Multi-modal content security detection method and device, storage medium and product
Through the multimodal content security detection method, combining text, image and audio data, the characterization vectors are unified and aligned, which solves the problems of missed detection and missed detection in the face of complex and hidden content, and achieves higher detection accuracy and robustness.
Patent Information
- Application Number
- CN202411738344.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-29
- Publication Date
- 2025-02-28
AI Technical Summary
Traditional single-modal content security detection methods are prone to missed or missed when facing complex, diverse and hidden unsafe content, and are poorly adaptable to new or rapidly changing content security threats.
The multimodal content security detection method is adopted, and by obtaining multimodal data such as text, images and audio, inputting a pre-trained multimodal representation model, obtaining corresponding representation vectors, and unifying them to preset dimensions and alignments. Finally, the fused multimodal feature vector is input to the detection model to classify potential spam information and normal information.
It improves the ability to identify complex and hidden content, enhances the accuracy and robustness of detection, reduces false positives and missed reports, adapts to diverse content types, supports rapid monitoring and response, and optimizes user experience.
Smart Images
Figure CN119544343A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of data processing technology, and in particular to a multimodal content security detection method, device, computer-readable storage medium, and computer program product. Background Art
[0002] With the explosive growth of Internet content, ensuring the security of information content in cyberspace has become particularly important. Traditional single-mode content security detection can no longer meet the needs of an increasingly complex network environment.
[0003] The following problems exist in single-mode content security detection:
[0004] Limitations of single-modal detection: Traditional content security detection methods usually rely on a single data modality, such as only images, videos, audio or text, and these methods are prone to missed detections or false detections when faced with comprehensive and cross-modal information, thus failing to fully capture potentially unsafe content.
[0005] Difficulty in detecting hidden content: Hidden unsafe content, such as false information or malicious content interwoven through different modalities, is often difficult to be identified by single-modality detection methods.
[0006] Poor model adaptability: Traditional detection methods have poor adaptability to new or rapidly changing content security threats and lack the ability to automatically adjust and optimize. Summary of the invention
[0007] The object of the present invention is to provide a multimodal content security detection method, device, computer-readable storage medium and computer program product, so as to solve the above-mentioned problems existing in the prior art.
[0008] In order to achieve the above object, the technical solution adopted by the present invention is as follows:
[0009] A first aspect of the present invention provides a multimodal content security detection method, the method comprising:
[0010] Acquire multimodal data from the content to be detected, wherein the multimodal data includes: at least one of text, image, and audio;
[0011] Input the multimodal data into a pre-trained multimodal representation model to obtain a corresponding representation vector;
[0012] When the multimodal data includes at least two modal data, the representation vectors of each modal data are unified to a preset dimension and aligned, and the aligned representation vectors of each modality are fused to obtain a final representation vector of the multimodal data;
[0013] Inputting the final representation vector of the multimodality into a pre-trained extraction model to obtain a multimodal feature vector, wherein the extraction model is based on: a convolutional neural network CNN and an artificial neural network MLP training;
[0014] The multimodal feature vector is input into a pre-trained detection model to obtain a classification of potential spam information and normal information.
[0015] Exemplarily, the multimodal representation model includes: a text representation model, an image representation model and an audio representation model; the step of inputting the multimodal data into a pre-trained multimodal representation model to obtain a corresponding representation vector includes:
[0016] When the multimodal data includes text, the text is input into a pre-trained text representation model to obtain a text representation vector of a first preset dimension, wherein the text representation model is based on: BCE Embeddingbase training;
[0017] When the multimodal data includes an image, the image is input into a pre-trained image representation model to obtain an image representation vector of a second preset dimension, wherein the image representation vector is based on: ResNet34 training;
[0018] When the multimodal data includes audio, the audio is input into a pre-trained audio representation model to obtain an audio representation vector of a first preset dimension of the audio at each moment, wherein the audio representation vector is based on: Wav2Vec 2.0 training.
[0019] Exemplarily, unifying the representation vectors of the multimodal data into a preset dimension includes:
[0020] When the image representation vector of the second preset dimension is obtained, up-sampling is performed to obtain the image representation vector of the first preset dimension;
[0021] When the audio representation vector of the first preset dimension is obtained, mean normalization processing is performed on it to obtain the audio representation vector of the first preset dimension over the complete audio time.
