Detecting Fake News on Social Media via Multimodal Semantic Understanding and Enhanced Transformer Architectures

On social media platforms, news content is typically presented in a combination of images and text, with the images and text complementing each other to convey information. In the spread of fake news, images and text often work together, and fake information may be more deceptive and misleading through the combination of images and text. Therefore, relying solely on text analysis or image recognition to detect fake news often fails to achieve ideal results. To address this, this paper conducts research on image-text matching for social media news based on multimodal semantic understanding, aiming to effectively combine image and text information and fully explore the semantic correlations between them, thereby improving the accuracy of fake news detection. Image-text matching research helps to understand the interaction and semantic consistency between images and text in news reports, and by comparing the semantics of the images and text in the news content, it determines whether there are contradictions or misleading information, thus assessing the authenticity of the news.

The model constructed is based on an end-to-end image-text matching architecture, with a focus on improving fake news recognition by extracting key samples and analyzing their semantic features. To effectively extract the key semantics of the samples, this research adopts a sample diversion strategy based on Gaussian Mixture Models, filtering the image-text data to retain the core semantic information of each sample. The training of the model is divided into two stages. In the first stage, the model predicts the negative impact of key samples and evaluates their influence on fake news detection. In the second stage, the image-text matching model is formally trained by combining key semantics and negative impact. This phased training method enables the model to gradually optimize its image-text matching ability and improve the accuracy of fake news recognition.

In the specific implementation of image-text matching, this paper introduces a Siamese network structure to ensure that the image-text matching learning process is more stable and consistent. In each training step, the Siamese network X and X' models share parameters, ensuring the consistency of the structure and internal parameters of the two models, which is crucial for handling the semantic associations in image-text information. During the training process, the X' model is first pre-trained using the original image-text data to compute the matching loss function and predict the negative impact of each sample, i.e., the weights. These weights reflect the importance of the sample in fake news detection, further helping the model focus on key samples for learning. In the second stage, the X model retrains using the original image-text data, key samples, and negative impact weights, ultimately obtaining the optimized image-text matching model.

2.1 Image-text feature extraction and similarity measurement

In the process of feature extraction, this paper uses the pre-trained model CLIP as the baseline model. CLIP consists of two independent Transformer encoders, which are used to process visual and textual information, respectively. The visual encoder $d(\cdot)$ uses the pre-trained Vision Transformer encoder ViTB/32, which adds a layer of normalization before output to enhance the stability and accuracy of feature extraction. The text encoder $h(\cdot)$ consists of 12 layers of Transformer structures with a width of 512, equipped with 8 attention heads, to capture the semantic details in the text. Both encoders were pre-trained on the large-scale WIT dataset, ensuring their strong feature extraction capability. In specific applications, the visual encoder $h(\cdot)$ extracts high-dimensional visual features from the image, while the text encoder $h(\cdot)$ extracts high-dimensional textual features from the text. These features will be used in subsequent steps to calculate the similarity between the image and text. Let $F=(U, S)$ represent a social media news image-text dataset, and the similarity calculation between modalities can be represented as follows:

$T\left(U_u, S_k\right)=\frac{d\left(U_u\right) \cdot h\left(S_k\right)}{\left\|d\left(U_u\right)\right\| \times\left\|h\left(S_k\right)\right\|}$                      (1)

This paper uses cross-entropy loss to measure the matching degree between image-text features. During training, the image and text are first input into the visual encoder and text encoder, respectively, to obtain the corresponding feature vectors. Then, the cosine similarity between the image and text feature vectors is calculated to evaluate the proximity of the image-text pair in the semantic space. Suppose the image or text is represented by $a_u$ and $b_u$, with a batch size denoted by $v$, then we have:

