Rumor detection method and system for social network confrontation scene, medium and program product

Abstract

The invention discloses a social network confrontation scene-oriented rumor detection method and system, a medium and a program product, a teacher network is utilized, source tweet text features and propagation structure features are aligned through an attention mechanism to construct effective classification features, a student network is calibrated based on the classification features, and a social network confrontation scene-oriented rumor detection result is obtained. The distance between the extended sample and the original sample is reduced. The teacher network aligns the source tweet features and the propagation structure features through a cross attention mechanism to realize more effective feature fusion, so that rumors and non-rumors are distinguished more accurately, and effective standard aggregation classification features are provided for the student network; the student network continuously corrects the own feature space through the standard aggregation classification features generated by the teacher network to reduce the distance between the extended sample and the original sample, so that the sensitivity of the deep learning model to disturbance is reduced, and the model detection performance is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of rumor detection, and specifically relates to a rumor detection method, system, medium and program product for social network confrontation scenarios. Background Art

[0002] Traditional rumor detection methods mainly include manual methods and machine learning-based methods. Manual methods refer to the determination of whether information is a rumor by relevant experts. With the rapid popularization of social media, the amount of social media data has increased exponentially. Both the detection capability and the detection speed of manual methods cannot meet the current actual needs of fast and accurate detection of social network rumors. Machine learning-based rumor detection methods also have similar problems. Machine learning methods rely on labor-intensive feature engineering, which requires researchers to spend a lot of time screening features suitable for rumor detection tasks. With the increase of data in social networks, the difficulty and complexity of feature engineering also increase. In addition, the effect of machine learning methods depends on the quality of the feature set used, and it is difficult for humans to understand high-dimensional, complex, and abstract features, resulting in the inability to fully extract effective information from the data, which seriously limits the effect of machine learning-based rumor detection methods. Neither manual methods nor machine learning-based methods can meet the needs of quickly and effectively identifying rumors in massive social media data. In this context, researchers began to study rumor detection methods based on deep learning and made significant progress.

[0003] However, deep learning models have serious security issues. Attackers can easily use deep learning attack techniques to attack deep learning models, causing them to output incorrect results. For example, spammers only need to add some tiny adversarial perturbations that are difficult for humans to detect to evade detection by deep learning-based spam detection systems. Rumor detection models based on deep learning also face serious security threats. If rumor makers use adversarial perturbations to modify rumors, existing rumor detection models based on deep learning will not be able to accurately identify these specially processed rumors. Therefore, it is of great significance to design a rumor detection method that can effectively identify adversarial rumors. Summary of the invention

[0004] The purpose of the present invention is to provide a rumor detection method, system, medium and program product for social network confrontation scenarios. The method generates standard aggregate classification features through a teacher network and uses a student network to correct the feature space, thereby improving the robustness of the model in confrontation attack scenarios.

[0005] The purpose of the present invention is achieved through the following technical solutions:

[0006] A rumor detection method for a social network adversarial scenario includes the following steps:

[0007] Step 1: Data acquisition and preprocessing: Obtain data packets through social media platforms, preprocess the text and communication structure, and obtain the information propagation path and network structure characteristics; clean, encode and extract features from original tweets, forwarded comments and user interaction data;

[0008] Step 2: Construction of the teacher network; extracting features through the source tweet text feature branch and the fusion feature branch to generate standard aggregated classification features; using the attention mechanism to align the source tweet text features and the propagation structure features to improve the expressiveness of the classification features;

[0009] Step 3: Construction of the student network; based on the teacher network, an extended sample generation module, a text embedding reconstruction module, and an aggregated classification feature correction module are added to optimize the feature space and improve robustness; the student network reduces the feature offset between the extended samples and the original samples through adversarial learning;

[0010] Step 4: Data input and training; input the standard aggregated classification features generated by the teacher network into the student network. The student network corrects the feature space through adversarial learning to enhance the model's ability to detect adversarial disturbances; and uses multi-task joint training to optimize network performance;

[0011] Step 5: Results and Verification: Verify the detection performance of the model in adversarial attack scenarios using real social media datasets.

