Sensitive information identification method and device, server and computer readable storage medium

Abstract

The embodiment of the invention provides a sensitive information identification method and device, a server and a computer readable storage medium, and relates to the technical field of natural language processing. Collecting network parameters and behavior parameters released by each user in various modal forms from at least one network platform; carrying out attribute feature extraction fusion on the network parameters corresponding to the users to obtain fusion feature vectors corresponding to the users, and carrying out intention feature extraction on the network parameters and behavior parameters corresponding to the users to obtain intention features corresponding to the users; and determining whether sensitive information exists in the network parameters corresponding to the users or not according to the fusion feature vectors and the intention features corresponding to the users. Therefore, potential risks can be recognized in time, and potential safety hazards are reduced.

Description

Technical Field

[0001] This application relates to the technical field of natural language processing. Specifically, it relates to a method, device, server, and computer-readable storage medium for sensitive information recognition. Background Art

[0002] In the digital age, multi-modal information interaction has become the main form of Internet information dissemination. Multi-modal data such as text and images are intertwined, greatly enriching information expression, but also bringing huge challenges to sensitive information management.

[0003] Existing technologies often can only perform simple sensitive information recognition on single-modal data, but this will cause a large number of potential risks to not be recognized in time, thus causing security hazards. Summary of the Invention

[0004] In view of this, the purpose of this application is to provide a method, device, server, and computer-readable storage medium for sensitive information recognition to timely identify potential risks and reduce security hazards.

[0005] To achieve the above purpose, the technical solutions adopted in the embodiments of this application are as follows:

[0006] In the first aspect, this application provides a method for sensitive information recognition, and the method includes:

[0007] Collect network parameters and behavior parameters in various modal forms published by each user from at least one network platform;

[0008] Extract and fuse attribute features of the network parameters corresponding to each user to obtain a fused feature vector corresponding to each user, and extract intention features of the network parameters and behavior parameters corresponding to each user to obtain an intention feature corresponding to each user;

[0009] Determine whether there is sensitive information in the network parameters corresponding to the user according to the fused feature vector and intention feature corresponding to each user.

[0010] In an optional implementation manner, the extracting and fusing attribute features of the network parameters corresponding to each user to obtain a fused feature vector corresponding to each user includes:

[0011] For each user, determine a target extraction and fusion method according to a preset recognition task and the network parameters corresponding to the user;

[0012] Extract and fuse attribute features of the network parameters corresponding to the user according to the target extraction and fusion method to obtain a fused feature vector corresponding to the user.

[0013] In an alternative embodiment, the target extraction fusion method includes early fusion;

[0014] Performing attribute feature extraction and fusion on the network parameters corresponding to the user according to the target extraction fusion method to obtain a fusion feature vector corresponding to the user, including:

[0015] Performing splicing and fusion on the network parameters in each of the modal forms according to a preset splicing method;

[0016] Inputting the spliced and fused network parameters into a pre-trained feature extraction model for attribute feature extraction to obtain a fusion feature vector corresponding to the user.

[0017] In an alternative embodiment, the target extraction fusion method includes late fusion;

[0018] Performing attribute feature extraction and fusion on the network parameters corresponding to the user according to the target extraction fusion method to obtain a fusion feature vector corresponding to the user, including:

[0019] Performing attribute feature extraction on the network parameters corresponding to each of the modal forms of the user to obtain attribute features corresponding to the network parameters in each modal form;

[0020] Performing splicing and fusion on the attribute features corresponding to the network parameters in each of the modal forms according to the weight values corresponding to each of the modal forms to obtain a fusion feature vector corresponding to the user.

[0021] In an alternative embodiment, the network parameters include text parameters, image parameters, and audio parameters, and the attribute features include text features, image features, and audio features; the performing attribute feature extraction on the network parameters corresponding to each of the modal forms of the user to obtain attribute features corresponding to the network parameters in each modal form, including:

[0022] Inputting the text parameters into a pre-trained first text feature extraction model for processing to obtain the text features;

[0023] Inputting the audio parameters into a pre-trained speech feature extraction model for processing to obtain the audio features;

[0024] If the image parameters include text parameters, inputting the image parameters into a pre-trained image feature extraction model for processing to obtain initial image features, and inputting the text parameters in the image parameters into a pre-trained second text feature extraction model for processing to obtain image text features;

[0025] Concatenate and fuse the initial image features and the image text features to obtain the image features;

[0026] If the image parameters do not include text parameters, input the image parameters into a pre-trained image feature extraction model for processing to obtain the image features.

[0027] In an alternative embodiment, the intention features include emotional features, semantic logic features, and behavioral features; the extracting intention features from the network parameters and behavioral parameters corresponding to each user to obtain the intention features corresponding to each user includes:

[0028] For each user, perform emotional feature extraction and semantic logic feature extraction on the text parameters in the network parameters corresponding to the user, and perform behavioral feature extraction on the behavioral parameters corresponding to the user to obtain the emotional features, semantic logic features, and behavioral features corresponding to the user.

[0029] In an alternative embodiment, the sensitive information includes sensitive content and sensitive intention; the determining whether there is sensitive information in the network parameters corresponding to the user according to the fusion feature vectors and intention features corresponding to each user includes:

[0030] For each user, input the fusion feature vector corresponding to the user into a pre-trained sensitive content recognition model for processing to obtain a sensitive probability; the sensitive probability is the probability that the network parameters include sensitive content;

[0031] In the case where the sensitive probability exceeds a preset probability, determine that there is sensitive content in the multi-modal network parameters corresponding to the user;

[0032] Input the intention feature corresponding to the user into a pre-trained intention classification model for feature fusion and feature analysis to obtain the user intention of the user;

[0033] In the case where the user intention is a preset intention, determine that there is a sensitive intention in the multi-modal network parameters corresponding to the user.