[0022] Exemplarily, when an image representation vector of the second preset dimension is obtained, upsampling it to obtain an image representation vector of the first preset dimension includes:
[0023] The image representation vector of the second preset dimension is up-sampled by Fourier transform, specifically including:
[0024] The image representation vector p[n] of the second preset dimension is transformed into a frequency domain X[k] expression:
[0025]
[0026] Where N is the second preset dimension, and the frequency domain is expanded to a new frequency domain expression X[k]→X'[k]:
[0027]
[0028] Where M is the preset dimension. The new frequency domain X'[k] is subjected to an inverse Fourier transform to obtain a vector p[m] of the preset dimension length:
[0029]
[0030] Exemplarily, aligning representation vectors of multimodal content includes:
[0031] Input the representation vectors of each multimodal content of uniform length into the pre-trained transformation model VVM for alignment, and obtain the corresponding aligned representation vectors;
[0032] The structures of each transformation model VVM are the same, which are:
[0033] y[m]=VecVariaModel(x[m])
[0034] Among them, x(m) can be p[m], s[m], a′[m], y(m) is the corresponding vector after the output transformation, and the loss function includes the following:
[0035]
[0036] Among them, D(x,y) is used to calculate the vector x i and vector y i The distance between them, x and y are any two modal data;
[0037]
[0038] Among them, the specific loss function loss(VMM,(l,x,y)) combines the loss function of label l to deal with the situation that there may be incomplete pairs in the training data, and M represents the preset maximum distance threshold;
[0039]
[0040] Among them, the comprehensive loss function Loss(VMM,(l,p,s,a)) averages the specific loss function loss(VMM,(l,x,y)) under different input modes;
[0041] The audio transformation model VVM is obtained respectively a ,Image Transformation Model VVM p, Text Transformation Model VVM s .
[0042] Exemplarily, the convolutional neural network CNN includes: 3 convolutional layers and 3 pooling layers, and the artificial neural network MLP includes: 3 linear layers.
[0043] Exemplarily, the pre-trained detection model includes: a binary classifier MMD.
[0044] A second aspect of the present invention provides a multimodal content security detection device, the device comprising:
[0045] An acquisition module, configured to acquire multimodal data from the content to be detected, wherein the multimodal data includes at least one of text, image, and audio;
[0046] A vectorization module is used to input multimodal data into a pre-trained multimodal representation model to obtain a corresponding representation vector;
[0047] When multimodal content includes at least two modal contents, it also includes:
[0048] An alignment and fusion module, which is used to unify and align the representation vectors of each modal data to a preset dimension when the multimodal data includes at least two modal data, and to fuse the aligned representation vectors of each modal data to obtain a final representation vector of the multimodal data;
[0049] An extraction module is used to input the final representation vector of the multimodal into a pre-trained extraction model to obtain a multimodal feature vector, wherein the extraction model is based on: a convolutional neural network CNN and an artificial neural network MLP training;
[0050] The detection module is used to input the multimodal feature vector into a pre-trained detection model to obtain a classification of potential junk information and normal information.
[0051] A third aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which implements the steps of the above-mentioned method when executed by a processor.
[0052] A fourth aspect of the present invention provides a computer program product, comprising a computer program, which implements the steps of the method described above when executed by a processor.
[0053] The beneficial effects of the present invention are:
[0054] The present invention solves the limitations of traditional single-modal content security detection methods when processing complex, diverse and hidden unsafe content, and has the characteristics of high detection accuracy and fast speed. BRIEF DESCRIPTION OF THE DRAWINGS
[0055] Figure 1It is a schematic flow chart of a multimodal content security detection method proposed by the present invention;
[0056] Figure 2 It is a schematic diagram of a framework of a multimodal content security detection method proposed by the present invention;
[0057] Figure 3 It is a schematic diagram of the architecture of a multimodal content security detection device proposed by the present invention;
[0058] Figure 4 It is a schematic diagram of the structure of a multimodal content security detection device proposed by the present invention. DETAILED DESCRIPTION
[0059] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings. It should be understood that the specific implementation methods described herein are only used to explain the present invention and are not used to limit the present invention.
[0060] like Figure 1 As shown, the present invention provides a method for security detection of multimodal content on one hand, the method comprising:
[0061] Step S100: Acquire multimodal data from the content to be detected, wherein the multimodal data includes: at least one of text, image and audio.