[0012] Furthermore, the step 1 is specifically as follows:

[0013] Step 1.1: Data sources;

[0014] Obtain data from social media platforms such as Twitter and Facebook, including tweets, comments, reposts, and timestamp information posted by users; capture the interaction in social networks through these data, and provide effective text content and propagation structure features for rumor detection;

[0015] Step 1.2: Data preprocessing;

[0016] First, the text is cleaned to remove irrelevant URLs, emoticons, and special characters to reduce noise. Then, the text is segmented and word embeddings are generated using the BERT pre-trained model to convert the text into numerical form. Finally, the propagation structure features are constructed, and the forwarding and commenting behaviors in social media are used to build a social network graph to extract the information propagation path and network structure features.

[0017] Furthermore, the step 2 is specifically as follows:

[0018] Step 2.1: Source tweet text feature branch;

[0019] First, use BERT to generate an embedding vector Then, select the first L embedding vectors as the source tweet text embedding vector where d r is the text embedding dimension, and finally a convolutional neural network is used to utilize X r Extract source tweet text features X st ;

[0020] Step 2.2: Fusion feature branches;

[0021] Construct propagation structure features, and capture the network structure information of rumor propagation by analyzing the node relationships, propagation paths, and information flows in social networks;

[0022] Use BERT to model tweet text and generate tweet text features Using node features of the joint graph in Rumor2vec as user features Tweet node feature x j The formula is as follows:

[0023]

[0024] Where: || is the connection operation;

[0025] Next, use the multi-head cross attention mechanism, the formula is as follows:

[0026] The multi-head attention mechanism is as follows:

[0027]

[0028] Where: N h It is the head of many heads of attention;

[0029] The source tweet text features are fused with the propagation structure features. The specific calculation process of the fusion features is as follows:

[0030] MHead=MH(X gi ,X r ,X gi ,N ch )

[0031]

[0032] Where: are trainable model parameters;

[0033] Finally, based on the generated fusion features and the source tweet text features, the standard aggregate classification feature X is generated. scf , the formula is as follows:

[0034] X scf =X ff ‖X st

[0035] Step 2.3: Train the model and optimize;

[0036] Teacher Network Utilization X scf The specific calculation process of the judgment result is as follows:

[0037]

[0038] Where: W tfc are trainable model parameters;

[0039] In the teacher network, the predicted labels are minimized And the cross entropy loss of the true label y is used to train all model parameters;

[0040] By optimizing, we can obtain an optimal standard clustering classification feature X scf , this feature can help the student network correct its feature space; the generation process of standard clustering classification features is as follows:

[0041] X scf =Teacher(G,X N ,X r ,W tn )

[0042] Where: W tn are the weights of the teacher network.

[0043] Furthermore, the step 3 is specifically as follows:

[0044] Step 3.1: Extend the sample generation module EEG;

[0045] The classification keywords are selected through the gradient of the teacher network, and the most influential keywords for the classification task are identified according to the sensitivity of the model to different categories; then, the selected keywords are re-segmented and split into more fine-grained word fragments to better capture the information at the lexical level, and finally the extended sample t' is generated through the BERT model 0 Features of X rexd ;

[0046] Step 3.2: Text embedding reconstruction module TER;

[0047] Use a seq2seq network with an attention mechanism to correct and reconstruct the extended samples through context features;

[0048] In the encoder, h iThe specific calculation process is as follows:

[0049]

[0050] Where: W e are all the trainable parameters involved in GRU, h 0 is a randomly initialized matrix;

[0051] Step 3.3: Aggregate classification feature correction module ACFC;

[0052] Minimize the loss between the standard aggregate classification features and the aggregate classification features, and adjust the parameters of the student network by calculating the difference between the two, so that it is closer to the standard features of the teacher network in the feature space. Then, correct the feature space of the student network and further improve the classification performance by optimizing the weight distribution during the learning process, so that the student network can more accurately distinguish between rumors and non-rumors.

[0053] Furthermore, the step 4 specifically includes:

[0054] The standard aggregate classification features generated by the teacher network are used as input to train the student network through the cross entropy loss function; the student network reduces the impact of adversarial disturbances by optimizing the adversarial learning between the extended sample generation module, the text embedding reconstruction module and the aggregate classification feature correction module, thereby improving the rumor detection performance; the training process combines multiple rounds of refinement optimization mechanism to ensure model convergence by dynamically adjusting the learning rate.