[0034] In a second aspect, the present application provides a sensitive information recognition device, the device includes:

[0035] An acquisition module, configured to acquire network parameters and behavioral parameters in various modal forms published by each user from at least one network platform;

[0036] A processing module, configured to perform attribute feature extraction and fusion on network parameters corresponding to each of the users to obtain a fusion feature vector corresponding to each of the users, and perform intention feature extraction on the network parameters and behavior parameters corresponding to each of the users to obtain an intention feature corresponding to each of the users;

[0037] A determination module, configured to determine whether there is sensitive information in the network parameters corresponding to the users according to the fusion feature vectors and intention features corresponding to each of the users.

[0038] In a third aspect, the present application provides a server, including a processor and a memory. The memory stores a computer program that can be executed by the processor, and the processor can execute the computer program to implement the method according to any one of the foregoing embodiments.

[0039] In a fourth aspect, the present application provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the method according to any one of the foregoing embodiments is implemented.

[0040] The sensitive information recognition method, device, server, and computer-readable storage medium provided by the embodiments of the present application can obtain a fusion feature vector that combines attribute features in multiple modal forms and an intention feature that characterizes the user's intention by performing attribute feature extraction and fusion on multi-modal network parameters published by the user on the network platform, and performing intention feature extraction on the user's network parameters and behavior parameters. Thus, it is possible to determine whether there is sensitive information in the user's network parameters based on the fusion feature vector and the intention feature. In this way, network parameters in multiple modal forms can be comprehensively analyzed for feature analysis, and at the same time, the user's intention can be deeply analyzed. Therefore, potential risks can be identified in a timely manner and security hazards can be reduced.

[0041] To make the above objects, features, and advantages of the present application more obvious and understandable, the following specific embodiments are given in conjunction with the accompanying drawings and are described in detail as follows. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] To more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings required in the embodiments. It should be understood that the following drawings only show some embodiments of the present application, and therefore should not be regarded as limiting the scope. For those of ordinary skill in the art, other related drawings can be obtained based on these drawings without creative efforts.

[0043] Figure 1 Shows a block diagram of the server provided by the embodiments of the present application;

[0044] Figure 2Shows a schematic flowchart of the sensitive information recognition method provided by the embodiments of the present application;

[0045] Figure 3 Shows a schematic flowchart of extracting attribute features from network parameters;

[0046] Figure 4 Shows a schematic flowchart of the processing of the fusion feature vector by the sensitive content recognition model;

[0047] Figure 5 Shows a functional module diagram of a sensitive information recognition device provided by the embodiments of the present application.

[0048] Icons: 10 - Server; 100 - Memory; 110 - Processor; 120 - Communication module; 200 - Acquisition module; 210 - Processing module; 220 - Determination module. Detailed implementation manners

[0049] Next, the technical solutions in the embodiments of the present application will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Usually, the components of the embodiments of the present application described and shown in the accompanying drawings here can be arranged and designed in various different configurations.

[0050] Therefore, the following detailed description of the embodiments of the present application provided in the accompanying drawings is not intended to limit the scope of the present application claimed, but merely represents the selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative efforts fall within the scope of protection of the present application.

[0051] It should be noted that relational terms such as "first" and "second" are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variant thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements, but also includes other elements not expressly listed, or elements inherent to such process, method, article or device. Without further limitation, the element defined by the statement "including a..." does not exclude the existence of additional identical elements in the process, method, article or device including the element.

[0052] In the digital age, multimodal information interaction has become the main form of Internet information dissemination. Multimodal data such as text and images are intertwined, greatly enriching information expression, but also bringing huge challenges to the management of sensitive information.

[0053] In the prior art, generally, simple sensitive information recognition can be performed on single-modal data. For example, for text data, predefined sensitive words in the text can be quickly located through keyword matching algorithms, and for image data, some images with obvious sensitive features can be recognized through image recognition technology.

[0054] However, this method can only achieve basic sensitive word screening functions and preliminary image screening functions, and is only applicable to some scenarios with low real-time requirements and relatively fixed vocabulary libraries. It cannot comprehensively capture cross-modal sensitive information, such as sensitive information published in the form of speech-to-text, sensitive information with text mixed in images, etc. Therefore, a large number of potential risks cannot be identified in time, resulting in security hazards.

[0055] In addition, the prior art often only focuses on the screening of sensitive words, sensitive features, etc., and has serious deficiencies in the ability to recognize the true intentions behind the information. However, in actual applications, malicious users may also spread sensitive information through implicit expressions, sarcastic tones, etc. These contents may not directly contain sensitive features, but have certain sensitive intentions, and these sensitive intentions are often difficult to recognize. Therefore, a large number of potential risks cannot be identified in time, resulting in security hazards.

[0056] Based on this, the embodiments of the present application provide a sensitive information recognition method, device, server, and computer-readable storage medium to solve the above problems.

[0057] Figure 1 For the block diagram of the server 10 provided by the embodiments of the present application, please refer to Figure 1 . The server includes a memory 100, a processor 110, and a communication module 120. The components of the memory 100, the processor 110, and the communication module 120 are directly or indirectly electrically connected to each other to achieve data transmission or interaction. For example, these components can be electrically connected to each other through one or more communication buses or signal lines.

[0058] Among them, the memory 100 is used to store computer programs or data that can be executed by the processor. The memory 100 can be, but is not limited to, a random access memory (RAM), a read only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), etc.

[0059] The processor 110 is used to read / write the data or computer programs stored in the memory and execute the computer program to implement the sensitive information recognition method provided by the embodiments of the present application.

[0060] The communication module 120 is used to establish a communication connection between the server and other communication terminals through the network and is used to send and receive data through the network.

