Intrusion detection method and system

Through the collaborative work of edge nodes and cloud nodes, Bayesian deep learning model and sliding window adaptive adjustment technology are used to solve the problem of extended intrusion detection response time in cloud environments, achieving efficient and accurate intrusion detection, and reducing cloud Nodes communicate pressure and improve detection accuracy.

CN114448599BActive Publication Date: 2024-11-29GCI SCI & TECH +2

Patent Information

Application Number
CN202210151457.7
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-02-18
Publication Date
2024-11-29
Estimated Expiration
2042-02-18

AI Technical Summary

Technical Problem

In a cloud environment, intrusion detection requires processing massive data, resulting in an increase in response time. Traditional cloud computing architecture cannot meet the needs of low latency and efficient network access, and the computing power of edge nodes is weak, and there are geographically distributed and heterogeneous connections. It is difficult to achieve efficient intrusion detection due to entry problems.

Method used

Through edge nodes, the edge intrusion detection model is constructed through multiple rounds of training on the device feature data, and the model input and output are used as global training samples. The cloud node trains global model parameters based on Bayesian deep learning model, adjusts the sliding window size according to the detection error Adapt to adjust the data volume and realize accurate prediction of cloud node intrusion detection results.

Benefits of technology

It effectively reduces the communication transmission pressure of cloud nodes, improves the response service capabilities, and improves the accuracy of intrusion detection by updating global model parameters online with the data distribution of edge nodes.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114448599B_ABST
    Figure CN114448599B_ABST
Patent Text Reader

Abstract

The present invention discloses an intrusion detection method and system, wherein an edge node constructs an edge intrusion detection model based on the device feature data of the edge side itself, and a cloud node trains the global model parameters of the cloud node based on the model input and model output of the edge intrusion detection model of all edge nodes, wherein the global model is a model based on Bayesian deep learning; at the same time, the cloud node adaptively adjusts the amount of sampled data input of the global model based on the error accuracy predicted by the global model, and the present invention can not only reduce the pressure of communication transmission of the cloud node, but also improve the ability of the cloud node to respond to services. At the same time, the present invention updates the intrusion detection model every other detection cycle, and the parameter update of the global model will change with the change of the data distribution of the edge node, thereby improving the accuracy of intrusion detection.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of network security technology, and in particular to an intrusion detection method and system. Background Art

[0002] Intrusion detection technology is a technology designed and configured to ensure the security of computer systems. It can detect and report unauthorized or abnormal phenomena in the system in a timely manner. It is a technology used to detect violations of security policies in computer networks. The combination of software and hardware for intrusion detection is an intrusion detection system. Through intrusion detection technology, abnormal data on the edge side such as hosts and networks can be monitored. At present, intrusion detection based on cloud environments is very mature, but it takes time to transmit massive data in cloud environments and process data in cloud centers, which increases response time. Therefore, it is urgent to study a set of efficient algorithms to achieve the purpose of intrusion detection. Summary of the invention

[0003] The embodiments of the present invention provide an intrusion detection method and system, which can effectively reduce the pressure of cloud node communication transmission and improve the ability of cloud nodes to respond to services.

[0004] A first aspect of the present invention provides an intrusion detection method, comprising:

[0005] In the current detection cycle, for each edge node, the edge node performs multiple rounds of training on the pre-built edge intrusion detection model based on the collected device feature data, obtains the input and output of each round of edge intrusion detection model and uses it as a global training sample;

[0006] The cloud node trains the parameters of the pre-built global model based on the global training samples to obtain a global model with optimized parameters; wherein the global model is a detection model based on Bayesian deep learning;

[0007] The cloud node inputs the sampling data collected in the current detection cycle into the global model after parameter optimization to obtain the intrusion detection result of the current detection cycle; wherein, the data volume of the sampling data collected in the current detection cycle is determined according to the sliding window size of the current detection cycle sampling, and the sliding window size of the current detection cycle is obtained by adjusting the size of the sliding window of the previous detection cycle based on the error of the global model of the previous detection cycle.

[0008] A second aspect of the present invention provides an intrusion detection system, comprising a cloud node and a plurality of edge nodes; wherein:

[0009] The edge node is used to perform multiple rounds of training on the pre-built edge intrusion detection model based on the collected device feature data in the current detection cycle, obtain the input and output of each round of the edge intrusion detection model and use it as a global training sample;

[0010] The cloud node is used to train the parameters of the pre-built global model based on the global training sample to obtain the global model with optimized parameters;

[0011] The cloud node is also used to input the sampling data collected in the current detection cycle into the global model after parameter optimization to obtain the intrusion detection result of the current detection cycle; wherein, the data volume of the sampling data collected in the current detection cycle is determined according to the sliding window size of the current detection cycle sampling, and the sliding window size of the current detection cycle is obtained by adjusting the size of the sliding window of the previous detection cycle based on the error of the global model of the previous detection cycle.