[0055] Furthermore, the training and verification of the model in step 5 are completed through real data sets, as follows:

[0056] Step 5.1: Dataset construction; Based on the public dataset, combine the attack method to generate a dataset containing different proportions of adversarial samples; stratify the generated adversarial samples according to the perturbation intensity to ensure the applicability of the model in different scenarios.

[0057] Step 5.2: Model training and comparison; evaluate the effectiveness of the ITSC method through ten-fold cross validation and compare it with the baseline method to verify the robustness and performance advantages of the model in the anti-rumor scenario; compare the indicators including accuracy, precision, recall and F1 value, and further evaluate the overall performance of the model through macro indicators.

[0058] Step 5.3: Robustness evaluation: Verify the robustness of the model on datasets containing different proportions of adversarial samples, evaluate the performance degradation of the model by gradually increasing the proportion of adversarial samples, and visualize the stability of the model feature space.

[0059] A computer device / equipment / system includes a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of a rumor detection method for a social network confrontation scenario.

[0060] A computer-readable storage medium stores a computer program / instruction thereon, which, when executed by a processor, implements the steps of a rumor detection method for a social network confrontation scenario.

[0061] A computer program product includes a computer program / instruction, which, when executed by a processor, implements the steps of a rumor detection method for a social network confrontation scenario.

[0062] The beneficial effects of the present invention are:

[0063] The present invention mainly includes two parts: a teacher network and a student network. Using the teacher network, the source tweet text features and the propagation structure features are aligned through the attention mechanism to construct effective classification features, and the student network is calibrated based on the classification features to reduce the distance between the extended sample (generated in the same way as the adversarial sample) and the original sample. The advantages of the present invention are: first, the teacher network aligns the source tweet features and the propagation structure features through the cross-attention mechanism to achieve more effective feature fusion, thereby more accurately distinguishing rumors from non-rumors and providing effective standard aggregation classification features for the student network; secondly, the student network continuously corrects its own feature space through the standard aggregation classification features generated by the teacher network to reduce the distance between the extended sample and the original sample, thereby reducing the sensitivity of the deep learning model to disturbances, thereby improving the model detection performance. The present invention trains a robust rumor detection model through a teacher-student network architecture, which can still perform effective rumor detection in adversarial attack scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 Flow chart of the present invention

[0065] Figure 2 A typical application environment for implementing the present invention.

[0066] Figure 3 A structural block diagram of the system described in the present invention.

[0067] Figure 4 The overall architecture of the teacher network in the teacher-student network based rumor detection method.

[0068] Figure 5 Text embedding reconstruction module of the student network in the teacher-student network based rumor detection method. Figure 6Comparison of the recall rates of the models on the BERT-Dataset and Glove-Dataset datasets. DETAILED DESCRIPTION

[0069] The present invention is further described below in conjunction with the accompanying drawings.

[0070] according to Figure 1 The present invention finally verifies the rumor detection model through data acquisition and preprocessing, construction and training of teacher network and student network, and improves its robustness in anti-attack scenarios. The specific steps are as follows:

[0071] Step 1: Data acquisition and preprocessing

[0072] The data sources are mainly social media platforms (such as Twitter, Facebook, etc.), and the content on these platforms can reflect the spread of rumors and their impact on users. In the data preprocessing stage, steps such as cleaning and feature extraction are used to ensure that the model can extract valuable information from it for effective rumor detection.

[0073] Step 2: Construction of teacher network;

[0074] Features are extracted through the source tweet text feature branch and the fusion feature branch to generate standard aggregated classification features; the attention mechanism is used to align the source tweet text features and propagation structure features to improve the expressiveness of classification features.

[0075] Step 3: Construction of student network;

[0076] Based on the teacher network, an extended sample generation module, a text embedding reconstruction module, and an aggregate classification feature correction module are added to optimize the feature space and improve robustness; the student network reduces the feature offset between the extended samples and the original samples through adversarial learning;

[0077] Step 4: Data input and training;

[0078] The standard aggregated classification features generated by the teacher network are input into the student network. The student network corrects the feature space through adversarial learning to enhance the model's ability to detect adversarial disturbances. A multi-task joint training method is used to optimize network performance.