[0061] Optionally, the server can be a high-performance server cluster equipped with multiple high-performance CPUs to meet the needs of large-scale data processing and complex model operations. At the same time, a large-capacity memory and high-speed solid-state drives can be configured to ensure fast reading and writing of data.

[0062] In this embodiment, the server needs to have powerful graphics processing capabilities and can install high-performance GPUs, such as professional GPUs of NVIDIA, to accelerate the training and inference processes of deep learning models.

[0063] Optionally, the operating system in the server can be a Linux system (such as Ubuntu 20.04) to provide a stable and efficient operating environment. The development language can be Python, and its rich open-source libraries and tools, such as deep learning frameworks like TensorFlow and PyTorch, can be used to implement the construction and training of the model. In addition, relevant data processing libraries need to be installed, such as Pandas for data cleaning and analysis and Scikit-learn for the implementation of machine learning algorithms.

[0064] Optionally, considering the continuous changes in future application scenarios and data types, new detection algorithms, data processing modules, etc. can also be conveniently integrated through the expansion interface. Whether facing emerging social media platforms or customized requirements in special industries, the functions can be quickly expanded to adapt to different business scenarios and data characteristics. Moreover, the server can adopt technical means such as distributed computing and parallel processing to significantly improve the processing speed of multimodal information, so as to ensure that in some scenarios such as live barrage and real-time comments, information can be quickly screened and reviewed, and the spread of sensitive information can be blocked in a timely manner.

[0065] It should be understood that Figure 1 The structure shown is only a schematic diagram of the server structure, and the server may also include more or fewer components than those shown in Figure 1 or have a different configuration from that shown in Figure 1 The components shown in Figure 1 can be implemented by hardware, software, or a combination thereof.

[0066] Next, taking the server in the above Figure 1 as the execution subject, the sensitive information recognition method provided by the embodiments of the present application will be introduced exemplarily in combination with the flowchart.

[0067] Specifically, Figure 2 Please refer to Figure 2 for a flowchart of a sensitive information recognition method provided by an embodiment of the present application. The method includes:

[0068] Step S20: Collect network parameters and behavior parameters in various modal forms published by each user from at least one network platform.

[0069] Optionally, the network platform may include social media platforms, forums, news websites, etc.

[0070] Optionally, the network parameters refer to the content published by the user, and the behavior parameters may include the frequency of the user's published content, the theme, the interaction with other users, etc.

[0071] Optionally, the server may collect the network parameters and behavior parameters from at least one network platform every preset time period.

[0072] Optionally, the network parameters can be obtained through a web crawler and a data interface adaptation component. Among them, the web crawler can be used to collect network parameters in various modal forms from at least one network platform, and the data interface adaptation component can be used to ensure seamless docking with the data interfaces of different platforms to achieve stable data acquisition.

[0073] Step S21: Extract and fuse the attribute features of the network parameters corresponding to each user to obtain the fused feature vectors corresponding to each user, and extract the intention features from the network parameters and behavior parameters corresponding to each user to obtain the intention features corresponding to each user.

[0074] Step S22: Determine whether there is sensitive information in the network parameters corresponding to the user according to the fused feature vectors and intention features corresponding to each user.

[0075] Understandably, the fused feature vector fuses the attribute features in multiple modal forms, and the intention feature can reflect the user's behavior intention. Therefore, the server can conduct in-depth analysis on the network parameters based on the fused feature vector and the intention feature.

[0076] In this embodiment, although the features of different modalities seem irrelevant, through the fusion of the attribute features in the multi-modal form, it is beneficial to capture the comprehensive information that is difficult to express in a single modality, pay attention to the internal connection between the network parameters of different modalities, and improve the accuracy and robustness of the recognition or classification task.

[0077] The sensitive information recognition method provided by the embodiment of the present application can obtain the fused feature vectors that fuse the attribute features in multiple modal forms and the intention features that characterize the user's intention by extracting and fusing the attribute features of the multi-modal network parameters published by the user on the network platform and extracting the intention features from the user's network parameters and behavior parameters. Thus, it is possible to determine whether there is sensitive information in the network parameters of the user based on the fused feature vector and the intention feature. In this way, the network parameters in multiple modal forms can be comprehensively analyzed for feature analysis, and at the same time, the user's intention can be deeply analyzed. Therefore, potential risks can be identified in a timely manner and security hazards can be reduced.

[0078] Optionally, in order to facilitate the processing of network parameters, after the server collects the network parameters, it is also necessary to perform initialization processing on the network parameters.

[0079] In this embodiment, a data cleaning tool can be used to preprocess the collected network parameters to remove duplicate and invalid data.

[0080] In a possible implementation manner, the network parameters may include text parameters, image parameters, and audio parameters.

[0081] Optionally, the server can use a network crawler framework to collect text parameters from major social media platforms, news websites, forums, etc. During the data collection process, reasonable crawling rules and frequencies can be set according to actual application requirements to ensure that the collected data is representative and does not violate the platform regulations. In addition, it is also possible to cooperate with some platforms to obtain data through official data interfaces to ensure the legality and stability of the data.

[0082] Optionally, the server may collect voice parameters from an audio sharing platform, a voice assistant application, etc. Among them, for publicly available voice resources, they can be directly downloaded and format-converted. For some data that requires authorization, cooperation can be carried out with relevant institutions or enterprises to obtain it.

[0083] Optionally, the server may collect image parameters from a picture sharing website through a web crawler.

[0084] In this embodiment, the server may clean the collected text parameters, including removing HTML tags, special characters, stop words, etc. In addition, a Python library can be used for text tokenization. For example, using the NLTK (Natural Language Toolkit) library to split the text into individual words to obtain a word sequence, and then perform stemming or lemmatization on the words to unify the word form for subsequent feature extraction. Among them, the Jieba tokenization tool can be used for Chinese text tokenization.

