fully connected layer, W t is the weight parameter of the fully connected layer, b t is the offset;

[0065] S22. Use the Vision Transformer (ViT) model to extract image visual features and design a classifier structure consisting of two layers of fully connected networks to perform dimensionality transformation to generate a d×1-dimensional visual feature representation R of the input image. I :

[0066] R I =FC(W i (ViT(image)+b i )

[0067] Among them, FC is the fully connected layer, W i is the weight parameter of the fully connected layer, b i is the offset;

[0068] S23, use the multi-branch feature extraction network based on wavelet-enhanced Feature ExtractionConvolution (WFEConv) to extract frequency domain features, its structure is as follows Figure 2As shown in the figure, we first build an initial feature extraction module consisting of a frequency domain convolution module consisting of a wavelet feature enhancement convolution, a two-dimensional batch normalization layer (BatchNorm2d) and a ReLU activation function layer, and then downsample through maximum pooling (MaxPool2d):

[0069] x'=MaxPool(F ReLU (F BN (F WFEConv (x))))

[0070] Where x is the input image, F WFEConv represents wavelet feature enhanced convolution, F BN is batch normalization, F ReLU is the activation function, and the output x' is the feature map;

[0071] After obtaining the initial feature x', the multi-scale frequency domain feature response is obtained through the four feature extraction branches of Branch1, Branch2, Branch3, and Branch4. Branch1 uses 1×1 convolution; Branch2 uses 1×1 convolution followed by a 3×3 convolution; Branch3 uses 1×1 convolution followed by two 3×3 convolutions in series; Branch4 uses 3×3 maximum pooling followed by 1×1 convolution. All use the WFEConv module for convolution operations. The outputs of the four branches are spliced ​​in the channel dimension, expressed as:

[0072] x"=Concat(F B1 (x'),F B2 (x'),F B3 (x'),F B4 (x'))

[0073] Among them, F B1 (x'),F B2 (x'),F B3 (x'),F B4 (x') represents the output of the four branches respectively. Finally, the features are compressed by 1×1 convolution, and the frequency domain feature representation R is adjusted using the fully connected layer. F The spatial dimension of is d×1:

[0074] R F =FC(W f x”+b f )

[0075] Among them, FC is the fully connected layer, W f is the weight parameter of the fully connected layer, b f is the offset.

[0076] 3. The method for detecting fake news based on adaptive fusion of multimodal features according to claim 1, wherein step S3 comprises the following steps:

[0077] S31, for the multimodal features obtained in S2, use the multimodal feature fusion attention module (MFA) to efficiently integrate text, image and frequency domain features. Its structure is as follows Figure 3 As shown in (a), the multimodal feature fusion attention consists of a single-head attention and a feedforward network, both of which contain a residual connection. The single-head attention is used for cross-modal attention calculation to achieve feature fusion. The single-head attention structure is as follows Figure 3 (b) is shown in the figure, and the calculation formula is as follows:

[0078]

[0079] b rel is the relative position encoding, τ(R) is the input sequence length, and the attention function f is ReLU 2 ;

[0080] S32, the present invention adopts dual-path MFA to cross-fuse the modalities, and the first stage realizes text R T and frequency domain R F Cross-fusion, stage two realizes the preliminary fusion feature R TF With image R I In each stage, different modal features are used as the source of Q, K, and V to achieve cross-modal information interaction. The fusion process is as follows:

[0081] R TF =MFA(R T ,R F )

[0082] R TFI =MFA(R TF ,R I )

[0083] In order to achieve adaptive modal feature fusion, two global gating modules are introduced to calculate the weight coefficients of inter-modal fusion respectively:

[0084] α=Softmax(MLP([R T ; R F ]))

[0085] β=Softmax(MLP([R TF ; R I ]))

[0086] Final fusion features It is obtained by the following formula, thus realizing the adaptive fusion of multimodal features:

[0087]

[0088] Among them, α1, α2, β1, β2∈[0,1], and α1+α2=1, β1+β2=1, respectively represent the importance weight of each modality in the gating mechanism. This mechanism enables the model to have the ability to dynamically adjust between modalities during the fusion process, reflecting its sensitivity and adaptability to the contribution of different modalities.