[0079] Step 5: Results and verification;

[0080] Verify the detection performance of the model, especially in adversarial attack scenarios.

[0081] according to Figure 2 ,The rumor maker of the present invention modifies the source tweet through adversarial perturbations, so that the rumor successfully escapes the detection of the rumor detection model while keeping the semantics unchanged.

[0082] according to Figure 3 , the structural block diagram of the system described in the present invention includes:

[0083] Teacher Network: Aligns source tweet features and propagation structure features through a cross-attention mechanism to achieve more effective feature fusion, thereby more accurately distinguishing rumors from non-rumors and providing effective standard aggregate classification features for the student network.

[0084] Student network: It continuously corrects its own feature space through the standard aggregated classification features generated by the teacher network to narrow the distance between the extended samples and the original samples, thereby reducing the sensitivity of the deep learning model to perturbations and improving the detection performance of the model in adversarial scenarios.

[0085] according to Figure 4 , is the overall architecture of the teacher network in the rumor detection method based on the teacher-student network, as follows:

[0086] The present invention uses BERT and convolutional neural network to construct source tweet text features. First, BERT is used to generate an embedding vector Then, select the first L embedding vectors as the source tweet text embedding vector where d r is the text embedding dimension, and finally a convolutional neural network is used to utilize X r Extract source tweet text features X st .

[0087] X st =CNN(X r )

[0088] In the teacher network, for each tweet node v in the propagation structure graph j =(u j ,t j ), the present invention uses the joint representation of user features and tweet text features as its node features. Specifically, for tweet text, the present invention uses BERT to model the tweet text and generate tweet text features For user features, the present invention does not use unreliable attributes that ordinary users can modify at will, such as user age, gender, region, etc., but uses the node features of the joint graph in Rumor2vec as user features. Compared with the attribute information that users can modify freely, the user features generated based on a large amount of propagation structure data are more reliable. It should be noted that the edge weights in the joint graph constructed by the present invention are different from the edge weights in the joint graph in Rumor2vec. The edge weight of the joint graph in Rumor2vec is the number of propagation graphs containing the edge, while the edge weight of the joint graph used in the present invention is the sum of the edge weights in the propagation graph containing the edge. The final tweet node feature x j As shown below:

[0089]

[0090] Where: || is the connection operation

[0091] This section uses X N ={x 0 ,x 1 ,…,x N-1} represents the matrix of node feature vectors in the propagation graph. Unlike graph convolutional neural networks that treat each neighbor node equally in the process of aggregating neighbor node features, graph attention networks can assign different weights to different neighbor nodes. In the field of rumor detection, some tweets contain important information, such as tweets posted by professionals, while some tweets themselves are not very meaningful for rumor detection, such as tweets posted by someone who is not familiar with the content of the current topic. Therefore, the present invention uses a graph attention network to extract propagation structure features. The graph attention network consists of several graph attention layers. The specific calculation process of the lth graph attention layer is as follows:

[0092]

[0093] e′ i,j =LeakyReLU(e i,j )

[0094]

[0095] Where: and is a trainable model parameter, || is a connection operation, σ is an activation function, is node v i The neighbor node set of . For the convenience of the following text, the present invention uses Represents the above formula.

[0096] According to the single-head attention layer in the above formula, the multi-head attention layer is formally expressed as follows:

[0097]

[0098] In the teacher network, the present invention uses a graph attention network with two attention layers, the first layer is a multi-head attention layer, and the second layer is a single-head attention layer. The specific calculation process of the graph attention network used in the present invention is as follows:

[0099] X'=SMH(X N ,N gh )

[0100] Xg=AttLayer(X',W gin ,W ga )

[0101] Where: N gh is the number of attention heads used by the multi-head attention layer.

[0102] In the teacher network, the σ activation function used by the graph attention network is the elu

[40] activation function, the negative input slope of LeakyReLU() is set to 0.2, and N gh Set to 8.

