Intelligent networked terminal multi-attack intrusion detection method based on OTA signature update

Abstract

The application discloses an intelligent network connection terminal multi-attack intrusion detection method based on OTA signature update, first injects multi-type combined attacks into an ICTs platform, then performs signature update based on OTA upgrade, next designs a CNN IDS based on automatic coding design, and finally realizes classification of attack modes after training. The application improves attack detection success rate and reduces calculation cost.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of Internet of Things, and particularly relates to a multi-attack intrusion detection method for intelligent networked terminals. BACKGROUND

[0002] With the rapid development of wireless communication and digital terminal technology, intelligent networked terminals (ICTs) have become an important part of future human life. Modern ICTs have strong capabilities in complex environment perception and intelligent control, mainly through the cooperation of various advanced chips and sensors. However, with the development of information communication technology and the increasing information exchange and open interfaces between terminals and the external world, more and more attack surfaces appear in the internal network of ICTs and the interactive network between terminals. These malicious attacks are likely to cause information leakage of terminal owners and terminal out-of-control security threats. Therefore, it is necessary to implement a multi-type attack intrusion detection system (IDS) in ICTs.

[0003] Through the summary of previous related researches, it is found that the current related researches on intrusion detection of ICTs mainly focus on identity authentication, machine learning (ML), deep learning (DL) and other directions. However, most of the current related researches focus on the IDS of the internal network of the terminal, and most of the IDSs can only detect a single type of attack. Therefore, the intrusion detection of multi-type combined attacks on the internal network of the terminal and the interactive network between the terminal and the external world is very important for improving the security of ICTs. However, the multi-attack intrusion detection system needs to consider the signature update between the internal network of the terminal and the terminal and the cloud, which will not only cause huge computational cost, but also cause security problems.

[0004] Therefore, how to implement multi-type combined attack intrusion detection in ICTs while guaranteeing the real-time performance of internal information interaction of the terminal and the accuracy of attack detection is a difficult problem faced by the current ICTs security. SUMMARY

[0005] In order to overcome the shortcomings of the prior art, the application provides a multi-attack intrusion detection method for intelligent networked terminals based on OTA signature update, which first injects multi-type combined attacks into the ICTs platform, then updates the signature based on OTA upgrade, next designs a CNN IDS based on automatic coding design, and finally realizes the classification of attack modes after training. The application improves the attack detection success rate and reduces the computational cost.

[0006] The technical scheme adopted by the present application to solve its technical problems comprises the following steps:

[0007] Step 1: Multi-type combined attack injection simulation ICTs platform;

[0008] Inject network attacks into the simulation ICTs platform alone or in different combinations;

[0009] Step 2: Signature update based on OTA upgrade;

[0010] After the attack injection in step 1 is completed, collect attack data through packet sniffing and data preprocessing;

[0011] Filter the collected attack data traffic based on signature detection to match the data traffic stored in the signature table, and the detection result is divided into the following three cases:

[0012] Case 1: If no attack is detected, the detection engine will return the information directly to the management log;

[0013] Case 2: If a single type of attack is detected, a new attack log is sent directly to the cloud-based management platform;

[0014] Case 3: If a combined attack is detected, the detection engine will use a CNN classifier based on automatic encoding to classify the traffic;

[0015] The classified attack log is sent to the cloud-based security management system, and for new attack signatures, the security management system sends the new attack signatures to all OTA networked intelligent terminals;

[0016] Each intelligent terminal is regarded as a packet inspector, and the packet inspector is used to find attack signatures and update the cloud signature database, which is then extended to the following intrusion detection architecture;

[0017] Step 3: CNN IDS design based on automatic encoding;

[0018] Step 3-1: Design an automatic encoder;

[0019] The automatic encoder is an unsupervised neural network composed of an encoder and a decoder; first, the encoder compresses the input data into a low-dimensional space, and then the decoder restores it to the original data; the encoder and the decoder are connected to form a neural network, and the network parameters are optimized through a backpropagation algorithm;

