Broadcast television data safety monitoring management system

Abstract

The invention discloses a broadcast television data safety monitoring management system. The method comprises the following steps: acquiring wide initial multi-source broadcast television data; recognizing abnormal information in the image information based on a CNN convolutional neural network, recognizing abnormal character information in the voice information by using a DeBERTa-V3 character and word recognition model, and establishing a data security recognition model; using an improved IPSO particle swarm algorithm to optimize hyper-parameters in the data security identification model, inputting the initial multi-source broadcast television data into a target data security identification model for identification, and obtaining a broadcast television security evaluation index; and performing state evolution prediction according to the radio and television safety evaluation index by utilizing an HMM (Hidden Markov Model). The broadcast security risk change trend can be mastered in real time, and potential security threats can be pre-judged in advance.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of big data analysis, in particular to a broadcast television data security monitoring management system. BACKGROUND

[0002] The broadcast television industry has fully entered the era of digitization and networking, and the scale of data transmission and storage has grown explosively. Traditional rule-based network security protection methods are difficult to cope with new attack patterns and cannot accurately identify abnormal network traffic and protocol vulnerabilities. In terms of content, images, voice and text data in broadcast television programs may contain illegal content and false information. Existing content review methods rely heavily on manual sampling, which is inefficient and prone to omissions. Meanwhile, user behavior data contains personal privacy information, and improper use or disclosure can have serious consequences. However, there is currently a lack of systematic monitoring and analysis mechanisms for user behavior security risks. SUMMARY

[0003] The purpose of the present application is to solve the above problems by designing a broadcast television data security monitoring management system.

[0004] To achieve the above purpose, the technical solution of the present application is as follows: further, in the above broadcast television data security monitoring management system, the broadcast television data security monitoring management system comprises: a broadcast data acquisition module for acquiring network layer data, content layer data and behavior layer data in a broadcast television tower, obtaining multi-source broadcast television data, preprocessing the multi-source broadcast television data to obtain initial multi-source broadcast television data; an identification model establishment module for identifying abnormal information in image information based on a CNN convolutional neural network, identifying abnormal text information in voice information using a DeBERTa-V3 word recognition model, and establishing a CNN-DeBERTa data security identification model; an identification model optimization module for optimizing hyperparameters in the CNN-DeBERTa data security identification model using an improved IPSO particle swarm algorithm, and adding a Transformer encoder in the security identification model to obtain a target CNN-DeBERTa data security identification model; a broadcast security evaluation module for inputting the initial multi-source broadcast television data into the target CNN-DeBERTa data security identification model for identification to obtain a broadcast television security evaluation index; a security monitoring management module for using a HMM hidden Markov model to perform state evolution prediction based on the broadcast television security evaluation index, and calculating a risk evolution probability value in real time, and triggering a broadcast television intervention strategy when the risk evolution probability value is greater than a threshold value.

[0005] Further, in the above broadcast television data security monitoring management system, the broadcast data acquisition module comprises the following sub-modules: A deletion sub-module is configured to delete duplicate data packets and error check data in network layer data in multi-source data, delete damaged image data, audio data that cannot be decoded, and text information with garbled codes in content layer data, and delete duplicate operation records and abnormal login logs in behavior data. A correction sub-module is configured to fill in missing data using a data interpolation method, identify and correct abnormal values, and obtain corrected broadcast television data. A unification sub-module is configured to convert data of different sources and different formats in the corrected broadcast television data into a unified format to obtain unified broadcast television data. A normalization sub-module is configured to normalize the unified broadcast television data by a Z-score normalization method to obtain initial multi-source broadcast television data.

[0006] Further, in the above broadcast television data security monitoring management system, the identification model establishment module comprises the following sub-modules: An input layer sub-module is configured to receive preprocessed image data based on an input layer of a CNN convolutional neural network, and convert the image into a three-dimensional tensor. A convolution layer sub-module is configured to set a plurality of convolution layers, and extract image features using convolution kernels of different numbers and sizes. A pooling layer sub-module is configured to use a maximum pooling layer after each convolution layer, with a pooling kernel size of 2x2 and a step size of 2, to downsample the feature map. A fully connected layer sub-module is configured to flatten the feature map into a one-dimensional vector after a plurality of convolution and pooling layers, and input the one-dimensional vector into a fully connected layer.

[0007] Further, in the above broadcast television data security monitoring management system, the identification model establishment module comprises the following sub-modules: A conversion sub-module is configured to convert speech data into a text sequence based on a DeepSpeech end-to-end speech recognition model, post-process the speech recognition result, and correct errors in character recognition and sentence segmentation. An encoding sub-module is configured to input the converted text sequence into a DeBERTa-V3 model, input text Token embedding, segment embedding, and position embedding, encode the text through a multi-layer Transformer encoder, and obtain a context representation of each word. A classification sub-module is configured to use a classifier to classify the context representation, and determine whether there is abnormal text information.