$\begin{aligned} & Z R\left(a_u, b_u\right)=-\log \left(\frac{\exp \left(T\left(a_u, b_u\right)\right)}{\sum_{k=1}^v \exp \left(T\left(a_u, b_u\right)\right)}\right) \\ & \operatorname{LOSS}_{Z R}=\frac{1}{v} \sum_{u=1}^v\left(Z R\left(U_u, S_u\right)+Z R\left(S_u, U_u\right)\right)\end{aligned}$                      (2)

2.2 Key semantic retention

In the task of social media image fake news detection, the image-text matching model faces complex challenges, including the interference of noise samples and hard samples. Noise samples are those where the semantics between the image and text are completely inconsistent, while hard samples refer to those where there is partial inconsistency in the semantic information between the image and text. When these noise and hard samples are included during the training process, the model may overfit these erroneous samples, leading to deviations in the prediction of the actual image-text correspondence. According to the learning characteristics of deep neural networks (DNN), the model tends to first learn and memorize the obvious correspondences in simple samples, while for complex hard samples and noise samples, it may produce erroneous memories, thus affecting the overall model's performance. In image-text matching tasks, erroneous memories will make matching between some image-text pairs more difficult, especially when these erroneous samples have a high similarity to other image-text samples, making the model more susceptible to interference and resulting in incorrect fake news judgments. Therefore, to prevent the negative impact of hard and noise samples on the training process, retaining the key semantic information of simple samples helps the model focus better on those image-text pairs with semantic consistency and strong discriminative power, thus effectively improving the accuracy of fake news detection.

The model identifies simple and hard samples by calculating the loss function value of the sample. Since DNNs tend to learn simple samples first, the loss value of simple samples is usually low, while the loss value of hard and noise samples is high. Based on this characteristic, this paper uses a Gaussian Mixture Model (GMM) to fit the loss distribution of the training dataset and uses the difference in sample loss to identify which samples are simple. This step provides the foundation for subsequent semantic retention. By analyzing the loss distribution, the model can accurately filter out those simple samples with clear semantics and low loss, which will be used as the core of key semantics. Let the mixture coefficient be represented by $x_j$, and the probability density of the $j$-th component be represented by $\psi\left(c\mid\phi_j\right)$, then we have:

$o(c \mid \varphi)=\sum_{j=1}^J x_j \psi\left(c \mid \varphi_j\right)$                   (3)

Assume that the component with the smaller mean in the two-component Gaussian Mixture Model is represented by $\phi_j$. The probability that the $u$-th image-text pair is correctly matched is the posterior probability $o_u$, and the calculation process is as follows:

$o_u=o\left(\varphi_j \mid c_u\right)=\frac{o\left(\varphi_j\right) o\left(c_u \mid \varphi_j\right)}{o\left(c_u\right)}$                     (4)

Next, the model uses the selected simple sample set $F_z$ to construct the key semantic library of the training set. In this process, the model matches each training sample pair ($U_u, S_u$) with the image and text that have the highest semantic similarity. This similarity measurement is performed by calculating the visual similarity $T$ between the image $U_u$ and all images in $F_z$, selecting the image $U^U{ }_u$ that is most similar to $U_u$, and pairing the corresponding text $S^U{ }_u$ with the original text $S_u$. In this way, the model can use $U_u$ and $S^U{ }_u$ as key semantic pairs to enhance the visual modal semantic information of the original image $U_u$. This process ensures the semantic consistency between the visual information and the corresponding text information for each sample, thereby improving the model's ability to understand the visual content. The threshold is represented by $\pi$. The selection process is as follows:

$F_z=\left\{\left(U_k, S_k\right) \mid o_k \geq \pi\right\}$                   (5)

In the semantic retention of the text modality, the model also calculates the similarity $T_s$ between the text $S_u$ and all texts in $F_z$, and selects the text $U^S{ }_u$ that is most similar to $T_s$. In this way, the textual information between the original text $S_u$ and $U^{\mathcal{S}}{ }_u$ can be strengthened, ensuring a more accurate semantic understanding of the text modality. By mutually enhancing the text and image modalities, the model can better learn the semantic association between image and text, which is crucial for fake news detection. This is because fake news on social media often exhibits semantic mismatches, and the model needs to be able to recognize and correct such mismatches. Through this bidirectional enhancement, the model can not only better understand the independent semantics of images and texts, but also find deep connections between them, thus improving the accuracy of matching.