[0020] The encoder is represented as The decoder is represented as The network structure of the automatic encoder is represented as:

[0021]

[0022] wherein X is the input ID data of the autoencoder, and Z is the intermediate feature data extracted and saved to a local file after the model training is completed;

[0023] The autoencoder inputs the terminal identification ID data, passes through multiple hidden neurons, and outputs feature data after compression processing, and then converts the feature data into a p*p data grid, uses q*q p*p data grids to form a p 2 *q 2 grid data as the input of the subsequent CNN model;

[0024] During the training process, the output X Pred and the input X True are used for error calculation, and the mean square error is used as the error function, and the training result is calculated by formula (2); after the training is completed, the intermediate feature data Z, i.e. the p 2 *q 2 grid data, is extracted;

[0025]

[0026] Step 3-2: CNN model design:

[0027] The CNN model is composed of an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer in sequence; the input is the p 2 *q 2 grid data obtained in step 3-1; in the convolutional layer, c convolutional kernels are used for operation, and the output size of the convolutional layer is calculated by formula (3):

[0028]

[0029] wherein O is the output size, W is the input data size, K is the kernel size, the padding value P is used to ensure that the data size after convolution is the same as the input data size, and S is the step size;

[0030] The input of the convolutional layer depends on the output of the last layer, and its output determines the input of the next layer. When there are multiple convolutional layers, the output of each convolutional layer can be calculated by formula (4).

[0031] a i =σ(z)=σ(a i-1 *W i +b i ) (4)

[0032] wherein a i is the output of a convolutional layer, σ is the activation function, a i-1 is the output of the last layer, W i is the connection weight, and bi is the bias of the layer. Here, the bias b i The initial value of b is set to 0.1.

[0033] The dimension of the pooling layer F is d*d, which is used to take the average value of the data in each non-repeated d*d area, and the dimension of the output data after pooling is reduced to half of the dimension of the input data;

[0034] The following is a full connection layer and an output layer; the output layer adopts a softmax classifier for classification, as shown in formula (5):

[0035]

[0036] In the formula, n is the number of classifications, o i is the output of the output layer neuron i, P i is the probability of the output layer neuron i; the maximum output neuron is taken as the type of the output of the CNN model;

[0037] In the optimization of the CNN model, a cross-entropy loss function is selected as the loss function, and an adaptive moment estimation optimization method is used to minimize the loss function.

[0038] Preferably, the network attack includes denial of service attack, fuzzy attack and fraud attack.

[0039] Preferably, the cloud-based security management system includes an attack signature generator, a database, complete information about attack signatures, and a signature list of each connection terminal.

[0040] Preferably, the autoencoder uses an Adam optimizer during training, and the learning rate is set to 0.001 for parameter optimization; a small batch gradient descent method is used, and the batch size is 200 for error calculation, and the model is trained for 5 epochs.

[0041] Preferably, 128 neurons are used in the full connection layer, the bias value b is set to the initial value 0.1, a dropout layer is added after the full connection layer, and the failure probability in the dropout layer is set to 0.5 during the training process, and the failure probability in the dropout layer is set to 0 during the testing process.

[0042] Preferably, p=3, q=8, c=20, and d=2.

[0043] The beneficial effects of the present application are as follows:

[0044] The present application overcomes the shortcomings of the previous ICTs IDS, which only considers the detection of internal network attacks of terminals and can only detect one type of attack at a time. In the simulation of ICTs, multiple types of combined attacks are realized, the success rate of attack detection is improved, and the calculation cost is reduced. BRIEF DESCRIPTION OF DRAWINGS

[0045] Figure 1 is the overall flowchart of the present application.

[0046] Figure 2 is the confusion matrix diagram of the embodiment of the present application under 11-bit CAN ID input. (a) the result of one attack type existing; (b) the result of two attack types existing; (c) the result of three attack types existing; (d) the result of four attack types existing.