[0008] Further, in the above broadcast television data security monitoring management system, the identification model optimization module comprises the following sub-modules: An improved sub-module is used for introducing a dynamic inertia weight based on the PSO particle swarm algorithm, and the inertia weight is linearly decreased from 0.9 to 0.4 with the increase of the number of iterations, and a Gaussian mutation operator is used to mutate the position of the particle to obtain an improved PSO particle swarm algorithm. A determination sub-module is used for determining the hyperparameters to be optimized, including the learning rate of the CNN convolutional neural network, the decay rate of the BatchNormalization, the number of neurons of the fully connected layer, the fine-tuning learning rate of the DeBERTa-V3 model, the number of attention heads, and the hidden layer dimension. An optimization sub-module is used for representing each particle as a combination of hyperparameters, and the position vector of the particle corresponds to the value of the hyperparameter. The fitness value of each particle is calculated in each iteration, and the speed and position of the particle are updated. The optimal hyperparameter combination is obtained through iteration.

[0009] Further, in the above broadcast television data security monitoring management system, the broadcast safety evaluation module comprises the following sub-modules: An input sub-module is used for inputting the initial multi-source broadcast television data into the target CNN-DeBERTa data security identification model to obtain image anomaly probability, voice anomaly probability and content anomaly probability respectively. A calculation sub-module is used for calculating network security indicators and user behavior security indicators based on network layer data and behavior data, and a broadcast television safety evaluation index is calculated by using weighted average calculation.

[0010] Further, in the above broadcast television data security monitoring management system, the safety monitoring management module comprises the following sub-modules: A definition sub-module is used for defining the state set of HMM, and the safe state represents that the broadcast television system is running normally, the early warning state represents that the system has certain security risks, and the dangerous state represents that the system faces high security risks. A setting sub-module is used for setting the broadcast television safety evaluation index as the observation value of HMM, and the observation value sequence is composed of the safety evaluation index output by the model in real time. An evaluation sub-module is used for estimating the state transition probability matrix and the observation probability matrix through the HMM model, calculating the probability of evolving from the current state to the dangerous state according to the current observation value sequence and the model parameters, and obtaining the risk evolution probability value.

[0011] Further, in the method for implementing the above broadcast television data security monitoring management system, the method comprises the following steps: Obtaining network layer data, content layer data and behavior layer data in a broadcast television tower, obtaining multi-source broadcast television data, preprocessing the multi-source broadcast television data, and obtaining initial multi-source broadcast television data; Based on the CNN convolutional neural network, abnormal information in image information is recognized, and the DeBERTa-V3 word recognition model is used to recognize abnormal text information in voice information, and a CNN-DeBERTa data security recognition model is established. The improved IPSO particle swarm algorithm is used to optimize the hyperparameters in the CNN-DeBERTa data security recognition model, and a Transformer encoder is added to the security recognition model to obtain a target CNN-DeBERTa data security recognition model. The initial multi-source broadcast television data is input into the target CNN-DeBERTa data security recognition model for recognition, and a broadcast television security evaluation index is obtained. Using the HMM hidden Markov model, state evolution prediction is performed according to the broadcast television security evaluation index, and a risk evolution probability value is calculated in real time. When the risk evolution probability value is greater than a threshold value, a broadcast television intervention strategy is triggered.

[0012] Further, in the method for implementing the above broadcast television data security monitoring management system, the method comprises the following steps: The repeated data packets and error check data in the network layer data of the multi-source data are deleted, the damaged image data, the audio data that cannot be decoded, and the text information with garbled codes in the content layer data are deleted, and the repeated operation records and abnormal login logs in the behavior data are deleted. The missing data is filled by using the data interpolation method, and the abnormal values are identified and corrected to obtain corrected broadcast television data. The data of different sources and different formats in the corrected broadcast television data are converted into a unified format to obtain unified broadcast television data. The unified broadcast television data is normalized by using the Z-score normalization method to obtain the initial multi-source broadcast television data.

[0013] Further, in the method for implementing the above broadcast television data security monitoring management system, the method comprises the following steps: Based on the DeepSpeech end-to-end speech recognition model, the voice data is converted into a text sequence, and the speech recognition result is post-processed to correct errors and sentence errors. The converted text sequence is input into the DeBERTa-V3 model, the input text Token embedding, segment embedding and position embedding are input, and the text is encoded by a multi-layer Transformer encoder to obtain the context representation of each word. classifying the context representation using a classifier to determine whether there is anomalous textual information.