[0103] At present, the present invention has generated the initial propagation structure characteristics It should be noted that the number of tweets N for different topics is different. Since a fixed-dimensional propagation structure feature matrix is ​​required in the subsequent calculation, the present invention calculates the initial propagation structure feature X g Pad or truncate. Previous rumor detection methods usually directly pad or truncate the tweet sequence according to the release time of the tweet. However, important tweets do not necessarily appear in the first part of the tweet sequence. If the tweet sequence is cut directly according to the release time, those tweets that are released later and contain important information may be lost. Therefore, the present invention cuts the tweet sequence according to the importance of the tweet. j The importance calculation formula is as follows:

[0104]

[0105] Where: α i,j It's a tweet v j With tweets v i Attention scores in the second attention layer of the graph attention network.

[0106] For each topic, the present invention retains the most important N t Specifically, for a tweet sequence longer than N t The present invention sorts the tweets according to their importance and retains the top N t tweets. For a tweet sequence shorter than N t The present invention uses zero vectors to supplement, and the final characteristic matrix Among them, dgi is the dimension of the node embedding vector in the propagation graph. gi It is called the propagation structure characteristic matrix.

[0107] The present invention utilizes the source tweet text embedding feature X based on the multi-head cross attention mechanism r and propagation structure characteristics X gi Generate fusion feature X ff The multi-head attention mechanism is composed of multiple single-head attention mechanisms. The specific calculation process of the single-head attention mechanism is as follows:

[0108]

[0109] The single-head attention mechanism in the above formula is represented as Attn(X k W k ,X q W q ,X v W v ). On this basis, the multi-head attention mechanism is as follows:

[0110]

[0111] Where: N h is the number of heads of multi-head attention.

[0112] The present invention uses the multi-head attention mechanism shown in the above formula to generate fusion features. The specific calculation process of the fusion features is as follows:

[0113] MHead=MH(X gi ,X r ,X gi ,N ch )

[0114]

[0115] Where: are trainable model parameters.

[0116] The present invention utilizes the fusion feature X ff and text feature X st Generate standard aggregated categorical features X scf , standard aggregate classification feature X scf The specific calculation process is as follows:

[0117] X scf =X ff ‖X st

[0118] Teacher Network Utilization X scf The specific calculation process of the judgment result is as follows:

[0119]

[0120] Where: W tfc are trainable model parameters.

[0121] In the teacher network, the present invention minimizes the predicted label All model parameters are trained using the cross entropy loss between y and the true label y.

[0122] Through optimization, the present invention obtains an optimal standard clustering classification feature X scf , which can help the student network correct its feature space. For ease of use, the present invention describes the generation process of the standard clustering classification feature through the following equation:

[0123] X scf =Teacher(G,X N ,X r ,W tn )

[0124] Where: W tn are the weights of the teacher network.

[0125] according to Figure 5 , a schematic diagram of the text embedding reconstruction module of the student network in the rumor detection method based on the teacher-student network is given, as follows:

[0126] In the encoder, h i The specific calculation process is as follows:

[0127]

[0128] Where: W e are all the trainable parameters involved in GRU, h 0 is a randomly initialized matrix, X rexd To generate extended samples t' through BERT 0 characteristics.

[0129] In the decoder, h' i The specific calculation process is as follows:

[0130] h' i =GRU(s i-1 ,h' i-1 ,W d )

[0131] Where: h' 0 =h L ,s i It is the aggregated feature generated by applying the multi-head attention mechanism.

[0132] TER uses a multi-head attention mechanism to generate aggregate features s i .s i The specific generation process is as follows:

[0133] s i =MH(H,h' i ,H,N dh )

[0134] Where: H = {h 1 ,h 2 ,…,h L}.

[0135] Get the reconstructed source tweet text embedding features (reconstructed features) X rec ={h' 1 ,h' 2 ,…,h' L}. Embed X based on the source tweet text r and reconstruct feature X rec , the present invention calculates the reconstruction loss L of the student network rec The specific calculation process of reconstruction loss is as follows:

[0136]

[0137] Where: For smoothing L 1 Loss function.

[0138] The following describes in detail the experimental example scenarios of the present invention, and analyzes the implementation results in combination with the advantages of the present invention.