2701.png

Figure 1. Schematic diagram of the construction process of the key semantic library

Figure 1 shows the schematic diagram of the construction process of the key semantic library. Finally, the constructed key semantic library $L$ contains the key image-text pairs selected from the simple sample set. These pairs are used to enhance the visual and textual semantic information of the original samples during training. Let $L=\left\{\left(U^U{ }_u, S^U{ }_u\right),\left(U^{\mathcal{S}}{ }_u, S^{\mathcal{S}}{ }_u\right)\right\}$ represent the enhanced semantic information on the visual and text modalities for each training sample pair, and this information helps the model better capture the subtle differences between images and texts. Through this key semantic retention strategy, the model not only improves the robustness of fake news detection but also enhances the deep understanding of social media news content. Especially on social media platforms, the authenticity of news content often involves complex semantic changes, and the construction of the key semantic library effectively enhances the model's ability to handle complex image-text matching tasks, thereby helping to more accurately identify fake news.

2.3 Model training method

The optimization of the constructed model is divided into two stages, aiming to reasonably evaluate the impact of sample pairs on the model and adjust accordingly, thereby improving the accuracy of fake news detection. The goal of the first stage is to predict the negative impact of each sample pair on the model’s performance and calculate the influence factor of each sample. This stage focuses on identifying which image-text pairs interfere with the model’s training, especially those hard samples and noise samples related to fake news. The goal of the second stage is to optimize the main model based on the influence factors calculated in the first stage. In this stage, the model will adjust the training process with weighted samples, assigning higher weights to samples with smaller impacts to help the model better learn the semantic information from key samples.

2702.png

Figure 2. Model training strategy diagram

The main reason for predicting the negative impact of samples during model training lies in the memory characteristics of DNNs. In the learning process, DNNs tend to prioritize learning the obvious patterns in simple samples, while for hard samples, they are prone to learning incorrect correspondences. These erroneous learning results not only directly affect the model’s prediction ability for hard samples, but may also negatively impact other samples that should have been correctly predicted. Specifically, when the model during training matches noise samples with the most similar simple samples in the key semantic library, these simple samples should play a positive role in the training. However, if the model learns incorrect matching relationships, it will reduce the loss of these simple samples, which in turn increases the loss of these simple samples, thereby affecting the model’s prediction of other correct samples. Figure 2 shows a schematic diagram of the model training strategy.

The basic idea of training is to introduce an independent copy model $X^{\prime}$, which shares all parameters with the main model $X$ but remains independent during updates. The core purpose of this design is to predict the negative impact each training sample has on the model by calculating the difference in loss values of the simple sample set $L_y$ before and after training for both model $X$ and $X^{\prime}$. The samples in the simple sample set $L_y$ generally have obvious semantic alignment and are more sensitive to fluctuations in model performance. Therefore, by analyzing the loss value changes of these samples, the model can indirectly assess whether there is interference from noise or difficult samples during training. Specifically, the loss value for image-text matching is represented by $o_i$, the loss value for text-image matching is represented by $w_j$, and $X_t^{\prime}$ represents the model before the $t$-th round of training. The loss for key samples before training is denoted as $o^t{ }_j$ and $w^t{ }_j$, and after updating the parameters, the model $X_{t+1}^{\prime}$ is obtained. The loss for key samples after training is represented as $o^{t+1}{ }_j$ and $w^{t+1}{ }_j$. For the training sample ($U_j$, $S_j$) corresponding to the key samples ($U^{\mathcal{U}}{ }_j$, $S^{\mathcal{U}}{ }_j$) and ($U^j{ }_j$, $S^S{ }_j$), the loss calculation process between them is as follows:

