Social media fraudulent content identification method and system based on deep learning

Abstract

The invention discloses a social media fraudulent content identification method and system based on deep learning. The method comprises the following steps: step 1, acquiring multi-modal data; step 2, carrying out multi-modal feature fusion; step 3, hierarchical feature extraction; step 4, distributing weights through an attention mechanism; step 5, carrying out fraudulent classification identification and judgment through an adversarial learning model; the recognition system comprises a data acquisition module, a multi-modal feature fusion module, a hierarchical feature extraction module, an attention mechanism module and an adversarial learning classification module. The method has the advantages that the model can be automatically focused on key fraudulent areas or vocabularies in the content, the attention degree on important features is improved, meanwhile, the generalization ability of the model is enhanced through the adversarial learning technology, the model can better cope with fraudulent content of different types and forms, and the method has good application prospects. The problem that a traditional model is insufficient in capturing complex semantics and long-distance dependence is effectively solved.

Description

Technical Field

[0001] This invention relates to the field of social offensive content recognition, specifically to a method and system for social media offensive content recognition based on deep learning. Background Technology

[0002] With the rapid development of social media, user-generated content has exploded, and offensive content within it has severely impacted the online environment and user experience. Existing methods for identifying offensive content are mainly based on traditional machine learning, such as support vector machines and Naive Bayes. These methods rely on manually designed features, and their effectiveness is limited for complex social media content, especially content containing multimodal information such as text, images, and emojis. Furthermore, traditional methods struggle to capture long-distance semantic dependencies and contextual information, resulting in insufficient accuracy and generalization ability. Therefore, there is an urgent need for a deep learning-based method that can automatically learn multimodal features and effectively identify offensive content in social media. Summary of the Invention

[0003] The purpose of this invention is to solve the problems mentioned above, and therefore proposes a method and system for identifying offensive content on social media based on deep learning.

[0004] The objective of this invention can be achieved through the following technical solution: a deep learning-based method for identifying offensive content on social media, comprising: Step 1: Multimodal data acquisition. Collect content data from social media platforms, including text, images, and emojis. Perform word segmentation, stop word removal, and part-of-speech tagging on the text data to convert it into a word vector sequence. Perform noise reduction and normalization on the image data and extract image features through a convolutional neural network. Encode the emojis and convert them into corresponding vector representations. Remove noisy data through data cleaning. Step 2: Multimodal feature fusion. Input the preprocessed text word vectors, image features, and emoji encoding vectors, and fuse the features of different modalities to generate preliminary multimodal fusion features; Step 3: Hierarchical feature extraction. The preliminary multimodal fusion features are input into a hierarchical deep neural network. Semantic features are extracted through hierarchical feature extraction. Step 4: Assign different weights to different features using an attention mechanism; Step 5: Input the features processed by the attention mechanism into the adversarial learning model, classify and identify the features, and judge the offensiveness of the identified content.

[0005] This invention also provides a deep learning-based social media offensive content recognition system, applied to the above-mentioned recognition method. The system includes a data acquisition module, a multimodal feature fusion module, a hierarchical feature extraction module, an attention mechanism module, and an adversarial learning classification module.

[0006] Furthermore, the data acquisition module is used to collect content data and perform classification and preprocessing on the content data.

[0007] Furthermore, the multimodal feature fusion module is used to receive multimodal data preprocessed by the data acquisition module, and to fuse the features of different modalities through a fusion algorithm to generate multimodal fused features.

[0008] Furthermore, the hierarchical feature extraction module is used to perform hierarchical feature extraction on the multimodal fusion features to extract semantic features.

[0009] Furthermore, the attention mechanism module assigns weights to different features based on the hierarchical features extracted by the hierarchical feature extraction module.

[0010] Furthermore, the adversarial learning classification module includes a generator, a discriminator, and a classifier. The generator acquires feature information output by the attention mechanism module, obtains the distribution pattern of real offensive content based on the input feature information, and generates features similar to real offensive content. The discriminator receives features similar to real offensive content and features of real offensive content, and trains the adversarial model using these features to optimize its discrimination ability. The classifier, based on the optimized adversarial model training, performs the final classification and recognition of the input features.

[0011] Compared with existing technologies, the beneficial effects of this invention are: it can effectively integrate multimodal information such as text, images, and emojis, and gradually extract features through a hierarchical network structure, thereby capturing offensive features in the content more comprehensively. This invention considers the complementarity of multimodal information, avoids the limitations of single-modal information, and improves the accuracy and comprehensiveness of feature extraction. The model can automatically focus on key offensive regions or words in the content, increasing attention to important features. At the same time, adversarial learning technology is used to enhance the model's generalization ability, enabling it to better cope with different types and forms of offensive content, effectively solving the problem of traditional models' insufficient capture of complex semantics and long-distance dependencies. Attached Figure Description

[0012] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings.

[0013] Figure 1This is a flowchart of the deep learning-based social media offensive content identification method of the present invention.