[0047] Figure 3 is the confusion matrix diagram of the embodiment of the present application under 29-bit CAN ID input. (a) the result of one attack type existing; (b) the result of two attack types existing; (c) the result of three attack types existing; (d) the result of four attack types existing.

[0048] Figure 4 is the result of the embodiment of the present application under different intrusion combinations. (a) the result generated by 11-bit CAN ID input; (b) the result generated by 29-bit CAN ID input.

[0049] Figure 5 is the running time and overall detection accuracy of the embodiment of the present application under different combination attacks with or without automatic encoder dimension reduction processing. (a) the test running time result; (b) the overall detection accuracy result. DETAILED DESCRIPTION

[0050] The present application is further illustrated below in conjunction with the drawings and embodiments.

[0051] The purpose of the present application is to provide a new IDS framework based on OTA (Over the Air) and automatic coding CNN (Convolutional Neural Network) for realizing multiple types of attacks in ICTs, OTA is used to solve the signature update problem of ICTs interacting with the cloud, and the CNN based on automatic coding is used to solve the detection accuracy problem of multiple type combination attacks.

[0052] To achieve the above purpose, the technical solution adopted by the present application is:

[0053] Step 1: multiple type combination attack injection;

[0054] Step 2: OTA-based signature update;

[0055] Step 3: CNN IDS design based on automatic coding.

[0056] (1) The specific implementation process of step 1 is as follows:

[0057] Taking intelligent connected vehicles as an example, Controller Area Network (CAN) is the primary communication method in modern vehicles. However, due to the lack of corresponding security mechanisms in the CAN bus network, hackers may be able to attack it. Common attack types on in-vehicle networks include:

[0058] (1) Denial of Service (DoS) attack. Because CAN bus messages have priority, in a DoS attack, the attacker will continuously send the highest priority message to occupy CAN communication resources, causing the system to be unable to send any normal messages.

[0059] (2) Fuzzy attack. Attackers enter the CAN bus by simulating legitimate messages and randomly inject messages into the CAN bus, which may cause problems such as steering wheel vibration, intermittent on / off of turn signals, and automatic gear shifting.

[0060] (3) Spoofing attack, including Gear / RPM. Attackers choose to inject messages with specific CAN IDs into the CAN bus to cause vehicle malfunctions. Compared to fuzzy attacks, spoofing attacks select CAN IDs that are normally transmitted on the CAN bus to attack the network, while fuzzy attacks can create any simulated ID to launch the attack.

[0061] Different combinations of the above attacks were injected into the simulated ICTs platform.

[0062] (2) The specific implementation process of step 2 is as follows:

[0063] After the attack is injected, attack data is collected through packet sniffing and data preprocessing. First, traffic is filtered using signature-based detection to match patterns stored in a signature table. If certain patterns are matched, alerts are triggered to the driver and security management system administrator. Possible detection results are as follows:

[0064] Scenario 1: If no attack is detected, the detection engine will return the information directly to the management log.

[0065] Scenario 2: If a single type of attack is detected, the new attack log will be sent directly to the cloud-based management platform;

[0066] Scenario 3: If a combined attack is detected, the detection engine will use an autoencoder-based CNN classifier to classify the traffic (see the implementation process of step 3).

[0067] Then, the classified attack logs are sent to a cloud-based security management system, which basically contains an attack signature generator, a database, complete information about attack signatures, a signature list for each connected terminal, etc. For new attack signatures, the system sends them to all OTA connected terminals. Note that the general OTA concept refers to a method of distributing new software updates, configuration settings, or encryption keys from a central unit to all connected devices or users. These devices or users can accept or reject these updates.

[0068] Currently, most smart terminal manufacturers have used OTA in vulnerability fixes and software updates, but there is no research on using OTA in ICTs IDS. First, each smart terminal is regarded as a package inspector, and the package inspector is used to find attack signatures and update the cloud signature database, and then it is extended to the new connected terminal framework to further improve the success rate and real-time performance of attack detection.