[0012] Compared with the prior art, the present invention has the following beneficial effects:

[0013] The intrusion detection method provided by the present invention is that the edge node constructs an edge intrusion detection model based on the device feature data of the edge side itself, and the cloud node trains the global model parameters of the cloud node based on the model input and model output of the edge intrusion detection model of all edge nodes, wherein the global model is a model based on Bayesian deep learning; at the same time, the cloud node adaptively adjusts the amount of sampled data input to the global model based on the error accuracy predicted by the global model. The present invention can not only reduce the pressure of communication transmission of the cloud node, but also improve the ability of the cloud node to respond to services. At the same time, the present invention updates the intrusion detection model every other detection cycle, and the parameter update of the global model will change with the change of the data distribution of the edge node, thereby improving the accuracy of intrusion detection. Accordingly, the present invention also provides an intrusion detection system. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 It is a schematic diagram of the architecture of the intrusion detection method provided by an embodiment of the present invention;

[0015] Figure 2 It is a flowchart of an intrusion detection method provided by an embodiment of the present invention;

[0016] Figure 3 is a schematic diagram of parameter distribution estimation based on Bayesian deep learning provided by an embodiment of the present invention;

[0017] Figure 4 It is a schematic diagram of the parameter estimation process of the global model provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0018] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0019] Although the traditional cloud computing architecture can provide resource-centralized remote services, with the advent of the Internet of Things era, the amount of data services caused by massive terminal access has increased dramatically, and the traditional centralized remote service model cannot meet the low-latency, intensive computing network access and service requirements. Edge computing sinks cloud services to the massive terminal side, overcoming the shortcomings of cloud services in real-time content delivery, low service latency, high network transmission bandwidth pressure, and location awareness. Compared with traditional cloud computing, the server computing power of edge computing is relatively weak, and each edge node has problems such as geographically distributed deployment and heterogeneous access. Therefore, it is urgent to study a set of efficient, lightweight, and high-detection accuracy algorithms to achieve the purpose of real-time perception of intrusion detection and adaptive control of data granularity. The architecture of the intrusion detection method provided in this application can be referenced. Figure 1 The technical solution of the present application is described in detail below.

[0020] See also Figure 2 , which is a flow chart of an intrusion detection method provided by an embodiment of the present invention.

[0021] The intrusion detection method provided by the embodiment of the present invention includes steps S11 to S13:

[0022] S11, in the current detection cycle, for each edge node, the edge node performs multiple rounds of training on a pre-built edge intrusion detection model based on the collected device feature data, obtains the input and output of each round of the edge intrusion detection model and uses it as a global training sample.

[0023] In the embodiment of the present invention, the cloud node updates the intrusion detection once every detection cycle, and in each detection cycle, the edge node and the cloud node coordinate and divide the work. For the edge side, the edge node collects its own edge device feature data as the edge training sample of the edge intrusion detection model. The input and output of the edge intrusion detection model of all edge nodes in the nth round are used as the global training samples of the cloud node's nth round global model.

[0024] It can be understood that, for the i-th edge node, the input of its corresponding edge intrusion detection model is the edge-side device feature data, and the output is the corresponding intrusion behavior detection result.

[0025] Step S12: The cloud node trains the parameters of the pre-built global model based on the global training samples to obtain a global model with optimized parameters; wherein the global model is a detection model based on Bayesian deep learning.

[0026] In a specific implementation, the parameters of the global model are trained for the first round according to the global training samples of the first round of the global model obtained in step S11, the parameters of the global model are trained for the second round according to the global training samples of the second round of the global model obtained, ..., the parameters of the global model are trained for the nth round according to the global training samples of the nth round of the global model obtained, and after reaching a preset number of training rounds, a global model with optimized parameters is finally obtained.

[0027] See also Figure 3 , Figure 3 The parameter distribution estimation based on Bayesian deep learning provided by the embodiment of the present invention is shown. In the embodiment of the present invention, Bayesian deep learning is used to realize the estimation of parameter distribution, and a multi-round distributed edge intrusion detection model is used to construct the estimation of global model parameter distribution.