[0014] The beneficial effects are: 1. The safety hazards can be captured from multiple dimensions, greatly improving the comprehensiveness and accuracy of data processing. 2. The recognition accuracy and generalization ability of the model for image abnormal information, voice abnormal dialogue and text illegal content are improved, effectively avoiding the missed detection and misjudgment of security vulnerabilities. 3. The safety risk change trend can be grasped in real time, and potential security threats can be predicted in advance. Compared with the traditional static evaluation method, dynamic monitoring and active defense of safety risks are realized. 4. The targeted broadcast television intervention strategy is automatically triggered, covering data blocking, device isolation, content review and other measures. This intelligent intervention mechanism can quickly respond to security incidents, effectively reduce the probability of safety accidents, reduce the loss caused by security incidents, and protect the safe and stable operation of the broadcast television system and the safety of user information. BRIEF DESCRIPTION OF DRAWINGS

[0015] Various other advantages and benefits will become apparent to those of ordinary skill in the art upon reading the following detailed description of the preferred embodiments with reference made to the accompanying drawings. The drawings are for purposes of illustration only and are not intended to be limiting of the present application.

[0016] Figure 1 The first embodiment of the broadcast television data security monitoring management system in the embodiment of the present application is shown in the schematic diagram. Figure 2 The second embodiment of the broadcast television data security monitoring management system in the embodiment of the present application is shown in the schematic diagram. Figure 3 The third embodiment of the broadcast television data security monitoring management system in the embodiment of the present application is shown in the schematic diagram. DETAILED DESCRIPTION

[0017] In order to make the purpose, technical scheme and advantages of the present application more clear, the present application is further described in detail below with reference to the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.

[0018] Those skilled in the art can understand that, unless specifically stated otherwise, the singular forms "a", "an" and "the" used herein can also include the plural forms. It should be further understood that the use of the word "comprise" in the specification of the present application means that the features, integers, steps, operations, elements and / or components exist, but do not exclude the existence or addition of one or more other features, integers, steps, operations, elements, components and / or groups thereof.

[0019] The present application will be described in detail below with reference to the drawings, such as Figure 1As shown, the broadcast television data security monitoring management system comprises: The broadcast data acquisition module is configured to acquire network layer data, content layer data and behavior layer data in a broadcast television tower, obtain multi-source broadcast television data, pre-process the multi-source broadcast television data, and obtain initial multi-source broadcast television data. Specifically, the embodiment further comprises a deletion sub-module configured to delete duplicate data packets and error check data in the network layer data of the multi-source data, delete damaged image data, audio data that cannot be decoded and text information with garbled codes in the content layer data, and delete duplicate operation records and abnormal login logs in the behavior data. The correction sub-module is configured to fill in missing data by using a data interpolation method, identify and correct abnormal values, and obtain corrected broadcast television data. The unification sub-module is configured to convert data of different sources and different formats in the corrected broadcast television data into a unified format, and obtain unified broadcast television data. The normalization sub-module is configured to normalize the unified broadcast television data by using a Z-score normalization method, and obtain initial multi-source broadcast television data.

[0020] Specifically, I. Data acquisition Network layer data: Customized data collection agent programs are deployed in core network devices of the broadcast television tower, such as core switches, routers, firewalls, etc. The running state information of the devices is obtained by using SNMP (Simple Network Management Protocol) at regular intervals, including IP address, port number, network protocol (TCP, UDP, HTTP, etc.), data transmission rate, data packet quantity, error packet rate, etc. At the same time, the network traffic is mirrored to the data collection server in real time by using the mirror port technology, and the traffic data is analyzed by using Wireshark and other packet capture tools to extract detailed information of the network layer.

[0021] Content layer data: Image information: Image data is collected from the encoder of the broadcast television signal, the streaming media server, the broadcast system and other devices. For live signals, the original video stream is obtained in real time by using a video capture card; for stored program content, video files are read from the video file server. Key frames of the image, as well as the resolution, encoding format (H.264, H.265, etc.), frame rate and other metadata of the image are extracted.

[0022] Voice information: Voice data is extracted from the audio encoder, live stream and program storage server. For live audio, the audio stream is obtained in real time by using an audio capture device; for stored audio files, the audio files are read from the audio file server. At the same time, the sampling rate, bit depth, number of channels and other metadata of the voice are obtained.

[0023] Text information: including program subtitles, program introductions, user reviews and other text content. Collect text data from data sources such as subtitle servers, program management systems, and user interaction platforms.

[0024] Behavior data: integrate behavior data collection modules on user terminal devices (set-top boxes, smart TVs, mobile phone APPs, etc.) to record user operation behaviors, including user login time, login account, program switching time, viewing duration, fast forward and fast backward operations, search keywords, and data access records (program URL accessed, access frequency, etc.). At the same time, record the operation behaviors of administrators in the management system of the broadcast television tower, such as system configuration modification, data backup and recovery, etc.