[0014] Figure 2 This is a diagram illustrating the components of the deep learning-based social media offensive content recognition system of this invention. Detailed Implementation

[0015] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0016] Please see Figure 1 As shown, deep learning-based methods for identifying offensive content on social media include: Step 1: Multimodal data acquisition. Collect content data from social media platforms, including text, images, and emojis. Perform word segmentation, stop word removal, and part-of-speech tagging on the text data to convert it into a word vector sequence. Perform noise reduction and normalization on the image data and extract image features through a convolutional neural network. Encode the emojis and convert them into corresponding vector representations. Remove noisy data through data cleaning. Step 2: Multimodal feature fusion. Input the preprocessed text word vectors, image features, and emoji encoding vectors. Fuse the features from different modalities to generate preliminary multimodal fusion features, as detailed below: S1: Multilayer perceptron (MLP) is used to perform nonlinear transformation on the features of each modality. For text word vectors, after MLP processing, the semantic relationship between words can be captured. For images, the semantic features of the image are extracted. S2: The features of each modality after MLP processing are concatenated to form a long vector. This method directly integrates the feature information of different modalities to achieve preliminary multimodal feature fusion and generate preliminary multimodal fusion features. By fusing multimodal information, the complementarity between each modality is fully utilized, enabling the model to understand the content from multiple perspectives and improve the ability to identify offensive content.

[0017] Step 3: Hierarchical Feature Extraction. The initial multimodal fusion features are input into a hierarchical deep neural network. Through hierarchical feature extraction, semantic features are extracted. The deep neural network includes a low-level convolutional neural network layer and a high-level recurrent neural network layer. The low-level convolutional neural network layer is used to extract local features, such as semantic features of words and local texture features of images. The high-level recurrent neural network layer is used to capture long-distance semantic dependencies and contextual information, generating hierarchical features. Through hierarchical feature extraction, the overall meaning of the content can be better represented.

[0018] Step 4: Assign different weights to different features using an attention mechanism. Specifically, the feature vectors output by the hierarchical deep neural network are linearly transformed, mapping them to query vectors, key vectors, and sum-value vectors, respectively. The original attention score is obtained by calculating the dot product or cosine similarity between the query vector and all key vectors, quantifying the degree of association between features. The original attention score is then scaled, for example, by dividing by... , The key vector dimension is used in the attention mechanism to calculate the similarity between the query vector and the key vector, thus obtaining the attention score, which is obtained by dividing by... Scaling makes the gradient more stable during backpropagation, avoiding gradient vanishing or exploding problems. Then, the gradient is normalized using the Softmax function to generate attention weights ranging from 0 to 1. These weights reflect the relative importance of each feature in identifying offensive content. The normalized attention weights are then weighted and summed with their corresponding value vectors. In this way, the model can dynamically focus on key offensive regions, suppress irrelevant information interference, and significantly improve recognition accuracy.

[0019] Step 5: Input the attention-processed features into the adversarial learning model for feature classification and offensive content judgment. Specifically: Input the attention-processed features into the generator. Through multi-layer neural network mapping, learn the features of real offensive content, and then generate simulated offensive content features. Simultaneously input the real and simulated offensive content features into the discriminator, calculate the similarity between the input features and the real offensive content features, and output a discrimination probability value to determine whether the input features come from real data or the generator. Based on the discriminator's output, construct an adversarial loss function. During adversarial training, the generator adjusts its parameters through backpropagation to reduce the probability that the discriminator correctly identifies its generated content. The goal is to enable the generator to generate realistic and diverse offensive content. The discriminator updates its parameters using a gradient ascent algorithm to improve its probability of correctly identifying genuine offensive content and content generated by the generator, aiming to enhance its ability to distinguish between different content features. By training the generator and discriminator alternately, the generator can capture more diverse offensive content feature patterns, while the discriminator improves its feature discrimination ability, ultimately enhancing the model's generalization performance in complex scenarios. Through the above methods, multimodal data acquisition and preprocessing ensure data quality, multimodal feature fusion fully utilizes modal complementarity, hierarchical feature extraction achieves a deep understanding of the content, attention mechanisms focus on key features, and adversarial learning and classification recognition enhance the model's generalization ability. The synergistic effect of each step significantly improves the accuracy and robustness of social media offensive content recognition, resulting in a substantial improvement in recognition performance when faced with complex multimodal social media content.

[0020] Reference Figure 2This invention also provides a deep learning-based system for identifying offensive content on social media. This system includes a data acquisition module, a multimodal feature fusion module, a hierarchical feature extraction module, an attention mechanism module, and an adversarial learning classification module. The data acquisition module collects content data and performs classification preprocessing on the content data. The multimodal feature fusion module receives the preprocessed multimodal data from the data acquisition module and fuses features from different modalities using a fusion algorithm to generate multimodal fused features. The hierarchical feature extraction module performs hierarchical feature extraction on the multimodal fused features to extract semantic features. The attention mechanism module assigns weights to different features based on the hierarchical features extracted by the hierarchical feature extraction module. The adversarial learning classification module includes a generator, a discriminator, and a classifier. The generator acquires feature information output by the attention mechanism module, and based on the input feature information, obtains the distribution pattern of real offensive content, generating features similar to real offensive content. The discriminator receives features similar to real offensive content and features of real offensive content, and uses these features to train an adversarial model to optimize its discrimination ability. The classifier, based on the optimized adversarial model, performs the final classification and recognition of the input features. The various modules of the system have clear division of labor and work closely together. The data acquisition module provides basic data, the multimodal feature fusion module enriches feature input, the hierarchical feature extraction module deeply understands the content, the attention mechanism module captures key information, and the adversarial learning classification module achieves accurate classification and recognition, providing strong support for content security management on social media platforms.