[0139] To verify the effectiveness of the present invention, we conducted experiments on the public dataset Pheme. The present invention first compared the detection performance of other baseline methods of the present invention on a dataset containing adversarial rumors. Then, the robustness of the model was evaluated on datasets containing different proportions of adversarial samples. Finally, the present invention analyzed the effective mechanism of the present invention's method from the perspective of feature shift.

[0140] The present invention uses the public dataset pheme to evaluate the present invention method. In order to effectively evaluate the performance of the present invention method in the adversarial attack scenario, the present invention processes the dataset to ensure that it contains enough adversarial rumors.

[0141] The present invention first constructs two attacked models (proxy models) BERTF and GloveRNNF.

[0142] (1) BERTF: This model mainly consists of two parts, BERT and a fully connected layer. The model directly inputs the sentence vector generated by BERT into the fully connected layer.

[0143] (2) GloveRNNF: The model mainly consists of three parts, Glove, RNN and fully connected layer. Glove is an open source word vector disclosed by Pennington et al. Since Glove cannot directly generate sentence vectors like BERT, the present invention first inputs the Glove word vector into the RNN, and then uses the hidden layer of the last time step of the RNN as the sentence vector input into the fully connected layer.

[0144] The present invention trains GloveRNNF and BERTF based on pheme, and then uses PWWS, PSO, TextBugger and TextFooler to attack GloveRNNF and BERTF to generate eight attack data sets. Taking PWWS attacking BERTF as an example, for the data in pheme, if the data is a rumor and BERTF also determines that it is a rumor, PWWS is used to attack the data. If the attack is successful, the data in pheme is replaced with the attacked data. For convenience, the present invention refers to the data set generated by the GloveRNNF attack as Glove-Dataset, and the data set generated by the BERTF attack as BERT-Dataset. The ratio of adversarial rumors contained in the eight attacked data sets is shown in Tables 1 and 2.

[0145] Table 1 Percentage of adversarial rumor samples in four datasets in Glove-Dataset

[0146]

[0147] Table 2 Percentage of adversarial rumor samples in four datasets in BERT-Dataset

[0148]

[0149] The present invention uses OpenAttack to implement PWWS, PSO, TextBugger and TextFooler. For fairness, the present invention uses the default parameters in OpenAttack for the four attack methods.

[0150] Comparison is made with some baseline methods, which are shown below:

[0151] (1) GRU-2: A two-layer GRU is used to learn the temporal changes of contextual information in tweet sequences.

[0152] (2) GAN-GRU: The generator is used to generate uncertain or conflicting voices to make the tweet sequence more complex, forcing the discriminator to learn more robust rumor features.

[0153] (3) CAMI: Use CNN to learn high-level interactions between tweets.

[0154] (4) Rumor2vec: It uses the node embedding of the joint graph and the text features of the source tweets to jointly detect rumors.

[0155] In the experiment, the present invention adopts ten-fold cross validation. In addition, for fairness, for all models including the baseline model and the rumor detection model proposed by the present invention, the present invention retains the model with the highest accuracy of the validation set as the final model during the training process. The present invention first trains the model with pheme, and then evaluates the effectiveness of the model on eight data sets containing a large number of adversarial rumors.

[0156] The specific experimental environment of the present invention is shown in Table 3.

[0157] Table 3 Operating environment

[0158]

[0159] In order to evaluate the effectiveness of the proposed method, the present invention is compared with the baseline method on eight datasets containing a large number of anti-rumor data. This section evaluates the model by accuracy (Acc), precision (Pre), recall (Rec) and F1. Table 4 and Table 5 show the performance of the proposed method and the baseline method on Glove-Dataset and BERT-Dataset, respectively. Figure 5 This result is more intuitively presented. Next, the present invention will analyze in detail the performance of the present invention method and the baseline method on different data sets.