[0069] The specific implementation process of step 3 is as follows:

[0070] Design of autoencoder:

[0071] Due to the limitations of computing and storage capacity of terminal devices, reducing the number of model parameters and computational complexity is an effective means to reduce data dimensionality. Autoencoder is an unsupervised neural network composed of an encoder and a decoder, and is a common data dimensionality reduction method. First, the encoder compresses the input data into a low-dimensional space, and then the decoder restores it to the original data. The encoder and decoder are connected to form a neural network, and the network parameters are optimized through the backpropagation algorithm. Therefore, the present application uses an autoencoder network to reduce the dimensionality of data, which effectively reduces the number of model parameters and computational complexity.

[0072] The encoder is represented as The decoder is represented as The structure of the network can be represented as:

[0073]

[0074] The detailed parameter settings of the autoencoder are shown in Table 1. The autoencoder input terminal identification ID (which means CAN ID in smart connected cars) data (29 bits), 17 hidden neurons, and output feature data (9 bits) after compression processing. Then convert the 9-bit data into a 3*3 grid, and use an 8*8 data grid to form a 24*24 element pattern as the input of the CNN model. During the training process, the input X Pred and X TrueFor error calculation, mean square error is used as error function, and the training result can be calculated by formula (2). The model uses Adam optimizer, and the learning rate is set to 0.001 for parameter optimization. In addition, the small batch gradient descent method is used, and the batch size is 200 for error calculation, and the model is trained for 5 epochs. After the model training is completed, the intermediate feature data Z is extracted and saved to the local file as the input of the CNN.

[0075]

[0076] Table 1: Autoencoder parameter settings

[0077]

[0078]

[0079] CNN model design:

[0080] The CNN model used by the application is composed of an input layer, a convolutional layer, a pooling layer, a fully connected layer and an output layer. The input is the grid data obtained in the previous step. In the convolutional layer, 20 convolutional kernels are used for operation. The output size of the convolutional layer can be obtained by formula (3).

[0081]

[0082] Where O is the output size, W is the data size, K is the kernel size, the padding value P is used to ensure that the data size after convolution is the same as the input data size, and S is the step size.

[0083] The input of the convolutional layer depends on the output of the last layer, and its output determines the input of the next layer. When there are multiple convolutional layers, the output of each convolutional layer can be calculated by formula (4).

[0084] a i =σ(z)=σ(a i-1 *W i +b i ) (4)

[0085] Where a i is the output of a convolutional layer, sigma is the activation function, a i-1 is the output of the last layer, W i is the connection weight, and b i is the bias of the layer. Here, the initial value of the bias b i is set to 0.1.

[0086] The dimension of the pooling window F is 2*2, and the average pooling is used to take the average of the data in each non-repeated 2*2 area. Therefore, the dimension of the output data after pooling is reduced to half of the original dimension.

[0087] Then comes the fully connected layer. In the fully connected layer, 128 neurons are used, which are connected to the neurons of the flat final pooling layer. The bias value b in the 128 neurons is also set to the initial value 0.1. In order to prevent overfitting in the model training process, a dropout layer is used, which is achieved by making the activation function in certain neurons fail with a certain probability. In the training process, the failure probability in the dropout layer is set to 0.5, but in the testing process, it is equal to 0, that is, there is no dropout at all in the testing stage.

[0088] Finally comes the output layer. Because there are five types of data in this embodiment, including normal data, DOS, Fuzzy, Gear and RPM attack data, the goal of this embodiment is to accurately classify them, so the output layer also contains five neurons. Equation (5) is a softmax classifier. This layer is used to classify the five categories because it is more suitable for multiple classification tasks than other activation functions.

[0089]

[0090] In the formula, n is the number of classifications (n = 5), o i is the output of neuron o i , and P i is the probability of neuron i. A hot encoding is used, with the maximum output neuron as the type of model output. In the optimization of the CNN, a cross-entropy loss function is selected, and an adaptive moment estimation optimization method is used to minimize the function. By constantly adjusting the learning rate, it is found that the best learning rate of the optimizer is 1e-4.