[0062] The content to be detected can come from social platforms such as web pages, mobile terminal apps, such as Weibo, Xiaohongshu, etc., or from video playback platforms. If it is video content, images can be extracted by extracting frames, and audio content can be extracted from the video using tools such as FFmpeg. If the content to be detected includes text and images, the text and images should be extracted separately.
[0063] Step S200: input the multimodal data into a pre-trained multimodal representation model to obtain a corresponding representation vector.
[0064] Here, the multimodal representation model is composed of multiple representation models, and different modal data are input into corresponding representation models to obtain corresponding vectors. In other words, each modal data extracted from the content to be detected in step S100 is input into a pre-trained multimodal representation model for vectorization representation.
[0065] Specifically, the multimodal representation model includes: a text representation model, an image representation model and an audio representation model; the multimodal data is input into a pre-trained multimodal representation model to obtain a corresponding representation vector; including:
[0066] Step S210: When the multimodal data includes text, the text is input into a pre-trained text representation model to obtain a text representation vector of a first preset dimension, wherein the text representation model is based on: BCE Embedding base training.
[0067] In step 210, the text is input into the BCE Embedding base to generate a text representation vector of a first preset dimension, that is, a 768-dimensional representation vector.
[0068] Step S220: When the multimodal data includes an image, the image is input into a pre-trained image representation model to obtain an image representation vector of a second preset dimension, wherein the image representation vector is based on: ResNet34 training.
[0069] In step S220, the image is input into ResNet34 to generate an image representation vector of a second preset dimension, that is, a 512-dimensional representation vector.
[0070] Step S230: When the multimodal data includes audio, the audio is input into a pre-trained audio representation model to obtain an audio representation vector of a first preset dimension of the audio at each moment, wherein the audio representation vector is based on: Wav2Vec 2.0 training.
[0071] In step S230, the audio at each moment of the audio input Wav2Vec 2.0 generates an audio representation vector of a first preset dimension, that is, a 768-dimensional representation vector.
[0072] Vector unified representation method (MMCMS), the present invention provides a method that supports one or more modal data of text, image, and audio as input, and vectorizes different modal data through different representation models. The representation models selected here are listed in the following table:
[0073] Modal Model Name Representation vector dimension Model Architecture Text BCE Embedding base 768 Transformer CV Image ResNet34 512 CNN Audio Wav2Vec 2.0 768 CNN+Transformer
[0074] Step S300, when the multimodal data includes at least two modal data, the representation vectors of each modal data are unified to a preset dimension and aligned, and the aligned representation vectors of each modality are fused using a mean pooling method to obtain a final multimodal representation vector.
[0075] When any two modal data are combined in the present invention, there is a problem of inconsistent representation vector dimensions. Therefore, it is necessary to unify vectors of different dimensions. In order to facilitate the unification of vectors during feature fusion to a preset dimension, such as 768 dimensions, if the representation vector of the image 512 dimensions is to be unified to the preset dimension, it is necessary to upsample the 512-dimensional vector output by the ResNet34 model. First, the formulaic expression of the vector corresponding to different modes after passing through each modality encoder is defined. The modality encoder refers to the modality representation model. For the image P, it can be represented as a 512-dimensional vector p, p = [p[n]] = [p[0], p[1 ],p[2],……,p
[511] ],n=0,1,2,……,511,N=512;p[n] represents the nth element in vector p; for text S, it can be represented as a 768-dimensional vector s, with s=[s[m]]=[s[0],s[1],s[2],……,s
[767] ],m=0,1,2,……,767,M=768,s[m] represents the mth element in vector s, for audio A, there is a vector expression A of A at time t t =a t [m]=[a t [0],a t [1],a t [2],……,a t
[767] ],m=0,1,2,…,768,t=0,1,…,T,A t [m] represents vector A t Therefore, it is only necessary to upsample the image P once and perform mean normalization on the audio A once to unify the three modalities in the vector dimension.