[0021] Example 1: Taking a social media platform as an example, user-posted posts, comments, and images are collected. For text data, word segmentation and part-of-speech tagging are performed. For image data, noise reduction is performed, and the image size is normalized to a fixed size. Image features are extracted using a pre-trained convolutional neural network. For emoticons, an emoticon dictionary is established, mapping each emoticon to a unique vector representation. The text word vectors, image features, and emoticon vectors are input into a multimodal feature fusion module. Assuming the dimension of the text word vectors is 100, the dimension of the image features is 512, and the dimension of the emoticon vectors is 50, then... The fused feature dimension is 100 + 512 + 50 = 662. A multilayer perceptron is used to perform a nonlinear transformation on the fused features to obtain preliminary multimodal fusion features. The bottom layer uses a convolutional neural network (CNN) to extract local features from the multimodal fusion features, with a kernel size of 3x3 and a number of 64. The top layer uses a Long Short-Term Memory (LSTM) network with a hidden layer dimension of 128. The feature sequence output from the CNN is input into the LSTM, and the LSTM's gating mechanism captures long-range semantic dependencies, generating hierarchical features. Attention weights are calculated at each time step. The attention weights are calculated using either dot-product attention or additive attention. Taking dot-product attention as an example, the hidden state output by the LSTM is multiplied by the query vector to obtain an attention score. This score is then normalized using a softmax function to obtain the attention weight at each time step. The hidden state is multiplied by the attention weight and summed to obtain the context vector after focusing on key features. The generator uses a multilayer perceptron, with random noise vectors as input and generated feature vectors as output, with the same dimension as the real features. The discriminator also uses a multilayer perceptron, with either real or generated features as input and a probability value representing the probability that the input feature is a real feature. Through adversarial training, the generator generates vectors that are as close to real features as possible, while the discriminator distinguishes between real and generated features as accurately as possible. Once the adversarial training reaches equilibrium, a classifier is used to classify and identify real features. The output layer of the classifier has two neurons, corresponding to offensive and non-offensive content respectively, and the classification probability is output using a softmax function. Through these methods, offensive content in social media can be effectively identified, improving recognition accuracy and generalization ability, and providing strong support for content security management on social media platforms.

[0022] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A deep learning-based method for identifying offensive content on social media, characterized in that, The method includes: Step 1: Multimodal data acquisition. Collect content data from social media platforms, including text, images, and emojis. Perform word segmentation, stop word removal, and part-of-speech tagging on the text data to convert it into a word vector sequence. Perform noise reduction and normalization on the image data and extract image features through a convolutional neural network. Encode the emojis and convert them into corresponding vector representations. Remove noisy data through data cleaning. Step 2: Multimodal feature fusion. Input the preprocessed text word vectors, image features, and emoji encoding vectors, and fuse the features of different modalities to generate preliminary multimodal fusion features; Step 3: Hierarchical feature extraction. The preliminary multimodal fusion features are input into a hierarchical deep neural network. Semantic features are extracted through hierarchical feature extraction. Step 4: Assign different weights to different features using an attention mechanism; Step 5: Input the features processed by the attention mechanism into the adversarial learning model, classify and identify the features, and judge the offensiveness of the identified content.

2. A deep learning-based social media offensive content identification system, applied to the method described in claim 1 above, characterized in that, The system includes a data acquisition module, a multimodal feature fusion module, a hierarchical feature extraction module, an attention mechanism module, and an adversarial learning classification module.

3. The deep learning-based social media offensive content recognition system according to claim 2, characterized in that, The data acquisition module is used to collect content data and perform classification and preprocessing on the content data.

4. The deep learning-based social media offensive content recognition system according to claim 3, characterized in that, The multimodal feature fusion module is used to receive multimodal data preprocessed by the data acquisition module, and to fuse the features of different modalities through a fusion algorithm to generate multimodal fused features.

5. The deep learning-based social media offensive content recognition system according to claim 4, characterized in that, The hierarchical feature extraction module is used to perform hierarchical feature extraction on multimodal fusion features to extract semantic features.

6. The deep learning-based social media offensive content recognition system according to claim 5, characterized in that, The attention mechanism module assigns weights to different features based on the hierarchical features extracted by the hierarchical feature extraction module.

7. The deep learning-based social media offensive content recognition system according to claim 6, characterized in that, The adversarial learning classification module includes a generator, a discriminator, and a classifier. The generator acquires feature information output by the attention mechanism module, obtains the distribution pattern of real offensive content based on the input feature information, and generates features similar to real offensive content. The discriminator receives features similar to real offensive content and features of real offensive content, and trains an adversarial model using these features to optimize its discrimination ability. The classifier, based on the optimized adversarial model training, performs the final classification and recognition of the input features.

Images