[0025] II. Data preprocessing Data cleaning: establish a data cleaning rule library to clean the collected multi-source data. For network layer data, remove duplicate data packets and incorrect checksum data; for content layer data, remove damaged image files, audio data that cannot be decoded, and text information with garbled characters; for behavior data, remove duplicate operation records and abnormal login logs (multiple failed login attempts in a short period of time). Use data interpolation method, mean filling method, etc. to fill in missing data, and identify and correct abnormal values (data transmission rate is negative).

[0026] Data format conversion: convert data of different sources and different formats into a unified format. For example, convert image data to JPEG, PNG, etc., convert voice data to WAV, MP3, etc., and convert text data to UTF-8 encoded plain text format. For network layer data, parse the binary data obtained by packet capture into JSON or CSV format for subsequent processing.

[0027] Data normalization: use the minimum-maximum normalization method or Z-score normalization method to normalize numerical data and eliminate the influence of data dimension. For example, normalize numerical data such as data transmission rate and viewing duration to the [0, 1] interval or convert it to standard normal distribution data with a mean of 0 and a standard deviation of 1.

[0028] The recognition model establishment module is used to recognize abnormal information in image information based on a CNN convolutional neural network, recognize abnormal text information in voice information using a DeBERTa-V3 word recognition model, and establish a CNN-DeBERTa data security recognition model. Specifically, the input layer submodule is used to receive preprocessed image data based on the input layer of the CNN convolutional neural network, and convert the image to a three-dimensional tensor. a convolutional layer submodule configured to set a plurality of convolutional layers to extract image features using different numbers and sizes of convolutional kernels; a pooling layer submodule configured to use a max pooling layer after each convolutional layer to down-sample a feature map using a pooling kernel size of 2x2 and a step size of 2; a fully connected layer submodule configured to flatten a feature map into a one-dimensional vector after a plurality of convolutional and pooling layers and input the one-dimensional vector into a fully connected layer.

[0029] a conversion submodule configured to convert voice data into a text sequence based on a DeepSpeech end-to-end speech recognition model, post-process a speech recognition result, and correct errors in characters and punctuation; an encoding submodule configured to input the converted text sequence into a DeBERTa-V3 model, input text Token embedding, segment embedding, and position embedding, encode the text through a multi-layer Transformer encoder, and obtain a context representation of each word; a classification submodule configured to classify the context representation using a classifier to determine whether there is abnormal text information.

[0030] Specifically; I. Image Abnormal Information Recognition (CNN Convolutional Neural Network); Network structure design: a deep CNN network is constructed, which includes an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer.

[0031] Input layer: receiving pre-processed image data, converting the image into a three-dimensional tensor (height, width, channel number), and for an RGB image, the channel number is 3.

[0032] Convolutional layer: a plurality of convolutional layers are set to extract image features using different numbers and sizes of convolutional kernels. For example, the first convolutional layer uses 64 3x3 convolutional kernels with a step size of 1 and padding of 1 to preserve edge information of the image; the second convolutional layer uses 128 3x3 convolutional kernels with a step size of 1 and padding of 1; and the third convolutional layer uses 256 3x3 convolutional kernels with a step size of 1 and padding of 1. A BatchNormalization layer and a ReLU activation function are connected after each convolutional layer to speed up model convergence and avoid gradient disappearance.

[0033] Pooling layer: a max pooling layer is used after each convolutional layer to down-sample a feature map using a pooling kernel size of 2x2 and a step size of 2, thereby reducing the amount of data while retaining important features.

[0034] Fully connected layer: After multiple convolution and pooling layers, the feature map is flattened into a one-dimensional vector and input into the fully connected layer. Two fully connected layers are set, the first layer has 1024 neurons, and the second layer has 2 neurons (normal and abnormal), and the output layer uses the Softmax activation function to calculate the probability of the image belonging to normal or abnormal.

[0035] Training data preparation: Collect a large amount of normal and abnormal image data, including images with illegal content, unclear, distorted screen, etc. Perform data augmentation on image data, such as random rotation, flipping, scaling, adding noise, etc. to expand the training data scale and improve the generalization ability of the model.

[0036] Model training: Use the cross-entropy loss function as the loss function, and use the Adam optimizer to update the model parameters. Set appropriate learning rate, batch size and training rounds, monitor loss value and accuracy during training, and avoid model overfitting through early stopping method.

[0037] II. Speech abnormal dialogue recognition (DeBERTa-V3 word recognition model); Speech to text conversion: Use advanced speech recognition technology, such as deep learning-based end-to-end speech recognition model (DeepSpeech, Transformer speech recognition model, etc.), to convert speech data into text sequences. Post-process the speech recognition results to correct recognition errors such as misspelling, sentence breakage, etc.