[0076] Next, we use the Fourier transform-based upsampling method to upsample the 512-dimensional representation vector p[n] of image P to a 768-dimensional vector. The formula is as follows:
[0077] First, define the length N of the original vector p[n] as the second preset dimension, that is, 512, the original vector is the image representation vector of the second preset dimension, define the target length M as 768, the target length is the preset dimension, first transform the original vector p[n] into the frequency domain expression X[k]:
[0078]
[0079] Among them, e -j2πkn / N represents a complex exponential function, p[n] represents the time domain signal, N represents the signal length, and then the frequency domain X[k] is expanded to a new frequency domain expression X[k]→X′[k] as follows:
[0080]
[0081] Next, the frequency domain extended expression X′[k] is subjected to inverse Fourier transform to obtain the restored vector p[m] and the target length vector. The formula is as follows:
[0082]
[0083] Next, we will perform compression mapping on the audio. In the above description, when the input is an audio of length T, we use the audio representation model Wav2Vec 2.0, which will be referred to as the audio representation model below, and input the audio A to represent the only ordered sequence A t =A0,A1,A2,……,A T ,A t Yes t =a[m], m = 0, 1, 2, ..., 767, therefore, we need to t Process it and operate it as a vector a′[m] that is independent of the sequence, and simply make AveragePooling apply the sequence data A t Perform mean pooling from the time dimension to obtain the vector a′[m]:
[0084] a′[m]=AveragePooling(A).
[0086] The above-mentioned vector unified representation method (MMCMS) and transformer joint training (MMVM) constitute a multimodal representation fusion generation system. Through the above-mentioned vector unified representation method, the inputs of image P, text S, and audio A have been successfully transformed into corresponding vector representations p[n], s[m], a[m] through the corresponding modal representation models, and the image representation vector is represented as a vector p[m] of a uniform preset length by sampling the above-mentioned Fourier transform, and the audio representation vector is unified from the time dimension to a preset length a′[m] by using the mean pooling method.
[0087] Next, we designed a network VecVariaModel, which performs identity transformation based on vectors p[m], s[m], and a′[m], that is, the vector dimensions of input and output remain consistent. In fact, it is simple to train three models with the same structure at the same time. For this, the standard Transformers Encoder Only mode is adopted, and 5 Transformers Blockers are stacked to form a model, which maintains the identity of input and output. In other words, the transformation model VVM includes: 5 TransformersBlockers are stacked.
[0088] Therefore, the model expression can be defined as follows, which is abbreviated as VVM below:
[0089] y[m]=VecVariaModel(x[m])
[0090] Among them, x(m) can be p[m], s[m], a′[m], and y(m) is the corresponding vector after the output transformation. There can be a predefined paired data set (p, s, a, l). For the training set Train, there is a case where the data is unpaired. The label l is defined at this time to indicate whether the data is paired. The loss function is defined as follows:
[0091]
[0092] Where D(x,y) is a simple Euclidean distance metric function used to calculate the vector x i and vector y i The difference between them is used to calculate the distance between each pair of data, where x and y are any two modal data in p, s, and a.
[0093]
[0094] Among them, the specific loss function loss(VMM,(l,x,y)) is a loss function combined with the label l, which is used to deal with the situation where there may be incomplete pairs in the training data and calculate the loss of each pair of data. Specifically: when l=1, it means that x and y should be paired data. At this time, the loss function mainly focuses on D(x,y) 2 When l = 0, it means that x and y are not paired data, and the loss function mainly focuses on (1-l)max(MD(x,y),0) 2 to penalize the case where the distance is too small. Here M can be understood as a preset maximum distance threshold. If D(x,y) is less than M, the model will be penalized because in this case, the model should not try to map two unrelated inputs too close. Paired data refers to data sample pairs constructed from manually annotated related pictures p, audio a, and text s, that is, l is 1. Unpaired data refers to pictures p, audio a, and text s that are manually marked as unrelated, that is, l is 0.
[0095]
[0096] According to the calculated loss value, the model parameters are updated by back propagation. The model VVM is trained by minimizing the comprehensive loss Loss(VMM,(l,p,s,a)), and finally three representation models VVM a, VVM p and VVM s adapted to different modalities are obtained. Loss(VMM,(l,p,s,a)) is a comprehensive loss function that averages the specific loss function loss(VMM,(l,x,y)) under different input modalities. This means that the model will consider the three input modalities of p[m], s[m] and a′[m] at the same time during training, and try to optimize the performance on all these modalities.
[0097] D(x,y) is a basic distance metric used to directly compare the similarity of two vectors. loss(VMM,(l,x,y)) combines label information and can flexibly handle paired and unpaired data, ensuring that the model can show good generalization ability when processing different types of data. Loss(VMM,(l,p,s,a)) is a comprehensive loss function designed to enable the model to learn the relationship between multiple input modalities at the same time and improve the model's multi-task learning ability.