[0085] In this embodiment, the server needs to convert the collected audio parameters into a unified sampling rate and format (such as 16 kHz, PCM format). The server can use the Librosa library to extract audio features such as Mel-Frequency Cepstral Coefficients (MFCC) to convert the audio signal into a feature vector. In addition, the server can also process the noise in the audio parameters through a noise reduction algorithm to improve the audio quality.

[0086] In this embodiment, the server can also perform size normalization processing on the image parameters to adjust all images to the same size (such as 224x224 pixels). Use the OpenCV library to perform image enhancement operations on the image parameters to improve the image clarity, such as contrast adjustment, brightness adjustment, etc. In addition, images containing text need to be screened out, text extraction is performed on this part of the images, and the extracted text is cleaned and preprocessed.

[0087] Next, a possible implementation manner is provided for how to extract and fuse the attribute features of the network parameters corresponding to each user to obtain the fused feature vector corresponding to each user.

[0088] Specifically, for each user, the server may determine the target extraction and fusion method according to the preset recognition task and the network parameters corresponding to the user, and perform attribute feature extraction and fusion on the network parameters corresponding to the user according to the target extraction and fusion method to obtain the fused feature vector corresponding to the user.

[0089] Optionally, the recognition task refers to the recognition goal to be achieved, such as high precision, high real-time performance, etc. In addition, when determining the target extraction and fusion method, it is also necessary to consider the strength of the data correlation of the network parameters in different modal forms, the size of the modal differences, etc.

[0090] Optionally, the correspondence between the recognition task and the characteristics of the network parameters and the extraction and fusion method can be set, so that the server can determine the extraction and fusion method that should be adopted currently according to this correspondence.

[0091] In a possible implementation manner, the target extraction and fusion method may include early fusion.

[0092] In this embodiment, if the data correlation of the network parameters in different modal forms is strong and the recognition task is high-precision recognition, the target extraction and fusion method may be early fusion.

[0093] In this case, the server can splice and fuse the network parameters in each modal form according to a preset splicing method, input the spliced and fused network parameters into a pre-trained feature extraction model for attribute feature extraction, and obtain the fused feature vector corresponding to the user.

[0094] Optionally, early fusion means that the original data after initialization processing is first spliced and fused, and then feature extraction is performed on it.

[0095] In this embodiment, the server can splice the word sequence of the text parameter, the feature vector of the audio parameter, and the pixel matrix of the image parameter into a large comprehensive vector according to a preset splicing method, and then input the comprehensive vector into a pre-trained feature extraction model for feature extraction.

[0096] Optionally, the feature extraction model can be a convolutional neural network or a recurrent neural network, etc.

[0097] Optionally, the splicing method can be set according to the actual application situation, such as splicing directly in sequence, or splicing with certain weights, etc.

[0098] In another possible implementation manner, the target extraction and fusion method may also include late fusion.

[0099] Optionally, if the modal forms corresponding to the network parameters have large differences and weak correlations, the target extraction and fusion method may be late fusion.

[0100] In this case, the server can separately extract the attribute features of the network parameters corresponding to each modality of the user, obtain the attribute features corresponding to the network parameters in each modality, and splice and fuse the attribute features corresponding to the network parameters in each modality according to the weight values corresponding to each modality to obtain the fused feature vector corresponding to the user.

[0101] Optionally, late fusion means that the feature extraction is first performed on the network parameters, and then the extracted features are spliced and fused.

[0102] Optionally, the weight values corresponding to each modality can be set in advance, and then the weight values can be updated according to the actual application situation. For example, during the training process of the sensitive content recognition model, the weight values are adjusted through the backpropagation algorithm or the like to minimize the loss function.

[0103] Optionally, the weighted sum method can be used to splice the attribute features according to the weight values to ensure that the attribute features in each modality are appropriately weighted according to their importance.

[0104] In an example, if there are three feature vectors A (text), B (voice), and C (image), and the corresponding weights w_A, w_B, and w_C, the spliced fused feature vector D can be calculated by the formula D = w_A * A + w_B * B + w_C * C. The weighted sum operation here ensures that the features of each modality are appropriately weighted according to their importance. Optionally, the network parameters may include text parameters, image parameters, and audio parameters, and the attribute features may include text features, image features, and audio features.

[0105] Next, a possible implementation method is provided for how to separately extract the attribute features of the network parameters corresponding to each modality of the user and obtain the attribute features corresponding to the network parameters in each modality.

[0106] In this embodiment, the server can input the text parameters into a pre-trained first text feature extraction model for processing to obtain text features, and input the audio parameters into a pre-trained speech feature extraction model for processing to obtain audio features.

[0107] Optionally, the first text feature extraction model can be a BERT (Bidirectional Encoder Representations from Transformers) model. The server can first load the pre-trained weights of the BERT model, input the initialized text parameters into the BERT model for processing, and obtain the output of the last hidden layer of the model as the feature vector of the text parameters.

[0108] Understandably, the text parameters may include multiple texts, where each text may correspond to a feature vector of a fixed length, and this feature vector can characterize the semantic, syntactic, and other attribute information of the text.

[0109] Optionally, the audio feature extraction model can be a convolutional neural network. The server can input the initialized audio parameters into the speech feature extraction model, and perform further feature extraction through convolutional layers, pooling layers, and fully connected layers.

[0110] Understandably, the audio features can characterize the attribute information such as the sound quality, pitch, and temporal dynamics of the audio. These features include the spectral characteristics of phonemes, the pitch changes of the speaker, the intonation pattern, and the continuity and rhythm of pronunciation, and can provide key acoustic information for speech recognition and speaker recognition.