[0038] DeBERTa-V3 model application: DeBERTa-V3 is a pre-trained language model that can capture deep semantic and contextual information in text. The converted text sequence is input into the DeBERTa-V3 model, which includes Token embedding, segment embedding and position embedding of the text. Through the multi-layer Transformer encoder, the text is encoded to obtain the context representation of each word. In the output layer of the model, a classifier is used to classify the text to determine whether there is an abnormal dialogue, such as containing sensitive words, threatening language, false information, etc.

[0039] Model fine-tuning: Based on the pre-trained DeBERTa-V3 model, fine-tune it using broadcast television domain speech dialogue data. Collect pairs of speech text containing normal and abnormal dialogues to construct a fine-tuning dataset. During fine-tuning, adjust all parameters of the model to make it better adapt to the broadcast television speech abnormal dialogue recognition task.

[0040] The identification model optimization module is configured to optimize hyperparameters in the CNN-DeBERTa data security identification model by using an improved IPSO particle swarm algorithm, and add a Transformer encoder in the security identification model to obtain a target CNN-DeBERTa data security identification model. Specifically, the improved sub-module is configured to introduce a dynamic inertia weight based on the PSO particle swarm algorithm, and the inertia weight is linearly decreased from 0.9 to 0.4 as the number of iterations increases, and a Gaussian mutation operator is used to perform mutation operation on the position of the particle to obtain an improved PSO particle swarm algorithm. The determination sub-module is configured to determine the hyperparameters to be optimized, including the learning rate of the CNN convolutional neural network, the decay rate of the BatchNormalization, the number of neurons of the fully connected layer, the fine-tuning learning rate of the DeBERTa-V3 model, the number of attention heads, and the hidden layer dimension. The optimization sub-module is configured to represent each particle as a combination of hyperparameters, and the position vector of the particle corresponds to the value of the hyperparameter. In each iteration, the fitness value of each particle is calculated, the speed and position of the particle are updated, and the optimal combination of hyperparameters is obtained through iteration.

[0041] Specifically; I. Hyperparameter optimization (improved IPOS particle swarm algorithm); IPOS algorithm improvement: The traditional particle swarm algorithm (PSO) is prone to fall into local optimum in the optimization process. To improve the optimization effect, the IPOS algorithm is improved. A dynamic inertia weight is introduced, and the inertia weight is linearly decreased from 0.9 to 0.4 as the number of iterations increases, to balance the global search and local development capabilities of the algorithm. At the same time, a Gaussian mutation operator is used to perform mutation operation on the position of the particle, increase the diversity of the population, and avoid premature convergence of the algorithm.

[0042] Hyperparameter definition: Determine the hyperparameters to be optimized, including the learning rate of the CNN convolutional neural network, the decay rate of the BatchNormalization, the number of neurons of the fully connected layer, the fine-tuning learning rate of the DeBERTa-V3 model, the number of attention heads, and the hidden layer dimension. Optimization process: Each particle is represented as a combination of hyperparameters, and the position vector of the particle corresponds to the value of the hyperparameter. In each iteration, the fitness value of each particle is calculated (based on the accuracy or loss value of the model on the validation set), and the speed and position of the particle are updated. Through continuous iteration, the optimal combination of hyperparameters is found, and the identification performance of the model is optimized.

[0043] II. Add a Transformer encoder to CNN; In the high-level feature extraction stage of the CNN convolutional neural network, a Transformer encoder is added. Specifically, after the last pooling layer, the feature map is flattened into a sequence as the input of the Transformer encoder. The Transformer encoder is composed of multiple self-attention layers and feedforward neural network layers, which can capture long-range dependencies between image features. Through the self-attention mechanism, the model can focus on important areas in the image, improving the ability to recognize complex image abnormal information. The connection mode is: the output feature map of the last pooling layer of the CNN → flatten into a sequence → input into the Transformer encoder → output context-related feature vector → input into the fully connected layer for classification.

[0044] The broadcast safety evaluation module is used to input the initial multi-source broadcast television data into the target CNN-DeBERTa data security identification model for identification to obtain a broadcast television safety evaluation index. Specifically, the input sub-module is used to input the initial multi-source broadcast television data into the target CNN-DeBERTa data security identification model to obtain image anomaly probability, voice anomaly probability, and content anomaly probability, respectively. The calculation sub-module is used to calculate network security indicators and user behavior security indicators based on network layer data and behavior data, and to calculate the broadcast television safety evaluation index using weighted average calculation.

[0045] Specifically, The preprocessed initial multi-source broadcast television data is input into the target CNN-DeBERTa data security identification model to obtain image anomaly probability, voice anomaly probability, and other content anomaly probability (text anomaly probability), respectively. At the same time, network security indicators (network attack frequency, data leakage risk, etc.) and user behavior security indicators (abnormal login frequency, illegal data access frequency, etc.) are calculated based on network layer data and behavior data. The broadcast television safety evaluation index is calculated using the weighted average method, The safety monitoring management module is used to perform state evolution prediction based on the broadcast television safety evaluation index using the HMM hidden Markov model, to calculate the risk evolution probability value in real time, and to trigger the broadcast television intervention strategy when the risk evolution probability value is greater than a threshold value.