[0098] The design of this loss function allows the model to effectively learn and generalize when faced with complex and potentially incompletely matching datasets.
[0099] Based on this, three VVM pairs of representation models adapted to different modalities can be trained and labeled as VVM a , VVM p , VVM s That is, unpaired datasets and paired datasets are prepared, each sample of which has a corresponding label l, for training the three models to ensure that they can show good performance when processing data of their respective modalities.
[0100] Through the above, three transformation models VVM adapted to different modalities are obtained. Through contrast learning training, the information of different modalities is aligned. When the user inputs image P, audio A, and text S, the representation model and the corresponding sampling method are used, and VVM a , VVM p , VVM s , the alignment of multimodal information is achieved. Next, the input vector needs to be classified into two categories through a joint method.
[0101] Assume that the user's input data is a (image, audio, text) data pair, expressed using a tuple (P, S, A). After the expression through MMCMS, a vector tuple (VVM a , VVM p , VVM s), simply average pool these three vectors to get a 768-dimensional vector, k refers to the tuple type of user input. When the user inputs (image, audio, text) data pairs, k num is 3, k is (p, a, s). When the user inputs (image, audio) data pair, k num is 2, k is (p, a), and so on. The formula is as follows:
[0102]
[0103] After VVM fusion, the final multi-modal representation vector is output.
[0104] Step S400: input the final representation vector of the multimodality into a pre-trained extraction model to obtain a multimodal feature vector, wherein the extraction model is based on: convolutional neural network CNN and artificial neural network MLP training.
[0105] Here, the extraction model is equivalent to a feature transformer, using a convolutional neural network CNN (3 layers of convolution + 3 layers of pooling) + artificial neural network MLP (3 linear layers) to quickly extract features from the aligned modal information. It is necessary to use a large-scale mixture of multimodal junk information and normal information for representation training, mainly to enhance the representation capabilities of the MMCMS system. The convolutional neural network CNN includes: 3 layers of convolution + 3 layers of pooling, and the artificial neural network MLP includes: 3 linear layers.
[0106] Step S500: input the multimodal feature vector into a pre-trained detection model to obtain classification of potential spam information and normal information.
[0107] The detection model is a binary classifier MMD, which is equivalent to a vector decoder. It uses a simple MMD classifier for fast classification, improves system response efficiency, quickly mines potential spam information, and provides a basis for subsequent interception.
[0108] The present invention is mainly applicable to services in the field of telecommunication operators and web dialing detection, mainly involving content security, multimodal (Multi Modal), and natural language processing (NLP). The detection model, that is, the binary classifier MMD detects potential spam and normal information, collects and classifies data according to actual conditions, and trains the binary classifier, which will not be repeated here. The disclosed embodiment uses a representation model to represent audio / video / image / text as a vector of uniform size through the MMCS method, and uses the fused multimodal vector representation information as the embedding layer input of the large model. The multimodal vector decoding model MMD is used to generate detection results. At the same time, the model return content and result feedback are fed back to the system, and the parameters and detection strategies of each module are automatically adjusted to continuously optimize the detection accuracy of the system.
[0109] Step S600: The detected spam and normal information may also be manually re-labeled and fed back to the system. The re-labeled information may be used as a training set to retrain the detection model.
[0110] The present invention solves the limitations of traditional single-modal content security detection methods in dealing with complex, diverse and hidden unsafe content. Specifically, the present invention aims to overcome the following shortcomings in the prior art:
[0111] 1. Limitations of single-modal detection: Traditional content security detection methods usually rely on a single data modality (such as images, videos, audio or text only), and these methods are prone to missed detections or false detections when faced with comprehensive and cross-modal information, thus failing to fully capture potentially unsafe content.
[0112] 2. Insufficient fusion of multimodal information: Existing technologies usually lack an effective fusion mechanism when processing information of different modalities and cannot fully utilize the comprehensive information of multiple data sources such as images, videos, audio and text, resulting in insufficient accuracy and comprehensiveness of content detection.
[0113] 3. Difficulty in detecting hidden content: Hidden unsafe content (such as false information or malicious content interwoven through different modalities) is often difficult to be identified by a single-modality detection method. The present invention improves the ability to identify such complex content through multi-modal fusion.
[0114] 4. Poor model adaptability: Traditional detection methods have poor adaptability to new or rapidly changing content security threats and lack the ability to automatically adjust and optimize. The present invention uses a feedback mechanism to enable the system to automatically learn and adapt to new content security challenges.