[0111] In addition, considering that the image may include text, the server needs to process different images.

[0112] Specifically, if the image parameters include text parameters, the image parameters are input into a pre-trained image feature extraction model for processing to obtain initial image features, and the text parameters in the image parameters are input into a pre-trained second text feature extraction model for processing to obtain image text features. The initial image features and the image text features are concatenated and fused to obtain image features. If the image parameters do not include text parameters, the image parameters are input into a pre-trained image feature extraction model for processing to obtain image features.

[0113] Optionally, the image feature extraction model can be a ResNet model, and the server can input the initialized image parameters into the model for processing.

[0114] Optionally, the second text feature extraction model can be an OCR model. The server can input the image containing text parameters into the OCR model. The OCR model can extract the text in the image and generate word vectors. In addition, the server can also input the image into the image feature extraction model for processing to obtain initial image features, and then concatenate the initial image features with the second text features to obtain the attribute features corresponding to the image containing text parameters.

[0115] Optionally, the server can directly concatenate and fuse the initial image features and the second text features in a certain order, or can perform weighted fusion on the initial image features and the second text features through an attention mechanism, and can be specifically selected according to the actual application situation.

[0116] Optionally, the server can use text detection technology, such as the text localization module of OCR, to automatically identify whether the image contains text content.

[0117] In one example, Figure 3 For the process schematic diagram of extracting attribute features of network parameters, please refer to Figure 3 , the preprocessed text parameters can be processed through an input layer, a multi-head self-attention layer, a feed-forward neural network layer, etc., so as to obtain text features; the preprocessed audio parameters can be processed through a convolutional layer, a pooling layer and a fully-connected layer, so as to obtain audio features; the preprocessed image parameters can be processed through a convolutional layer, a pooling layer, a residual module, etc., so as to obtain image features.

[0118] Optionally, in order to further improve the sensitive information recognition effect, the feature extraction fusion can also be carried out by simultaneously adopting the methods of early fusion and late fusion.

[0119] In one example, when the features of certain modalities have an obvious impact on the task performance, for example, when it is proved by experiments that the text features corresponding to the text parameters occupy the main influential position, after early fusion of the text parameters, audio parameters and image parameters, the text parameters can be separately subjected to feature extraction, and the comprehensive features extracted after early fusion are subjected to late fusion with the text features, so as to obtain the final fusion feature vector.

[0120] In addition, the weight allocation of early fusion and late fusion can also be determined through experiments. After obtaining the fusion feature vectors through early fusion and late fusion respectively, the fusion feature vector obtained after early fusion and the fusion feature vector obtained after late fusion are weighted and calculated according to the corresponding weights, so as to further fuse them and obtain the final fusion feature vector.

[0121] Next, a possible implementation method is provided for how to extract intent features from the network parameters and behavior parameters corresponding to each user to obtain the intent features corresponding to each user.

[0122] Specifically, the intent features may include emotional features, semantic logic features, and behavior features.

[0123] The server can, for each user, extract emotional features and semantic logic features from the text parameters in the network parameters corresponding to the user, and extract behavior features from the behavior parameters corresponding to the user, so as to obtain the emotional features, semantic logic features, and behavior features corresponding to the user.

[0124] Optionally, the extraction of emotional features and language logic features can be achieved by analyzing the text parameters.

[0125] In this embodiment, the server can extract emotional features from the text parameters through a pre-trained emotional feature extraction model.

[0126] Optionally, the sentiment feature extraction model can be a BERT model. The server can input the text parameters into the model for processing, so as to output the sentiment tendency of the text, such as positive, negative, neutral, etc.

[0127] Optionally, in order to classify the text sentiment, a fully connected layer and a softmax layer can be added on the basis of the BERT model.

[0128] In this embodiment, the server can use a semantic role labeling tool and a dependency syntax analysis tool to perform semantic logic analysis on the text parameters. Among them, semantic role labeling can be used to determine the roles of each predicate in the sentence, such as subject, object, adverbial, etc.; dependency syntax analysis can be used to analyze the dependency relationships between the words in the sentence, such as subject-predicate relationship, verb-object relationship, etc.

[0129] It can be understood that through these analyses, the semantic structure and logical relationship of the text can be deeply understood, providing support for intention recognition.

[0130] In this embodiment, the server can process the behavior parameters through a clustering algorithm, classify users according to their behavior patterns, so as to construct a user behavior pattern model. For users with similar behavior patterns, the server can analyze their common features and potential intentions. Therefore, when new information is released, the server can combine the user's behavior pattern information, extract the user's behavior features, so as to judge the intention of the new information.

[0131] Next, a possible implementation manner is provided for how to determine whether there is sensitive information in the network parameters corresponding to the user according to the fusion feature vector and intention feature corresponding to each user.

[0132] Specifically, the sensitive information can include sensitive content and sensitive intention. The server can input the fusion feature vector corresponding to the user into a pre-trained sensitive content recognition model for each user for processing to obtain a sensitive probability. In the case where the sensitive probability exceeds the preset probability, it is determined that there is sensitive content in the multi-modal network parameters corresponding to the user, and the intention feature corresponding to the user is input into a pre-trained intention classification model for feature fusion and feature analysis to obtain the user's intention. In the case where the user's intention is a preset intention, it is determined that there is sensitive intention in the multi-modal network parameters corresponding to the user.

[0133] Among them, the sensitive probability is the probability that the network parameters include sensitive content.

[0134] Optionally, the sensitive content recognition model can be a model based on the combination of a convolutional neural network and a recurrent neural network. The server can input the fused feature vector into the sensitive content recognition model, extract local features through the convolutional neural network, capture context relationships through the recurrent neural network, and then perform classification through the fully connected layer and the softmax layer to determine the sensitive probability of the input data containing sensitive content.