[0046] Specifically, the definition sub-module is used to define the state set of the HMM, with the safe state representing normal operation of the broadcast television system, the warning state representing certain security risks in the system, and the dangerous state representing high security risks in the system. The setting sub-module is used to set the broadcast television safety evaluation index as the observation value of the HMM, and the observation value sequence is composed of the safety evaluation index output by the model in real time. The evaluation submodule is configured to estimate a state transition probability matrix and an observation probability matrix through the HMM model, calculate a probability of evolution to a dangerous state from a current state according to a current observation value sequence and model parameters, and obtain a risk evolution probability value.

[0047] Specifically; I. HMM Hidden Markov Model Construction State Definition: Define the state set of the HMM, including safe state, warning state and dangerous state. The safe state means that the broadcast television system is running normally and the risk is low; the warning state means that the system has certain security risks and needs close attention; the dangerous state means that the system faces high security risks and may have a security accident.

[0048] Observation Value Definition: The broadcast television safety evaluation index is taken as the observation value of the HMM, and the observation value sequence is composed of the safety evaluation index output by the model in real time.

[0049] Transition Probability Matrix Construction: The HMM model is trained through historical data to estimate the state transition probability matrix and the observation probability matrix. The state transition probability matrix represents the probability of transition from the current state to the next state, and the observation probability matrix represents the probability of generating a specific observation value in a certain state.

[0050] II. State Evolution Prediction and Risk Calculation The HMM model is used to predict the evolution of the broadcast television safety state, and the probability distribution of each state in the future is calculated according to the current observation value sequence and model parameters. The risk evolution probability value, i.e. the probability of evolution from the current state to the dangerous state, is calculated in real time. When the risk evolution probability value is greater than a preset threshold, the broadcast television intervention strategy is triggered.

[0051] III. Intervention Strategy Data Blocking: For detected abnormal data transmission, such as image, voice or text data containing illegal content, immediately block the data transmission channel to prevent the spread of abnormal data.

[0052] Device Isolation: If a network device or server is found to have security vulnerabilities or be attacked, it is isolated from other devices to avoid the expansion of security incidents.

[0053] Content Review: Manually review suspected abnormal program content to confirm whether there are security problems and timely delete or modify illegal content.

[0054] System Monitoring Upgrade: Strengthen real-time monitoring of the broadcast television system, increase monitoring frequency and monitoring indicators, and upgrade security protection systems such as firewall, intrusion detection system, etc. to improve the security of the system.

[0055] User notification and guidance: for the user's abnormal behavior, such as frequent illegal access attempts, send a security notification to the user, remind the user to pay attention to account security, and guide the user to perform legal operations.

[0056] The beneficial effects are that, by obtaining network layer data, content layer data and behavior layer data in the broadcast television tower, multi-source broadcast television data is obtained, the multi-source broadcast television data is preprocessed to obtain initial multi-source broadcast television data; CNN convolutional neural network is used to identify abnormal information in image information, DeBERTa-V3 word recognition model is used to identify abnormal text information in voice information, and a CNN-DeBERTa data security identification model is established; the improved IPSO particle swarm algorithm is used to optimize the hyperparameters in the CNN-DeBERTa data security identification model, and a Transformer encoder is added in the security identification model to obtain a target CNN-DeBERTa data security identification model; the initial multi-source broadcast television data is input into the target CNN-DeBERTa data security identification model for identification to obtain a broadcast television security evaluation index; the HMM hidden Markov model is used to perform state evolution prediction according to the broadcast television security evaluation index, and a risk evolution probability value is calculated in real time, and a broadcast television intervention strategy is triggered when the risk evolution probability value is greater than a threshold value. 1. It can capture security risks from multiple dimensions, greatly improving the comprehensiveness and accuracy of data processing. 2. Improve the recognition accuracy and generalization ability of the model for image abnormal information, voice abnormal dialogue and text illegal content, effectively avoid missing and misjudgment of security vulnerabilities. 3. It can master the trend of security risk changes in real time, and predict potential security threats in advance. Compared with the traditional static evaluation method, it realizes the dynamic monitoring and active defense of security risks. 4. Automatically trigger targeted broadcast television intervention strategies, including data blocking, device isolation, content review and other measures. This intelligent intervention mechanism can quickly respond to security incidents, effectively reduce the probability of security incidents, reduce the loss caused by security incidents, and ensure the safe and stable operation of the broadcast television system and the security of user information.