[0160] Table 4 Performance of the model in the Glove-Dataset dataset

[0161]

[0162] Table 5 Performance of the model in the Bert-Dataset dataset

[0163]

[0164] First, as shown in Table 4, Table 5, Figure 5As shown, the method of the present invention is significantly better than all baseline methods. On the four data sets included in Glove-Dataset, the method of the present invention improves by an average of 4.58%, 6.38%, 10.48% and 8.85% in accuracy, precision, recall and F1 compared with the Rumor2vec method with the best overall performance among the baseline methods. On the four data sets included in Bert-Dataset, the accuracy, precision, recall and F1 of the method of the present invention are improved by an average of 2.47%, 4.46%, 4.45% and 4.39% compared with Rumor2vec. The experimental results show that the method of the present invention has better performance than the baseline method on the data set containing adversarial rumors. This is because the method of the present invention uses the attention mechanism to align the source tweet text features with the tweet propagation structure features to achieve effective feature fusion, and the method of the present invention makes the rumor detection model insensitive to disturbances through adversarial learning between EEG, TER and ACFC. Compared with other baseline methods, the present invention can effectively detect adversarial rumors in the data set.

[0165] Secondly, as shown in Table 4, Table 5, Figure 6 As shown in the figure, the performance of the GRU-2 method, the GAN-GRU method and the CAMI method varies little in different datasets, and the maximum and minimum differences in the recall rates of these methods are 4.62%, 4.93% and 1.39% respectively. The performance of the Rumor2vec method in different datasets shows obvious fluctuations, with the maximum and minimum differences being 17.37%. The reason for this phenomenon is that although the GRU-2 method, the GAN-GRU method and the CAMI method use source tweets, they treat the source tweets equally with other tweets and do not pay enough attention to the source tweets. Compared with these methods, the Rumor2vec method and the method of the present invention pay enough attention to the source tweets, so their overall accuracy is still higher than other baseline methods even in the dataset containing the attacked data. However, this emphasis on source tweets makes the model more vulnerable to the adverse effects of adversarial source tweets, so the recall rate of the Rumor2vec method changes dramatically in different attack datasets. To meet this challenge, the method of the present invention reduces the distance between the extended samples and the original samples, making the rumor detection model insensitive to adversarial disturbances, thereby improving the model's detection accuracy for adversarial rumors.

[0166] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A rumor detection method for a social network confrontation scenario, characterized by: The following steps are involved: Step 1: Data acquisition and preprocessing; Obtain data packets through social media platforms, pre-process their text and communication structures, and obtain information propagation paths and network structure characteristics; Clean, encode and extract features from original tweets, forwarded comments and user interaction data; Step 2: Construction of the teacher network; extracting features through the source tweet text feature branch and the fusion feature branch to generate standard aggregate classification features; The attention mechanism is used to align the text features of the source tweets and the propagation structure features to improve the expressiveness of the classification features; Step 3: Construction of the student network; based on the teacher network, an extended sample generation module, a text embedding reconstruction module, and an aggregated classification feature correction module are added to optimize the feature space and improve robustness; the student network reduces the feature offset between the extended samples and the original samples through adversarial learning; Step 4: Data input and training; input the standard aggregated classification features generated by the teacher network into the student network. The student network corrects the feature space through adversarial learning to enhance the model's ability to detect adversarial disturbances; and uses multi-task joint training to optimize network performance; Step 5: Results and Verification: Verify the detection performance of the model in adversarial attack scenarios using real social media datasets.

2. According to claim 1, a rumor detection method for a social network confrontation scenario is characterized by: The step 1 is specifically as follows: Step 1.1: Data sources; Obtain data from social media platforms such as Twitter and Facebook, including tweets, comments, reposts, and timestamp information posted by users; capture the interaction in social networks through these data, and provide effective text content and propagation structure features for rumor detection; Step 1.2: Data preprocessing; First, the text is cleaned to remove irrelevant URLs, emoticons, and special characters to reduce noise. Then, the text is segmented and word embeddings are generated using the BERT pre-trained model to convert the text into numerical form. Finally, the propagation structure features are constructed, and the forwarding and commenting behaviors in social media are used to build a social network graph to extract the information propagation path and network structure features.