[0135] Optionally, the sensitive content can be sensitive words, sensitive characters, and so on.

[0136] In this embodiment, the sensitive content recognition model can be trained with labeled sensitive content and non-sensitive content, and the cross-entropy loss function and the Adam optimizer are used to update the model parameters. During the training process, the training effect of the model can be improved by adjusting hyperparameters such as the learning rate and the batch size.

[0137] Optionally, to solve the problem of lagging update of the vocabulary library in the prior art, the server can use a web crawler to monitor the hot word trends of major search engines, social media platforms, etc. in real time. Crawl the hot word list regularly every day, and perform semantic analysis and part-of-speech tagging on the newly emerged words. Use natural language processing tools such as WordNet, FastText, etc. to judge whether the new words are semantically similar to the known sensitive words. If the new words have potential bad meanings, add them to the sensitive word library automatically. At the same time, conduct a manual review of the word library once a week to ensure the accuracy of the newly added words.

[0138] At the same time, the newly added data in the word library can be used for iterative training of the sensitive content recognition model to ensure the accuracy of the sensitive content recognition model.

[0139] Optionally, the sensitive word library can also be stored through a MySQL database. Create a table in the database, including fields such as sensitive words, part-of-speech, semantic categories, etc. Considering possible retrieval requirements, in order to improve retrieval efficiency, an index can also be established for the sensitive word field. When the detection module needs to determine whether a word is a sensitive word, it can perform a quick search in the database through an SQL query statement. For example, use the SELECT statement to query whether the word exists in the sensitive word table.

[0140] Optionally, considering that in different application scenarios, the requirements for the recognition accuracy of sensitive content may be different, therefore, according to different application scenarios and data distributions, the threshold for the sensitive content recognition model to recognize sensitive content, that is, the preset probability, can be dynamically adjusted. For example, in a social media platform, for the real-time content posted by users, in order to discover sensitive information in time, the threshold can be appropriately lowered to improve the recall rate of detection; while for some scenarios with higher requirements for accuracy, such as news review, the threshold can be appropriately raised to reduce the false positive rate.

[0141] In this embodiment, the threshold can be adjusted manually or automatically by monitoring indicators such as the recall rate and accuracy of the monitoring model and combining with the actual business requirements.

[0142] In a possible implementation, the server can understand the model performance under different thresholds by analyzing historical data, run the sensitive content recognition model under different thresholds through A / B testing, and compare the model output results. Then, the threshold can be iteratively adjusted based on the results of user feedback or manual review.

[0143] In the actual application process, machine learning algorithms can be used to automatically optimize the threshold. For example, reinforcement learning can be used to adjust the threshold according to reward signals (such as reducing false positives and false negatives).

[0144] Optionally, after determining that there is sensitive content in the multi-modal network parameters corresponding to the user, the server can send the network parameters corresponding to the user to the review queue, and the reviewer can view relevant information such as the network parameters and detection results through the review interface, and confirm or modify the detection results according to their own judgment.

[0145] In this embodiment, in order to ensure the model accuracy of the sensitive content recognition model, the server can also further train and optimize the sensitive content recognition model based on the review results to improve the accuracy of the model.

[0146] In one example, Figure 4 For the schematic diagram of the processing flow of the fusion feature vector by the sensitive content recognition model, please refer to Figure 4 , the sensitive content recognition model can process the fusion feature vector through a convolutional layer, a pooling layer, a recurrent neural network layer, and a fully connected layer respectively, and then output the sensitive probability through the output layer. The server can determine whether there is sensitive content such as sensitive words in the network parameters through a preset probability, and the preset probability can be adjusted by a threshold adjustment module. When it is determined that there is sensitive content such as sensitive words, the network parameters are sent to the review queue for review, and the judgment result is corrected according to the review result and the sensitive content recognition model is iteratively trained.

[0147] It can be understood that, in order to more deeply understand the true meaning of words in complex contexts, the sensitive content recognition model can accurately grasp the semantic relationship between words, analyze the context, and thus accurately judge whether the words have sensitive intentions through iterative learning of a large amount of text data and review situations. For example, when processing a text containing metaphors and implications, the model can accurately identify hidden sensitive information through in-depth mining of the context, effectively reducing the situations of false positives and false negatives.

[0148] Optionally, the intent classification model can be a model based on a decision tree or a neural network, which is used to process intent features to classify user intent, such as malicious marketing, goodwill marketing, spreading rumors, etc.

[0149] Optionally, the preset intent may be a sensitive intent stored in advance according to actual application situations, such as malicious marketing, inciting emotions, spreading rumors, etc.

[0150] In one example, the intent classification model may include an input layer, a hidden layer, and an output layer.

[0151] Among them, the sentiment feature can include the probability of positive / negative / neutral classification, the semantic logic feature can include the context embedding vector (768 dimensions), and the behavior feature can be the state vector of the user's historical behavior sequence (such as posting frequency, interaction mode) after LSTM encoding. The input layer can splice these three types of features into a unified feature vector, for example: sentiment feature 3 dimensions + semantic logic feature 768 dimensions + behavior feature 128 dimensions = 899 dimensions for input.

[0152] The neural network in the hidden layer can adopt a 3-layer fully connected network (512→256→128 dimensions) and use the GELU activation function, which is superior to the nonlinear modeling capability of traditional decision trees. At the same time, the attention mechanism can be used to add a multi-head attention module to the first layer to dynamically weight the contribution of sentiment and semantic features (such as malicious marketing intentions are more dependent on behavioral patterns).

[0153] The output layer can use Softmax to output multi-classification probabilities (malicious marketing / inciting emotions / spreading rumors / normal), and the loss function is Focal Loss to alleviate the problem of class imbalance.