[0057] Please refer to Figure 2 In a DMTO device catalyst balance monitoring system based on fast calculation, the broadcast data acquisition module comprises: The deletion sub-module is used for deleting repeated data packets and error check data in the network layer data in the multi-source data; deleting damaged image data, audio data that cannot be decoded and text information with garbled codes in the content layer data; and deleting repeated operation records and abnormal login logs in the behavior data; The correction sub-module is used for filling the missing data by using the data interpolation method, identifying and correcting the abnormal values, and obtaining corrected broadcast television data. The uniform sub-module is configured to convert data of different sources and different formats in the modified broadcast television data into a uniform format to obtain uniform broadcast television data. The normalization sub-module is configured to normalize the uniform broadcast television data by a Z-score normalization method to obtain initial multi-source broadcast television data.

[0058] Please refer to Figure 3 In a DMTO device catalyst balance monitoring system based on rapid accounting, the identification model optimization module comprises: The improved sub-module is configured to introduce a dynamic inertia weight based on a PSO particle swarm algorithm, and the inertia weight is linearly decreased from 0.9 to 0.4 with the increase of the number of iterations, and a Gaussian mutation operator is used to perform mutation operation on the position of the particle to obtain an improved PSO particle swarm algorithm. The determination sub-module is configured to determine hyperparameters to be optimized, including a learning rate of a CNN convolutional neural network, an attenuation rate of BatchNormalization, a number of neurons of a fully connected layer, a fine-tuning learning rate of a DeBERTa-V3 model, a number of attention heads, and a hidden layer dimension. The optimization sub-module is configured to represent each particle as a combination of hyperparameters, a position vector of the particle corresponds to a value of the hyperparameter, calculate a fitness value of each particle in each iteration, update a speed and a position of the particle, and obtain an optimal combination of hyperparameters through iteration.

[0059] The above shows and describes the basic principles, main features and advantages of the present application. It should be understood by those skilled in the art that the present application is not limited to the above embodiments, and the above embodiments and descriptions in the specification are only preferred examples of the present application and are not intended to limit the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and all changes and improvements fall within the scope of the present application. The scope of protection of the present application is defined by the appended claims and their equivalents.

Claims

1. A broadcast television data security monitoring management system, characterized by, The broadcast television data security monitoring management system comprises: A broadcast data acquisition module, configured to acquire network layer data, content layer data and behavior layer data in a broadcast television tower, obtain multi-source broadcast television data, preprocess the multi-source broadcast television data, and obtain initial multi-source broadcast television data; An identification model establishment module, configured to identify abnormal information in image information based on a CNN convolutional neural network, identify abnormal text information in voice information by using a DeBERTa-V3 word recognition model, and establish a CNN-DeBERTa data security identification model; An identification model optimization module, configured to optimize hyperparameters in the CNN-DeBERTa data security identification model by using an improved IPSO particle swarm algorithm, add a Transformer encoder in the security identification model, and obtain a target CNN-DeBERTa data security identification model; A broadcast security evaluation module, configured to input the initial multi-source broadcast television data into the target CNN-DeBERTa data security identification model for identification, and obtain a broadcast television security evaluation index; A security monitoring management module, configured to perform state evolution prediction according to the broadcast television security evaluation index by using an HMM hidden Markov model, calculate a risk evolution probability value in real time, and trigger a broadcast television intervention strategy when the risk evolution probability value is greater than a threshold value.

2. The broadcast television data security monitoring management system of claim 1, wherein, The broadcast data acquisition module comprises the following submodules: A deletion submodule, configured to delete duplicate data packets and error check data in the network layer data of the multi-source data; delete damaged image data, audio data that cannot be decoded, and text information with garbled codes in the content layer data, and delete repeated operation records and abnormal login logs in the behavior data; A correction submodule, configured to fill in missing data by using a data interpolation method, identify and correct abnormal values, and obtain corrected broadcast television data; A unification submodule, configured to convert data of different sources and different formats in the corrected broadcast television data into a unified format, and obtain unified broadcast television data; A normalization submodule, configured to normalize the unified broadcast television data by using a Z-score normalization method, and obtain the initial multi-source broadcast television data.

3. The broadcast television data security monitoring management system of claim 1, wherein, The identification model establishment module comprises the following submodules: An input layer submodule, configured to receive preprocessed image data based on an input layer of a CNN convolutional neural network, and convert an image into a three-dimensional tensor; A convolution layer submodule, configured to set a plurality of convolution layers, and extract image features by using convolution kernels of different numbers and sizes; A pooling layer submodule, configured to use a maximum pooling layer after each convolution layer, set a pooling kernel size of 2x2 and a step size of 2, and downsample a feature map; A fully connected layer submodule, configured to flatten the feature map into a one-dimensional vector after a plurality of convolution and pooling layers, and input the one-dimensional vector to a fully connected layer.