$\begin{aligned} & o_j=Z R\left(U_j^U, S_j^U\right)+Z R\left(U_j^S, S_j^S\right) \\ & w_j=C E\left(S_j^U, U_j^U\right)+Z R\left(S_j^S, U_j^S\right)\end{aligned}$                    (6)

Next, by calculating the impact of the training sample ($U_j$, $S_i$) on the model performance, the loss change caused by this sample on the model can be quantified. In this process, the loss value change of the key samples ($U^U{ }_j$, $S^U{ }_j$) and ($U^s{ }_j$, $S^S{ }_j$) directly reflects the impact of the training sample ($U_j$, $S_j$) on the model performance. Specifically, if ($U_j$, $S_j$) is a difficult-to-predict noise sample, it may lead to a negative impact on the predictions of the key samples ($U^{\mathcal{U}}{ }_j$, $S^{\mathcal{U}}{ }_j$) and ($U^s{ }_j$, $S^S{ }_j$), causing their loss values to increase. In contrast, if ($U_j$, $S_j$) is an easily predictable simple sample, its impact should be positive, causing the loss values of the key samples to decrease. Specifically, the model performance change caused by ($U_j$, $S_j$) can be characterized by the change in the key sample loss values as follows:

$e_j=\frac{1}{2}\left(\frac{o_j^t}{o_j^{t+1}}+\frac{w_j^t}{w_j^{t+1}}\right)$                     (7)

When $e_j<1$, it indicates that the loss value of the key sample has increased; otherwise, when $e_j \geq 1$, it indicates that the sample has no negative impact. During the model training process, this paper quantifies the negative impact of each sample pair ($U_j$, $S_j$) on the model performance by calculating the influence factor $q_j$. The influence factor $q_j$ reflects the extent of the change in the loss values of the key samples caused by the training sample ($U_j$, $S_j$), thereby predicting the potential negative impact of this sample on the model. Specifically, the influence factor $q_j$ evaluates the interference with the model learning process caused by the loss change induced by the training sample ($U_j$, $S_j$). If the $q_j$ value is large, it indicates that the sample may severely affect the model's prediction performance, and the model will take corresponding measures to reduce its weight; if the $q_j$ value is small, it means the sample has little impact on the model's training and can continue to participate in training.

$q_j= \begin{cases}\tanh \left(e_j\right) & , e_j<1 \\ 1 & , \text { Others }\end{cases}$                        (8)

The principle of the retraining module is first reflected in retraining the main model $X$ and introducing a weighted crossentropy loss $L O S S_{Q Z D}$ to prevent training samples from causing performance degradation in the model. The influence factor $q_j$ of each training sample reflects the degree to which the sample affects the model's performance, and based on these influence factors, weights are assigned to the samples during training. Those samples with larger negative impacts will have lower weights during updates, thereby reducing their interference with the model. The calculation process for $L O S S_{Q Z D}$ is:

$\operatorname{LOSS}_{Q Z R}=\frac{1}{v} \sum_{j=1}^v q_j\left(Z R\left(U_j, S_j\right)+Z R\left(S_j, U_j\right)\right)$                        (9)

Furthermore, since the image-text correspondences in difficult and noisy samples are often unreliable, relying solely on these samples for learning may lead the model to learn incorrect semantic relationships. To enhance the model's understanding of these unreliable semantic relationships, this paper designs an auxiliary learning mechanism using the cross-entropy loss of key samples, $L O S S_L$. Key samples are those selected from the training set that have high semantic matching with the original samples, and they provide stable and correct semantic alignment in both visual and textual modalities. The calculation process for $L O S S_L$ is:

$\operatorname{LOSS}_L=\operatorname{LOSS}_{Z R}\left(U_j^U, S_j^U\right)+\operatorname{LOSS}_{Z R}\left(U_j^S, S_j^S\right)$                    (10)

The total loss function is as follows:

$L O S S=L O S S_{Q Z R}+L O S S_L$                      (11)                 