[0154] In this example, the intent classification model can be trained and optimized through data augmentation, joint training, and dynamic sampling.

[0155] Among them, data enhancement refers to the semantically preserved text rewriting and behavior pattern simulation of minority class samples (such as spreading rumors); joint training refers to the partial sharing of sentiment analysis (BERT) and intent classification model parameters in the pre-training stage to enhance feature alignment; dynamic sampling refers to adjusting the sampling weight of the training set according to the real-time business data distribution (such as increasing the sampling rate of malicious marketing classes when there has been a surge in recent cases).

[0156] In this example, lightweight deployment can be used when deploying the intent classification model. Through knowledge distillation, the 3-layer network is compressed into 2 layers (256→64 dimensions), which increases the inference speed by 3 times and is suitable for real-time review.

[0157] In addition, the model output probability can be linked with the dynamic threshold system (for example, reducing the malicious marketing classification threshold during high-risk periods to improve the recall rate). For example, in social platform audits, when a user posts "Low-price luxury goods purchasing agency [Behavior: High-frequency posting + External link] [Emotion: Neutral] [Semantic: Implicit inducement to trade]" and the model outputs a malicious marketing probability of 92%, it can be determined that it has a sensitive intention and trigger an automatic interception.

[0158] To execute the corresponding steps in the above embodiments and each possible manner, an implementation manner of a sensitive information recognition device is given below. Further, please refer to Figure 5 , Figure 5 FIG. is a functional module diagram of a sensitive information recognition device provided by an embodiment of the present application. It should be noted that the basic principle and the technical effects generated by the sensitive information recognition device provided in this embodiment are the same as those in the above embodiments. For a brief description, for parts not mentioned in this embodiment, reference can be made to the corresponding content in the above embodiments. The sensitive information recognition device includes: a collection module 200, a processing module 210, and a determination module 220.

[0159] The collection module 200 is used to collect network parameters and behavior parameters in various modal forms published by each user from at least one network platform.

[0160] It can be understood that the collection module 200 can also be used to execute the above step S20.

[0161] The processing module 210 is used to extract and fuse attribute features of the network parameters corresponding to each user to obtain a fused feature vector corresponding to each user, and extract intention features from the network parameters and behavior parameters corresponding to each user to obtain an intention feature corresponding to each user.

[0162] It can be understood that the processing module 210 can also be used to execute the above step S21.

[0163] The determination module 220 is used to determine whether there is sensitive information in the network parameters corresponding to the user according to the fused feature vector and intention feature corresponding to each user.

[0164] It can be understood that the determination module 220 can also be used to execute the above step S22.

[0165] Optionally, the processing module 210 is further used to, for each user, determine a target extraction and fusion method according to a preset recognition task and the network parameters corresponding to the user; extract and fuse the attribute features of the network parameters corresponding to the user according to the target extraction and fusion method to obtain a fused feature vector corresponding to the user.

[0166] Optionally, the processing module 210 is further configured to splice and fuse the network parameters in each modal form according to a preset splicing method; input the spliced and fused network parameters into a pre-trained feature extraction model to extract attribute features, and obtain a fused feature vector corresponding to the user.

[0167] Optionally, the processing module 210 is further configured to extract attribute features from the network parameters in each modal form corresponding to the user respectively, to obtain the attribute features corresponding to the network parameters in each modal form; splice and fuse the attribute features corresponding to the network parameters in each modal form according to the weight values corresponding to each modal form, to obtain a fused feature vector corresponding to the user.

[0168] Optionally, the processing module 210 is further configured to input the text parameters into a pre-trained first text feature extraction model for processing to obtain text features; input the audio parameters into a pre-trained speech feature extraction model for processing to obtain audio features; if the image parameters include text parameters, input the image parameters into a pre-trained image feature extraction model for processing to obtain initial image features, and input the text parameters in the image parameters into a pre-trained second text feature extraction model for processing to obtain image text features; splice and fuse the initial image features and the image text features to obtain image features; if the image parameters do not include text parameters, input the image parameters into a pre-trained image feature extraction model for processing to obtain image features.

[0169] Optionally, the processing module 210 is further configured to, for each user, extract emotional features and semantic logic features from the text parameters in the network parameters corresponding to the user, and extract behavioral features from the behavioral parameters corresponding to the user, to obtain the emotional features, semantic logic features, and behavioral features corresponding to the user.

[0170] Optionally, the determining module 220 is further configured to, for each user, input the fused feature vector corresponding to the user into a pre-trained sensitive content recognition model for processing to obtain a sensitive probability; the sensitive probability is the probability that the network parameters include sensitive content; in the case where the sensitive probability exceeds a preset probability, determine that there is sensitive content in the multi-modal network parameters corresponding to the user; input the intention feature corresponding to the user into a pre-trained intention classification model for feature fusion and feature analysis to obtain the user intention of the user; in the case where the user intention is a preset intention, determine that there is a sensitive intention in the multi-modal network parameters corresponding to the user.

[0171] Optionally, the above modules may be stored in the form of software or firmware Figure 1Stored in the memory shown or in the operating system (OS) of the server, and can be executed by the Figure 1 processor therein. At the same time, data, program codes, etc. required to execute the above modules can be stored in the memory.

[0172] The embodiments of the present application also provide a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the sensitive information recognition method provided by the embodiments of the present application can be implemented.

[0173] In several embodiments provided by the present application, it should be understood that the disclosed devices and methods can also be implemented in other ways. The device embodiments described above are merely illustrative. For example, the flowcharts and block diagrams in the accompanying drawings show the possible architectures, functions, and operations of devices, methods, and computer program products according to multiple embodiments of the present application. In this regard, each block in the flowchart or block diagram may represent a module, a program segment, or a part of code, and a module, a program segment, or a part of code contains one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than marked in the accompanying drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in the reverse order, depending on the functions involved. It should also be noted that each block in the block diagram and / or flowchart, and the combination of blocks in the block diagram and / or flowchart, can be implemented by a dedicated hardware-based system for performing the specified functions or actions, or can be implemented by a combination of dedicated hardware and computer instructions.