4. The broadcast television data security monitoring management system of claim 1, wherein, The identification model establishment module comprises the following submodules: A conversion submodule, configured to convert voice data into a text sequence based on a DeepSpeech end-to-end speech recognition model, post-process a speech recognition result, and correct errors in wrong characters and sentence breaks; The encoding submodule is configured to input the converted text sequence into a DeBERTa-V3 model, input text Token embedding, segment embedding and position embedding, encode the text through a multi-layer Transformer encoder, and obtain a context representation of each word; The classification submodule is configured to classify the context representation using a classifier to determine whether there is abnormal text information.

5. The broadcast television data security monitoring management system of claim 1, wherein, The recognition model optimization module includes the following submodules: The improvement submodule is configured to introduce a dynamic inertia weight based on a PSO particle swarm algorithm, linearly decrease the inertia weight from 0.9 to 0.4 as the number of iterations increases, and use a Gaussian mutation operator to mutate the position of the particle to obtain an improved PSO particle swarm algorithm; The determination submodule is configured to determine the hyperparameters that need to be optimized, including the learning rate of the CNN convolutional neural network, the decay rate of BatchNormalization, the number of neurons in the fully connected layer, the fine-tuning learning rate of the DeBERTa-V3 model, the number of attention heads, and the hidden layer dimension; The optimization submodule is configured to represent each particle as a combination of hyperparameters, the position vector of the particle corresponds to the value of the hyperparameter, calculate the fitness value of each particle in each iteration, update the speed and position of the particle, and obtain the optimal combination of hyperparameters through iteration.

6. The broadcast television data security monitoring management system of claim 1, wherein, The broadcast safety evaluation module includes the following submodules: The input submodule is configured to input the initial multi-source broadcast television data into a target CNN-DeBERTa data security recognition model to obtain image anomaly probability, voice anomaly probability and content anomaly probability respectively; The calculation submodule is configured to calculate network security indicators and user behavior security indicators based on network layer data and behavior data, and calculate a broadcast television safety evaluation index using weighted average.

7. The broadcast television data security monitoring management system of claim 1, wherein, The safety monitoring management module includes the following submodules: The definition submodule is configured to define a state set of the HMM, wherein a safe state represents that the broadcast television system is running normally, a warning state represents that the system has certain security risks, and a dangerous state represents that the system faces high security risks; The setting submodule is configured to use the broadcast television safety evaluation index as the observation value of the HMM, and the observation value sequence is composed of the safety evaluation indexes output by the model in real time; The evaluation submodule is configured to estimate the state transition probability matrix and the observation probability matrix through the HMM model, calculate the probability of evolving from the current state to the dangerous state according to the current observation value sequence and the model parameters, and obtain a risk evolution probability value. The method includes the following steps:

8. A method of implementing the broadcast television data security monitoring management system as claimed in claim 1, characterized in that, Obtain network layer data, content layer data and behavior layer data in a broadcast television tower to obtain multi-source broadcast television data, preprocess the multi-source broadcast television data to obtain initial multi-source broadcast television data; Analyze the abnormal information in the image information based on the CNN convolutional neural network, identify the abnormal text information in the voice information using the DeBERTa-V3 word recognition model, and establish a CNN-DeBERTa data security recognition model; ​ The improved IPSO particle swarm algorithm is used to optimize the hyperparameters in the CNN-DeBERTa data security identification model, and a Transformer encoder is added to the security identification model to obtain a target CNN-DeBERTa data security identification model; The initial multi-source broadcast television data is input into the target CNN-DeBERTa data security identification model for identification to obtain a broadcast television security evaluation index; The HMM hidden Markov model is used to perform state evolution prediction according to the broadcast television security evaluation index, and a risk evolution probability value is calculated in real time; when the risk evolution probability value is greater than a threshold value, a broadcast television intervention strategy is triggered.

9. A method of implementing the broadcast television data security monitoring management system as claimed in claim 1, characterized in that, The method comprises the following steps: Repeated data packets and error check data in the network layer data in the multi-source data are deleted; damaged image data, audio data that cannot be decoded, and text information with garbled codes in the content layer data are deleted; repeated operation records and abnormal login logs in the behavior data are deleted; Data interpolation is used to fill in the missing data, and the abnormal values are identified and corrected to obtain corrected broadcast television data; Different source and format data in the corrected broadcast television data are converted into a unified format to obtain unified broadcast television data; The unified broadcast television data is normalized by a Z-score normalization method to obtain initial multi-source broadcast television data.

10. A method of implementing the broadcast television data security monitoring management system as claimed in claim 1, characterized in that, The method comprises the following steps: The speech data is converted into a text sequence based on a DeepSpeech end-to-end speech recognition model, and the speech recognition result is post-processed to correct errors and sentence breaks; The converted text sequence is input into a DeBERTa-V3 model, and text Token embedding, segment embedding and position embedding are input; the text is encoded by a multi-layer Transformer encoder to obtain a context representation of each word; The context representation is classified by a classifier to determine whether there is abnormal text information.

Images