[0091] In summary, the process of processing data based on the CNN model of automatic coding is as follows:

[0092] Step 1: Convert CAN ID to two-dimensional data grid (including x-axis, y-axis);

[0093] Step 2: Put multiple two-dimensional data grids together to form a two-dimensional pattern;

[0094] Step 3: Grid data padding;

[0095] Step 4: Further set the two-dimensional data grid to form a larger two-dimensional pattern structure, where the numbers in the grid represent binary bits;

[0096] Step 5: For the two-dimensional pattern formed, maintain the time relationship between the sequence data features by arranging the data grid in sequence from the continuous data sequence, so as to realize the conversion of the original data grid to a larger two-dimensional pattern structure;

[0097] Step 6: The converted pattern data structure will be used as the training and test sets for the CNN to preserve the temporal characteristics of the original data.

[0098] For example: for a CAN ID of 11 bits, a 4x4 data grid is formed, where the 11-bit binary number is first filled in the 4 edges of the grid, and the remaining 5 elements are filled with 0. The reason for filling the data in this way is that the CNN can more easily identify the pattern around the grid. Then, the two-dimensional data grid is further set to form a larger two-dimensional pattern structure of 8x8 data patterns, where the numbers in the grid represent binary bits. In this way, the original 4x4 data grid is converted into a pattern of 32x32 elements. For a CAN ID of 29 bits, each piece of data is first converted into a 6*6 data grid. At present, since only the first 11 bits of the total 29 bits of data are not zero, they are placed near the center of the 6*6 grid to make full use of the ability of the CNN to identify two-dimensional pattern patterns. Since there are 36 elements in the 6*6 grid, 29 of which are filled by CAN ID, and the remaining 7 elements are filled with 0. Similarly, after the data is gridded, it is spliced together to form a pattern structure to maintain the temporal characteristics between each piece of data. The present application uses an 8*8 data grid to form a 48*48 two-dimensional pattern structure as the input of the network training set and test set. DETAILED EMBODIMENT

[0100] The model of the embodiment was trained and tested using vehicle attack data collected by the Korea Internet & Security Center. Four of the data sets were selected: DOS, Fuzzy, Spoofing Gear, and Spoofing RPM attacks.

[0101] To evaluate the effectiveness of the proposed model, a combination of the above four data sets was used as the training and test set to meet the requirements of multi-type attack detection, and the information of the four data sets is shown in Table 2. In the automatic encoder coding process, in order to ensure the temporal connectivity between the CAN ID sequences, the present application arranges the 3*3, 4*4 or 6*6 CAN ID in time sequence into a pattern. After all the CAN ID data is coded in the form of a pattern, in order not to be affected by the arrangement of the data, the model randomly selects a pattern during the training and testing process. The training set contains 75% of the randomly selected data, and the remaining 25% is used as the test set. Although the model is trained using a combination of all four data sets, during the testing phase, all possible combinations of different types of data sets are used for testing. Therefore, the test set contains one attack, one attack combination, one attack combination and The four attacks are combined, resulting in 15 different attack types. Each dataset and its index number is shown in Table 2. Through this type of training and testing, the present application will be able to detect all types of combined attacks in Table 3.

[0102] Table 2 Four attack data information

[0103]

[0104] Experimental results:

[0105] The performance of the proposed framework is tested on the 11th and 29th CAN ID, respectively, and the results are shown in the confusion matrices in Figure 2 and Figure 3 . In each figure, from left to right, there are four attack types (DOS, Fuzzy, Gear, RPM), three attack types, two attack types, and one attack type. In these figures, the number in each cell represents the number of predicted horizontal types (predicted types) to vertical types (actual types). For example, the number in the first column of the confusion matrix represents the number of different types of RPM attacks predicted by the model, and the number in the diagonal line represents the number of correct classifications by the model.

[0106] From Figure 2 and Figure 3 , it can be seen intuitively that in most cases, the trained model can correctly detect different attack types. However, when the dataset contains multiple attack types, one attack type may sometimes be misclassified as other types, and the model is most likely to misclassify the attack type as a normal attack type. On the other hand, when the dataset does not contain any attacks, the model may misclassify it as an attack. These misclassifications will slightly reduce the prediction accuracy of the model.