[0089] 4. The method for detecting fake news based on adaptive fusion of multimodal features according to claim 1, wherein step S4 comprises the following steps:

[0090] S41. According to the feature fusion module, multimodal fusion feature representation of text, image and frequency domain can be obtained. Then it will Input into the classifier to get the classification result:

[0091]

[0092] Among them, MLP represents the classifier, is a probability value between 0 and 1, and the model uses a binary cross entropy loss function:

[0093]

[0094] Among them, y∈{0,1} is the true label, is the model's predicted probability for label 1.

[0095] Simulation experiment

[0096] The proposed method compares its performance with competing models on two widely used multimodal fake news detection datasets, Weibo and Twitter. Accuracy, precision, recall, and F1-score, commonly used in classification tasks, were selected as performance evaluation metrics. Table 1 shows that the proposed MAFAN model surpasses the competing models in accuracy on both the Weibo and Twitter datasets and also achieves excellent results on other evaluation metrics. This further validates the effectiveness and robustness of MAFAN in fake news detection and demonstrates its significant advantages in multimodal information fusion and feature representation.

[0097] Table 1 Model comparison results on Weibo and Twitter datasets

[0098]

[0099] In order to further explore the role of the feature fusion layer, the multimodal feature distributions obtained by MAFAN and MAFAN-M without the feature fusion layer on the Weibo test set are mapped to a two-dimensional space using the t-SNE algorithm and visualized on a two-dimensional coordinate graph. Figure 5 As shown, the MAFAN-M model exhibits significant overlap between features from different labels, making it difficult to distinguish between true and false news in areas where both real and false news are mixed. In contrast, the features extracted by MAFAN are more compactly distributed, with clear boundaries between different labels, effectively reducing feature overlap. This suggests that the multimodal fusion method proposed in this paper can learn more accurate feature representations.

[0100] The above describes the method of the present invention. Those skilled in the art can implement the method of the present invention based on the description of this content. Based on the above content of the present invention, other embodiments obtained by those skilled in the art without making any creative work should fall within the scope of protection of the present invention.

Claims

1. A fake news detection method based on adaptive fusion of multimodal features, characterized by: The following steps are involved: S1. Construct a multimodal fake news detection model framework. The model framework includes three processes: feature extraction, feature fusion, and fake news detection. Three sub-models are used to extract visual features, frequency domain features, and text features, respectively. The fused features are obtained through a multimodal feature adaptive fusion module, and finally fake news classification is performed. S2. Use the text feature extractor to extract text semantic features and generate a fixed-length text embedding representation R T , use the image feature extractor to extract the visual features R in the image I , use the multi-scale feature extraction network to extract frequency domain features R F ; S3. Integrate multimodal features through a two-stage fusion mechanism. First, use the multimodal feature adaptive fusion attention module to fuse text features and frequency domain features to generate a preliminary multimodal representation. It is then further integrated with visual features to generate a global multimodal feature representation S4, the global multimodal features Input the news into the classifier for classification and use the softmax function to get the classification result.