[0174] In addition, the functional modules in each embodiment of the present application can be integrated together to form an independent part, or each module can exist alone, or two or more modules can be integrated to form an independent part.

[0175] When the above-mentioned functions are implemented in the form of software function modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art or a part of this technical solution can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of this application. The aforementioned storage medium includes: various media that can store program codes, such as USB flash drives, mobile hard disks, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), magnetic disks, or optical discs.

[0176] The above are only the preferred embodiments of this application and are not used to limit this application. For those skilled in the art, this application can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of this application shall be included within the protection scope of this application.

Claims

1. A sensitive information identification method, characterized in that: The method comprises: Collect network parameters and behavior parameters in multiple modal forms published by each user from at least one network platform; Performing attribute feature extraction and fusion on the network parameters corresponding to each of the users to obtain a fusion feature vector corresponding to each of the users, and performing intention feature extraction on the network parameters and behavior parameters corresponding to each of the users to obtain an intention feature corresponding to each of the users; According to the fused feature vectors and intention features corresponding to each of the users, it is determined whether sensitive information exists in the network parameters corresponding to the users.

2. The method according to claim 1, characterized in that: The extracting and fusing the attribute features of the network parameters corresponding to each of the users to obtain the fused feature vectors corresponding to each of the users includes: For each of the users, determining a target extraction and fusion method according to a preset recognition task and a network parameter corresponding to the user; According to the target extraction and fusion method, attribute feature extraction and fusion are performed on the network parameters corresponding to the user to obtain a fusion feature vector corresponding to the user.

3. The method according to claim 2, characterized in that The target extraction fusion method includes early fusion; The step of extracting and fusing attribute features of the network parameters corresponding to the user according to the target extraction and fusion method to obtain a fusion feature vector corresponding to the user includes: Splicing and fusing the network parameters in each of the modal forms according to a preset splicing method; The concatenated and fused network parameters are input into a pre-trained feature extraction model to extract attribute features, and a fused feature vector corresponding to the user is obtained.

4. The method according to claim 2, characterized in that: The target extraction fusion method includes late fusion; The step of extracting and fusing attribute features of the network parameters corresponding to the user according to the target extraction and fusion method to obtain a fusion feature vector corresponding to the user includes: Extracting attribute features of the network parameters in each of the modal forms corresponding to the user respectively, and obtaining attribute features corresponding to the network parameters in each of the modal forms; According to the weight values ​​corresponding to the various modal forms, the attribute features corresponding to the network parameters under the various modal forms are concatenated and fused to obtain a fused feature vector corresponding to the user.

5. The method according to claim 4, characterized in that The network parameters include text parameters, image parameters and audio parameters, and the attribute features include text features, image features and audio features; respectively extracting the attribute features of the network parameters in each modal form corresponding to the user to obtain the attribute features corresponding to the network parameters in each modal form includes: Inputting the text parameters into a pre-trained first text feature extraction model for processing to obtain the text features; Inputting the audio parameters into a pre-trained speech feature extraction model for processing to obtain the audio features; If the image parameters include text parameters, the image parameters are input into a pre-trained image feature extraction model for processing to obtain initial image features, and the text parameters in the image parameters are input into a pre-trained second text feature extraction model for processing to obtain image text features; The initial image feature and the image text feature are spliced ​​and fused to obtain the image feature; If the image parameters do not include text parameters, the image parameters are input into a pre-trained image feature extraction model for processing to obtain the image features.

6. The method according to claim 1, characterized in that The intention features include emotional features, semantic logic features and behavioral features; The extracting the intention feature of the network parameters and behavior parameters corresponding to each of the users to obtain the intention feature corresponding to each of the users includes: For each of the users, sentiment feature extraction and semantic logic feature extraction are performed on the text parameters in the network parameters corresponding to the user, and behavioral feature extraction is performed on the behavioral parameters corresponding to the user to obtain the sentiment feature, semantic logic feature and behavioral feature corresponding to the user.

7. The method according to claim 1, characterized in that The sensitive information includes sensitive content and sensitive intention; The determining, according to the fused feature vectors and the intention features corresponding to each of the users, whether there is sensitive information in the network parameters corresponding to the users includes: For each of the users, the fused feature vector corresponding to the user is input into a pre-trained sensitive content recognition model for processing to obtain a sensitive probability; the sensitive probability is the probability that the network parameters include sensitive content; When the sensitivity probability exceeds a preset probability, determining that sensitive content exists in the multimodal network parameters corresponding to the user; Inputting the intention features corresponding to the user into a pre-trained intention classification model for feature fusion and feature analysis to obtain the user intention of the user; In a case where the user intention is a preset intention, it is determined that a sensitive intention exists in the multimodal network parameters corresponding to the user.

8. A sensitive information identification device, characterized in that: The device comprises: A collection module, used to collect network parameters and behavior parameters in multiple modal forms published by each user from at least one network platform; A processing module, configured to extract and fuse attribute features of network parameters corresponding to each of the users to obtain fused feature vectors corresponding to each of the users, and to extract intention features of network parameters and behavior parameters corresponding to each of the users to obtain intention features corresponding to each of the users; The determination module is used to determine whether there is sensitive information in the network parameters corresponding to each user based on the fused feature vector and intention feature corresponding to each user.

9. A server, characterized in that: The method comprises a processor and a memory, wherein the memory stores a computer program executable by the processor, and the processor can execute the computer program to implement the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images