[0115] The multimodal fusion technology adopted in the present invention has the characteristics of small model and fast calculation.
[0116] like Figure 4 As shown, another aspect of the present invention provides a multimodal content security detection device, the device comprising:
[0117] The acquisition module 100 is used to acquire multimodal data from the content to be detected, wherein the multimodal data includes: at least one of text, image and audio;
[0118] A vectorization module 200 is used to input multimodal data into a pre-trained multimodal representation model to obtain a corresponding representation vector;
[0119] When multimodal content includes at least two modal contents, it also includes:
[0120] The alignment and fusion module 300 is used to unify and align the representation vectors of each modality data to a preset dimension when the multimodal data includes at least two modal data, and to fuse the aligned representation vectors of each modality using a mean pooling method to obtain a final representation vector of the multimodality;
[0121] The extraction module 400 is used to input the final representation vector of the multimodal state into a pre-trained extraction model to obtain a multimodal feature vector, wherein the extraction model is based on: a convolutional neural network CNN and an artificial neural network MLP training;
[0122] The detection module 500 is used to input the multimodal feature vector into a pre-trained detection model to obtain a classification of potential spam information and normal information.
[0123] The content security detection method based on the multimodal fusion of the present invention has significant beneficial effects in improving detection accuracy, enhancing robustness, reducing false positives and false negatives, adapting to diverse content types, supporting rapid monitoring and response, and optimizing user experience, and can play an important role in the field of content security, with the following specific aspects:
[0124] 1. Enhanced detection accuracy: This method performs content security detection by combining multiple modal information such as images, text, and voice. The image modality provides visual information, the text modality provides semantic content, and the voice modality provides audio context. By fusing information from these different modalities, it is possible to more comprehensively understand and identify potential security threats in the content, thereby significantly improving detection accuracy. For example, when detecting malicious content, a single modality may miss detection or misjudge, while multimodal fusion can supplement information missing and eliminate errors, improving overall recognition capabilities.
[0125] 2. Improve robustness: When facing variable input conditions or extreme environments, such as low-quality images, noisy text or fuzzy speech, multimodal fusion methods can effectively improve the robustness of content security detection systems. Even if the information quality of one modality is poor, the supplementary information of other modalities can make up for this deficiency, thereby maintaining stable system performance and high detection accuracy.
[0126] 3. Reduce false positives and false negatives: Combining data from multiple modalities helps to more accurately judge the safety of content. By comprehensively analyzing the visual features in images, the semantic information in text, and the sound features in speech, the incidence of false positives (normal content is mistakenly marked as harmful content) and false negatives (harmful content is not detected) can be reduced. This comprehensive judgment capability improves the overall reliability of the system.
[0127] 4. Adapt to diverse content types: Modern content security detection faces various types of content, including images, videos, texts, and voices. Multimodal fusion methods can adapt to the diversity of these contents and achieve unified detection of various content forms through comprehensive analysis of various modalities. This method can provide consistent security detection standards in multiple application scenarios to meet the needs of different industries and platforms.
[0128] 5. Support rapid monitoring and response: Multimodal fusion technology can realize real-time analysis and processing of content. Through efficient data fusion and feature extraction, potential threats in the content can be quickly detected, and alarms can be issued in time or necessary response measures can be taken. This real-time monitoring capability is of great significance for rapid response and prevention of content security.
[0129] 6. Optimize user experience: By improving detection accuracy and reducing false alarm rates, the multimodal fusion method can provide more accurate content screening services, avoid misjudgment and interference with normal content, and thus optimize user experience. While enjoying high-quality content, users can experience a safer and more undisturbed usage environment.
[0130] Another aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which implements the steps of the above method when executed by a processor.
[0131] Another aspect of the present invention provides a computer program product, comprising a computer program, which implements the steps of the above method when executed by a processor.
[0132] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be considered as the scope of protection of the present invention.
Claims
1. A method for security detection of multimodal content, characterized in that: The method comprises: Acquire multimodal data from the content to be detected, wherein the multimodal data includes: at least one of text, image, and audio; Input the multimodal data into a pre-trained multimodal representation model to obtain a corresponding representation vector; When the multimodal data includes at least two modal data, the representation vectors of each modal data are unified to a preset dimension and aligned, and the aligned representation vectors of each modality are fused to obtain a final representation vector of the multimodal data; Inputting the final representation vector of the multimodality into a pre-trained extraction model to obtain a multimodal feature vector, wherein the extraction model is based on: a convolutional neural network CNN and an artificial neural network MLP training; The multimodal feature vector is input into a pre-trained detection model to obtain a classification of potential spam information and normal information.