[0107] Performance analysis:

[0108] To more accurately evaluate the prediction performance of the model, similar evaluation indicators to the binary classification model, i.e. Accuracy, Recall, Precision and F1_score, are also obtained according to the confusion matrix of each class. In a multi-classification model, a certain class (such as Normal) is defined as a positive class, and the remaining classes (DOS, fuzzy, Gear and RPM) are defined as negative (intrusion) classes, and different positive and negative classes are selected in turn. Then, the true positive (T P ), false positive (F P ), false negative (F N ) and true negative (T N) represents the correct identification of normal data. The relevant evaluation metrics are calculated according to the evaluation metric formulas given in the bisection method in equations (6)-(9), where D represents the attack data set tested in each experiment.

[0109]

[0110]

[0111]

[0112]

[0113] Based on the confusion matrix described above, the evaluation metrics of the model in the data set containing different attack combinations can be calculated using equations (6)-(9). Equations (6)-(9) can be used to calculate the evaluation metrics of the model in the data set containing different attack combinations. From Figure 4 It can be seen from Table 6 that the autoencoder performs well under different attack combinations, whether it is 11-bit or 29-bit CAN ID. When the data set contains fuzzy attacks, the performance will decrease. From data set 1 to 2, data set 3, data set 5 to 6, data set 9, data set 10 to 13, the proportion of fuzzy attacks in the data set gradually increases, and the performance of the model decreases in turn, which shows that the fuzzy attack is the most difficult to detect.

[0114] After the training of the autoencoder is completed, the 29-bit CAN ID is compressed into 9-bit features, and then a 3*3 data grid is created using 9-bit data, and such a grid of 8*8 is used for CNN training. The trained CNN is used to test the data, and the relevant evaluation results are obtained. In order to evaluate the performance of the autoencoder in the model detection ability, we compare the results of the 8*8 spliced 3*3 data grid without the autoencoder feature extraction. Figure 5 (a) and Figure 5 (b) respectively show the comparison of test running time and overall model accuracy by adopting different combination attack types given in Table 2.

[0115] As Figure 5 (a) shows that for the model without autoencoder, the test running time is greatly affected by the number of combined attacks. Data set 15 contains four types of attacks, the largest number of samples, and the longest running time. However, data sets 1, 2, 3, and 4 contain only one type of attack, the smallest number of samples, and the shortest running time. However, when the autoencoder is adopted, the model test running not only greatly reduces the time, but also is not affected by the number of samples. Because the number of connections used in the model is much smaller. With the reduction of model parameters, the computational complexity of the model is greatly reduced, the detection time is effectively shortened, and the real-time detection ability of the model to attacks is also improved. InFigure 5 In (b), the overall detection accuracy using the autoencoder also slightly decreases, although it remains at an acceptably high level. Therefore, the dimension reduction of the autoencoder can effectively reduce the computational complexity of the model without affecting the detection performance, which is very important for devices with limited computing power and storage resources.

[0116] Table 3 Combination index of different types of attacks

[0117]