From the experimental results, the model proposed in this paper performs excellently in the social media news image-text matching task, especially in the two directions of "Image-to-Text Matching" and "Text-to-Image Matching." Specifically, the model achieves an R@1 of 81.2 in the "Image-to-Text Matching" task, significantly higher than other models, such as Self-Attention (78.9) and Stacked Cross Attention (77.3), demonstrating its strong ability in image and text matching. In the "Text-to-Image Matching" task, the model also achieves an R@1 of 67.9, clearly outperforming Perceptual Digital Quotient (55.6) and TMK+PDQF (61.2), indicating its efficiency in matching images and text. In addition, the model also outperforms most other models in the R@5 and R@10 metrics, especially in the text-to-image matching task, where it reaches R@5 and R@10 values of 93.5 and 97.5, demonstrating high accuracy and robustness (Table 1).

Table 1. Comparison results of different models for social media news image-text matching

Table 2. Ablation experiments of the proposed social media news image-text matching model

270501.png	270502.png
1) Training set	2) Test set

Figure 5. Comparison of simple sample set composition at different thresholds

From the ablation experiment results in Table 2, the proposed social media news image-text matching model demonstrates stable performance and strong advantages under different experimental setups. By comparing the different combinations of Siamese network structure and key semantic retention mechanism, it can be observed that models incorporating both show performance improvements across all three datasets. For example, on the Image Verification Corpus dataset, after adding the Siamese network structure and key semantic retention, the model achieves the highest R@1, R@5, and R@10 values in both image-to-text and text-to-image matching tasks, reaching 86.5, 96.5, 97.5 and 71.5, 91.2, 94.2, respectively. This is significantly better compared to other combinations where some factors are removed, such as using only the Siamese network or only retaining key semantics. A similar trend is observed on the Weibo Dataset and Top News Dataset, particularly on the Weibo Dataset, where the joint use of the Siamese network and key semantic retention increases the R@1 value to 78.9, compared to other combinations, such as using only the Siamese network (78.5) or only retaining key semantics (78.5), demonstrating its strong fusion effect.

From Figure 5, it can be seen that with different thresholds on the training and test sets, the changes in correct and incorrect correspondences show a clear trend. In the training set, as the threshold gradually increases from 0.1 to 0.99, the proportion of correct correspondences increases from 0.995 to 0.9995, indicating that as the threshold increases, the model's matching accuracy on the training data improves continuously. Meanwhile, the proportion of incorrect correspondences decreases from 0.005 to 0.0005, showing that the model gradually reduces the probability of incorrect matching. Correspondingly, on the test set, the proportion of incorrect correspondences increases with the threshold, from 0.981 to 0.999, reflecting that the model begins to make more incorrect matches on the test data. At the same time, the proportion of correct correspondences decreases from 0.019 to 0.004, suggesting that as the threshold increases, the model's generalization ability weakens, leading to a reduction in the number of correctly matched samples. Based on these results, it can be concluded that the setting of the threshold plays an important role in the performance and generalization ability of multimodal learning models. In the training set, a lower threshold means the model accepts more samples, leading to higher accuracy, but in the test set, a higher threshold, while improving matching accuracy on the training set, leads to an increase in incorrect correspondences on the test set. This indicates that, in the task of social media image fake news detection, to optimize the model's performance, it is necessary to balance the threshold setting during both training and testing. A threshold that is too high may lead to overfitting, thus affecting the model's accuracy on unknown data.

From the experimental results shown in Table 3, the model proposed in this paper performs very well in the task of social media image fake news detection, especially in terms of accuracy (Acc.) and the F1 score for different types of fake news detection. For example, the overall accuracy of the proposed model reaches 0.912, far surpassing other models such as RoBERTa (0.579) and BERT (0.425), demonstrating its high efficiency in detecting image-based fake news. Moreover, the model also excels in detecting different categories of fake news, particularly in the "Unverifiable Fake News" category, where the F1 score of the proposed model is 0.945, much higher than other models such as FLAVA (0.915) and Reformer (0.925). For the "Non-Fake News" and "Verified Fake News" categories, the F1 scores of the proposed model are 0.885 and 0.854, respectively, also higher than most comparison models, especially in the "Verified Fake News" category, where the performance shows a significant advantage over other models.