2. The method for detecting fake news based on adaptive fusion of multimodal features according to claim 1, characterized in that: The step S2 comprises the following steps: S21. Global semantic feature representation of text T bert It can be directly generated by the BERT model, and then the dimension of the text feature is adjusted to d×1 (d is the text feature embedding dimension) through a fully connected layer with a ReLU activation function, expressed as R T : R T =FC(W t T bert +b t ) Among them, FC is the fully connected layer, W t is the weight parameter of the fully connected layer, b t is the offset; S22. Use the Vision Transformer (ViT) model to extract image visual features and design a classifier structure consisting of two layers of fully connected networks to perform dimensionality transformation to generate a d×1-dimensional visual feature representation R of the input image. I : R I =FC(W i (ViT(image)+b i ) Among them, FC is the fully connected layer, W i is the weight parameter of the fully connected layer, b i is the offset; S23. Use a multi-branch feature extraction network based on wavelet-enhanced Feature Extraction Convolution (WFEConv) to extract frequency domain features. First, build an initial feature extraction module consisting of a frequency domain convolution module consisting of a wavelet-enhanced feature extraction convolution, a two-dimensional batch normalization layer (BatchNorm2d), and a ReLU activation function layer: x'=F ReLU (F BN (F WFEConv (x))) Where x is the input image, F WFEConv represents wavelet feature enhanced convolution, F BN is batch normalization, F ReLU is the activation function, and the output x' is the feature map, which is then downsampled through maximum pooling (MaxPool2d). Four feature extraction branches, Branch1, Branch2, Branch3, and Branch4, are set to obtain multi-scale frequency domain feature responses. All of them use the WFEConv module for convolution operations. Branch1 uses 1×1 convolution; Branch2 uses 1×1 convolution followed by a 3×3 convolution; Branch3 uses 1×1 convolution followed by two 3×3 convolutions in series; Branch4 uses 3×3 maximum pooling followed by 1×1 convolution. The outputs of the four branches are spliced ​​in the channel dimension, which is expressed as: x”=Concat(F B1 (x),F B2 (x),F B3 (x),F B4 (x)) Among them, F B1 (x),F B2 (x),F B3 (x),F B4 (x) represents the output of the four branches respectively. Finally, the features are compressed by 1×1 convolution, and the frequency domain feature representation R is adjusted using the fully connected layer. F The spatial dimension of is d×1: R F =FC(W f x”+b f ) Among them, FC is the fully connected layer, W f is the weight parameter of the fully connected layer, b f is the offset.

3. The method for detecting fake news based on adaptive fusion of multimodal features according to claim 1, characterized in that: The step S3 comprises the following steps: S31. For the multimodal features obtained in S2, the multimodal feature fusion attention module (MFA) is used to efficiently integrate text, image and frequency domain features. The multimodal feature fusion attention consists of a single-head attention and a feedforward network, both of which contain a residual connection. Single-head attention is used for cross-modal attention calculation to achieve feature fusion. The calculation formula is as follows: b rel is the relative position encoding, τ(R) is the input sequence length, and the attention function f is ReLU 2 ; S32, using dual-path MFA to cross-fuse modalities, the first stage realizes text R T and frequency domain R F Cross-fusion, stage two realizes the preliminary fusion feature R TF With image R I In each stage, different modal features are used as the source of Q, K, and V to achieve cross-modal information interaction. The fusion process is as follows: R TF =MFA(R T ,R F ) R TFI =MFA(R TF ,R I ) In order to achieve adaptive modal feature fusion, two global gating modules are introduced to calculate the weight coefficients of inter-modal fusion respectively: α=Softmax(MLP([R T ;R F ])) β=Softmax(MLP([R TF ;R I ])) Final fusion features It is obtained by the following formula, thus realizing the adaptive fusion of multimodal features: Among them, α1, α2, β1, β2∈[0,1], and α1+α2=1, β1+β2=1, respectively represent the importance weight of each modality in the gating mechanism. This mechanism enables the model to have the ability to dynamically adjust between modalities during the fusion process, reflecting its sensitivity and adaptability to the contribution of different modalities.

4. The method for detecting fake news based on adaptive fusion of multimodal features according to claim 1, characterized in that: The step S4 comprises the following steps: S41. According to the feature fusion module, multimodal fusion feature representation of text, image and frequency domain can be obtained. Then it will Input into the classifier to get the classification result: Among them, MLP represents the classifier, is a probability value between 0 and 1, and the model uses a binary cross entropy loss function: Among them, y∈{0,1} is the true label, is the model's predicted probability for label 1.

Images