3. The rumor detection method for a social network confrontation scenario according to claim 1 is characterized in that: The step 2 is specifically as follows: Step 2.1: Source tweet text feature branch; First, use BERT to generate an embedding vector Then, select the first L embedding vectors as the source tweet text embedding vector where d r is the text embedding dimension, and finally a convolutional neural network is used to utilize X r Extract source tweet text features X st ; Step 2.2: Fusion feature branches; Construct propagation structure features, and capture the network structure information of rumor propagation by analyzing the node relationships, propagation paths, and information flows in social networks; Use BERT to model tweet text and generate tweet text features Using node features of the joint graph in Rumor2vec as user features Tweet node feature x j The formula is as follows: Where: || is the connection operation; Next, use the multi-head cross attention mechanism, the formula is as follows: The multi-head attention mechanism is as follows: Where: N h It is the head of many heads of attention; The source tweet text features are fused with the propagation structure features. The specific calculation process of the fusion features is as follows: MHead=MH(X gi ,X r ,X gi ,N ch ) Where: are trainable model parameters; Finally, based on the generated fusion features and the source tweet text features, the standard aggregate classification feature X is generated. scf , the formula is as follows: X scf =X ff ‖X st Step 2.3: Train the model and optimize; Teacher Network Utilization X scf The specific calculation process of the judgment result is as follows: Where: W tfc are trainable model parameters; In the teacher network, the predicted labels are minimized And the cross entropy loss of the true label y is used to train all model parameters; By optimizing, we can obtain an optimal standard clustering classification feature X scf ,This feature can help the student network correct its feature space; The generation process of standard cluster classification features is as follows: X scf =Teacher(G,X N ,X r ,W tn ) Where: W tn are the weights of the teacher network.

4. The rumor detection method for a social network confrontation scenario according to claim 1 is characterized by: The step 3 is specifically as follows: Step 3.1: Extend the sample generation module EEG; The classification keywords are selected through the gradient of the teacher network, and the most influential keywords for the classification task are identified according to the sensitivity of the model to different categories; then, the selected keywords are re-segmented and split into more fine-grained word fragments to better capture the information at the lexical level, and finally the feature X of the extended sample t'0 is generated through the BERT model rexd ; Step 3.2: Text embedding reconstruction module TER; Use a seq2seq network with an attention mechanism to correct and reconstruct the extended samples through context features; In the encoder, h i The specific calculation process is as follows: Where: W e are all the trainable parameters involved in GRU, and h0 is a randomly initialized matrix; Step 3.3: Aggregate classification feature correction module ACFC; Minimize the loss between the standard aggregate classification features and the aggregate classification features, and adjust the parameters of the student network by calculating the difference between the two, so that it is closer to the standard features of the teacher network in the feature space. Then, correct the feature space of the student network and further improve the classification performance by optimizing the weight distribution during the learning process, so that the student network can more accurately distinguish between rumors and non-rumors.

5. The rumor detection method for social network confrontation scenarios according to claim 1 is characterized by: The step 4 specifically includes: The standard aggregate classification features generated by the teacher network are used as input to train the student network through the cross entropy loss function; the student network reduces the impact of adversarial disturbances by optimizing the adversarial learning between the extended sample generation module, the text embedding reconstruction module and the aggregate classification feature correction module, thereby improving the rumor detection performance; the training process combines multiple rounds of refinement optimization mechanism to ensure model convergence by dynamically adjusting the learning rate.

6. The rumor detection method for social network confrontation scenario according to claim 1 is characterized by: The training and verification of the model in step 5 are completed using a real data set, as follows: Step 5.1: Dataset construction; Based on the public dataset, combine the attack method to generate a dataset containing different proportions of adversarial samples; stratify the generated adversarial samples according to the perturbation intensity to ensure the applicability of the model in different scenarios. Step 5.2: Model training and comparison; evaluate the effectiveness of the ITSC method through ten-fold cross validation and compare it with the baseline method to verify the robustness and performance advantages of the model in the anti-rumor scenario; compare the indicators including accuracy, precision, recall and F1 value, and further evaluate the overall performance of the model through macro indicators. Step 5.3: Robustness evaluation: Verify the robustness of the model on datasets containing different proportions of adversarial samples, evaluate the performance degradation of the model by gradually increasing the proportion of adversarial samples, and visualize the stability of the model feature space.

7. A computer device / equipment / system comprising a memory, a processor and a computer program stored in the memory, characterized in that: The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 6.

8. A computer-readable storage medium having a computer program / instruction stored thereon, characterized in that: When the computer program / instructions are executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

9. A computer program product comprising a computer program / instructions, characterized in that: When the computer program / instructions are executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images