Table 3. Experimental results for social media image fake news detection

Table 4. Ablation experiment results for social media image fake news detection model

From the ablation experiment results in Table 4, the performance differences between models with and without the attention mechanism in social media image fake news detection are quite noticeable. In the model without the attention mechanism, the overall accuracy (Acc.) is 0.912, which is quite good, especially in the "Verified True Fake News" category, where the F1 score is 0.923, significantly higher than other categories. However, in the model with the attention mechanism, there is an improvement in several metrics, although the overall accuracy drops to 0.888. The F1 scores for "Non-Fake News" and "Verified Fake News" increase to 0.885 and 0.889, respectively, and the F1 score for "Unverifiable Fake News" also increases to 0.865. In comparison, the addition of the attention mechanism seems to improve the model's ability to identify certain categories, even though the overall accuracy slightly decreases. These results suggest that the inclusion of the attention mechanism contributes to the performance improvement of the proposed social media image fake news detection model. Specifically, the attention mechanism helps to better capture the deep associations between images and text, especially when dealing with "Non-Fake News" and "Verified Fake News," which may be due to the attention mechanism's ability to focus more accurately on the parts critical for determining fake news.

From the experimental data in Figure 6, it can be seen that the number of words in the social media news text has a certain impact on the model's performance. In Dataset 1, the accuracy remains stable between 0.855 and 0.905 as the number of words increases from 10 to 100, showing little fluctuation, indicating that an increase in text length does not significantly improve the model's performance. In contrast, Dataset 2 shows a more noticeable change in accuracy, increasing from 0.755 (10 words) to 0.822 (100 words), presenting a gradual upward trend, though the change is relatively small. In comparison, Dataset 3 demonstrates the most significant improvement, with accuracy increasing from 0.935 to 0.96, showing that as the number of words increases, the model's performance improves significantly, especially when the text length exceeds 70 words, where the accuracy becomes stable, further indicating that longer texts help improve the model’s ability to detect fake news.

2706.png

Figure 6. The impact of text length in social media news on experimental performance

2707.png

Figure 7. The impact of the number of propagation paths in social media news on experimental performance

From the experimental data in Figure 7, it can be seen that the number of propagation paths for social media fake news affects the model's accuracy to varying degrees. In Dataset 1, as the number of propagation paths increases, the accuracy fluctuates somewhat but remains stable between 0.896 and 0.906, showing a relatively steady performance. Even when the number of paths reaches 100, the accuracy only slightly decreases (0.898), indicating that the number of propagation paths has a limited impact on the model’s performance for this dataset. In Dataset 2, the accuracy changes little across different path numbers, ranging from 0.822 to 0.816, indicating that the increase in propagation paths has no significant effect on this dataset. In contrast, Dataset 3 shows a more noticeable improvement, with accuracy increasing from 0.938 to 0.958, particularly when the number of propagation paths exceeds 70, where the accuracy stabilizes. This shows that more propagation paths can help the model better capture the propagation characteristics of fake news, improving its detection ability.

These results indicate that the impact of the number of propagation paths for social media fake news on the model depends on the characteristics of the dataset. In Dataset 3, the increase in propagation paths significantly improved the model's performance, likely because the propagation network in this dataset is more complex, and more paths provide the model with richer semantic information, allowing for more accurate identification of fake news. In contrast, in Datasets 1 and 2, the effect of the number of propagation paths on the model's performance is weak, possibly because these datasets have fewer propagation paths or redundant information, and thus cannot fully utilize the additional path information. Therefore, the number of propagation paths for social media fake news does affect the detection effect, but the magnitude of the effect depends on the characteristics of the dataset and the complexity of the propagation network.