Claims

1. An intelligent networked terminal multi-attack intrusion detection method based on OTA signature update, characterized in that, The method comprises the following steps: Step 1: Multi-type combined attack injection simulation ICTs platform Inject network attacks into the simulation ICTs platform alone or in different combinations; Step 2: Signature update based on OTA upgrade After the attack injection in step 1 is completed, collect attack data through packet sniffing and data preprocessing; Filter the collected attack data traffic based on signature-based detection to match the data traffic stored in the signature table, and the detection result is divided into the following three cases: Case 1: If no attack is detected, the detection engine will return the information directly to the management log; Case 2: If a single type of attack is detected, a new attack log is sent directly to the cloud-based management platform; Case 3: If a combined attack is detected, the detection engine will use an auto-encoder-based CNN classifier to classify the traffic; The classified attack log is sent to the cloud-based security management system, and for new attack signatures, the security management system sends the new attack signatures to all OTA networked intelligent terminals; Each intelligent terminal is regarded as a packet inspector, and the packet inspector is used to find attack signatures and update the cloud signature database, which is then extended to the subsequent intrusion detection architecture; Step 3: Auto-encoder-based CNN IDS design Step 3-1: Design an auto-encoder The auto-encoder is an unsupervised neural network composed of an encoder and a decoder; first, the encoder compresses the input data into a low-dimensional space, and then the decoder restores it to the original data; the encoder and the decoder are connected to form a neural network, and the network parameters are optimized through the backpropagation algorithm; The encoder is represented as The decoder is represented as The network structure of the autoencoder is represented as: （1） wherein, is input ID data of the autoencoder, is intermediate feature data extracted and saved to a local file after the model training is completed; The auto-encoder inputs terminal identification ID data, passes through multiple hidden neurons, and outputs feature data after compression processing, and then converts the feature data into a p*p data grid, uses q*q p*p data grids to form a p 2 *q 2 Grid data as input for the subsequent CNN model; In the training process, the output and the input are used for error calculation, and the mean square error is used as the error function, and the training result is calculated by formula (2); after the training is completed, the intermediate feature data , that is, the grid data of p 2 *q 2 is extracted; （2） Step 3-2: CNN model design The CNN model is sequentially composed of an input layer, a convolution layer, a pooling layer, a full connection layer, and an output layer; the input is the p 2 *q 2 grid data obtained in step 3-1; in the convolution layer, c convolution kernels are used for operation, and the output size of the convolution layer is calculated by formula (3): （3） wherein is an output size, is an input data size, is a kernel size, a padding value for ensuring that the data size after convolution is the same as the input data size, is a stride; The input of the convolutional layer depends on the output of the last layer, and its output determines the input of the next layer; when there are multiple convolutional layers, the output of each convolutional layer can be calculated by formula (4); （4） wherein is the output of a convolutional layer, is an activation function, is the output of the last layer, is a connection weight, is a bias of the layer; here, the bias is set to 0.1 initially; The dimension of the pooling layer F is d*d, which is used to take the average value of the data in each non-repeated d*d region, and the dimension of the output data after pooling is reduced to half of the input data dimension; The following is a fully connected layer and an output layer; the output layer uses a softmax classifier for classification, as shown in formula (5): （5） In the formula, is the number of categories, is the output of the output layer neuron , is the probability of the output layer neuron ; the maximum output neuron is taken as the type of the CNN model output; In the optimization of the CNN model, the cross-entropy loss function is selected as the loss function, and the adaptive moment estimation optimization method is used to minimize the loss function. 2.The intelligent network terminal multi-attack intrusion detection method based on OTA signature update of claim 1, wherein The network attacks include denial-of-service attacks, fuzzing attacks, and spoofing attacks. 3.The intelligent network terminal multi-attack intrusion detection method based on OTA signature update of claim 1, wherein, The cloud-based security management system includes an attack signature generator, a cloud signature database, complete information about attack signatures, and a signature list for each connected terminal.

4. The intelligent networked terminal multi-attack intrusion detection method based on OTA signature update according to claim 1, characterized in that, The auto-encoder uses the Adam optimizer during training, with a learning rate of 0.001 for parameter optimization; the small batch gradient descent method is used, with a batch size of 200 for error calculation, and the model is trained for 5 epochs.

5. The intelligent networked terminal multi-attack intrusion detection method based on OTA signature update according to claim 1, characterized in that, The full connection layer uses 128 neurons, and the bias value is set to 0.1 The dropout layer is added after the full connection layer, and the failure probability of the dropout layer is set to 0.5 during the training process, and the failure probability of the dropout layer is set to 0 during the test process.

6. The intelligent networked terminal multi-attack intrusion detection method based on OTA signature update according to claim 1, characterized in that, The p=3, q=8, c=20, d=2.

Images