2. The method according to claim 1, characterized in that The multimodal representation model includes: a text representation model, an image representation model and an audio representation model; the multimodal data is input into a pre-trained multimodal representation model to obtain a corresponding representation vector; including: When the multimodal data includes text, the text is input into a pre-trained text representation model to obtain a text representation vector of a first preset dimension, wherein the text representation model is based on: BCE Embedding base training; When the multimodal data includes an image, the image is input into a pre-trained image representation model to obtain an image representation vector of a second preset dimension, wherein the image representation vector is based on: ResNet34 training; When the multimodal data includes audio, the audio is input into a pre-trained audio representation model to obtain an audio representation vector of a first preset dimension of the audio at each moment, wherein the audio representation vector is based on: Wav2Vec 2.0 training.
3. The method according to claim 2, characterized in that The step of unifying the representation vectors of the multimodal data into a preset dimension includes: When the image representation vector of the second preset dimension is obtained, up-sampling is performed to obtain the image representation vector of the first preset dimension; When the audio representation vector of the first preset dimension is obtained, mean normalization processing is performed on it to obtain the audio representation vector of the first preset dimension over the complete audio time.
4. The method according to claim 3, characterized in that When the image representation vector of the second preset dimension is obtained, upsampling is performed to obtain the image representation vector of the first preset dimension, including: The image representation vector of the second preset dimension is up-sampled by Fourier transform, specifically including: The image representation vector p[n] of the second preset dimension is transformed into a frequency domain X[k] expression: Where N is the second preset dimension, and the frequency domain is expanded to a new frequency domain expression X[k]→X'[k]: Where M is the preset dimension. The new frequency domain X'[k] is subjected to an inverse Fourier transform to obtain a vector p[m] of the preset dimension length:
5. The method according to any one of claims 1 to 4, characterized in that: Align the representation vectors of multimodal content, including: Input the representation vectors of each multimodal content of uniform length into the pre-trained transformation model VVM for alignment, and obtain the corresponding aligned representation vectors; The structures of each transformation model VVM are the same, which are: y[m]=VecVariaModel(x[m]) Among them, x(m) can be p[m], s[m], a′[m], y(m) is the corresponding vector after the output transformation, and the loss function includes the following: Among them, D(x,y) is used to calculate the vector x i and vector y i The distance between them, x and y are any two modal data; Among them, the specific loss function loss(VMM,(l,x,y)) combines the loss function of label l to deal with the situation that there may be incomplete pairs in the training data, and M represents the preset maximum distance threshold; Among them, the comprehensive loss function Loss(VMM,(l,p,s,a)) averages the specific loss function loss(VMM,(l,x,y)) under different input modes; The audio transformation model VVM is obtained respectively a ,Image Transformation Model VVM p, Text Transformation Model VVM s .
6. The method according to any one of claims 1 to 5, characterized in that: The convolutional neural network CNN includes: 3 convolutional layers and 3 pooling layers, and the artificial neural network MLP includes: 3 linear layers.
7. The method according to any one of claims 1 to 5, characterized in that: The pre-trained detection model includes: a binary classifier MMD.
8. A multimodal content security detection device, characterized in that: The device comprises: An acquisition module, configured to acquire multimodal data from the content to be detected, wherein the multimodal data includes at least one of text, image, and audio; A vectorization module is used to input multimodal data into a pre-trained multimodal representation model to obtain a corresponding representation vector; When multimodal content includes at least two modal contents, it also includes: An alignment and fusion module, which is used to unify and align the representation vectors of each modal data to a preset dimension when the multimodal data includes at least two modal data, and to fuse the aligned representation vectors of each modal data to obtain a final representation vector of the multimodal data; An extraction module is used to input the final representation vector of the multimodal into a pre-trained extraction model to obtain a multimodal feature vector, wherein the extraction model is based on: a convolutional neural network CNN and an artificial neural network MLP training; The detection module is used to input the multimodal feature vector into a pre-trained detection model to obtain a classification of potential junk information and normal information.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.
10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.