[0028] The training samples of all edge nodes in each round constitute the global training set of each round. The global training sets of multiple rounds include input x and output y: D = {(x1, y1), (x2, y2), ..., (x m ,y m )}.

[0029] In the specific implementation, multiple rounds of global training sets D are used to train the parameters of a Bayesian deep neural network, which is assumed to be the weight factor w. Then, given x and y, after multiple rounds of training, we get the distribution probability of w at each value:

[0030]

[0031] And p(y|x,w) and p(w) can be counted based on the edge intrusion detection model of the edge node. During multiple rounds of global model training, the value of the parameter w is 1 for some global models, 5 for some global models, and 2 for some global models, that is, it is distributed between 1-5, so this w may be 1, 2, or 5. Suppose 20 global models get w as 1, 50 global models get w as 2, and 30 models get w as 5. Then the distribution probability of w is that 20% of the values ​​are 1, 50% of the values ​​are 2, and 30% of the values ​​are 5. This distribution probability of w provides a data basis for the probability distribution of the global model. The parameter distribution result of a global model is obtained by edge computing the parameter distribution of the distributed model, and finally the distribution probability of w at each value is realized. The final global model can be referred to Figure 4 .

[0032] Of course, w may also be other parameters other than the weight factor, such as a bias factor, which is not limited in this application.

[0033] In an embodiment of the present invention, the distribution estimation of the learning parameters of the cloud node intrusion detection model is realized based on the Bayesian deep learning model, and the online update of the cloud node intrusion detection model is realized. The parameter update of the global model will change with the change of the data distribution of the edge node, which further reflects the nature of the online learning of the global model.

[0034] S13, the cloud node inputs the sampling data collected in the current detection cycle into the global model after parameter optimization to obtain the intrusion detection result of the current detection cycle; wherein, the data volume of the sampling data collected in the current detection cycle is determined according to the sliding window size of the current detection cycle sampling, and the sliding window size of the current detection cycle is obtained by adjusting the sliding window size of the previous detection cycle based on the error of the global model of the previous detection cycle.

[0035] In one implementation, the cloud node trains the parameters of the pre-built global model based on the global training sample to obtain a global model with optimized parameters, specifically:

[0036] After the cloud node performs multiple rounds of training on the global model based on global training samples, the distribution probability of the parameter under different values ​​is obtained.

[0037] Furthermore, the intrusion detection result corresponding to the current detection period is a distribution probability corresponding to different intrusion detection results.

[0038] The error of the global model is a comparison between the predicted value (i.e., the output value of the global model) and the actual value. In specific implementation, a set of data can be manually judged. If an attack behavior is simulated, it means that the predicted value is an attack (the value is 1). Then, the data is input into the global model through a sliding window to obtain the probability that the output value of the global model is an attack (the value is 1), and then the error of the global model can be calculated.

[0039] In the embodiment of the present invention, considering that a large number of devices generate a large amount of data, it is impossible for the cloud node to detect all device data. Therefore, an efficient data detection method is used to detect the data. A sliding window with a width of T is initialized in the cloud node, which can randomly select several data sets in each detection cycle to realize system intrusion detection. Considering that a fixed sliding cycle cannot cope with changes in data distribution, the embodiment of the present invention realizes adaptive adjustment of the sliding window based on the error of the global model.

[0040] The intrusion detection method provided by the embodiment of the present invention is that the edge node constructs an edge intrusion detection model for the device feature data of the edge side itself, and the cloud node trains the global model parameters of the cloud node based on the model input and model output of the edge intrusion detection model of all edge nodes, wherein the global model is a model based on Bayesian deep learning; at the same time, the cloud node adaptively adjusts the amount of sampled data input to the global model based on the error accuracy predicted by the global model. The present invention can not only reduce the pressure of communication transmission of the cloud node, but also improve the ability of the cloud node to respond to services. At the same time, the present invention updates the intrusion detection model every detection cycle, and the parameter update of the global model will change with the change of the data distribution of the edge node, thereby improving the accuracy of intrusion detection.

[0041] In an optional implementation, the sliding window size of the current detection cycle is obtained by adjusting the sliding window size of the previous detection cycle based on the error of the global model of the previous detection cycle, specifically:

[0042] The sliding window size of the current detection cycle is calculated according to the following formula:

[0043] T T= =C T-1 ×T T-1

[0044] C T-1 =(1-α×(e T-1 -e0))

[0045] Among them, T T is the sliding window size of the current detection cycle, C T-1 is the sliding window adaptive adjustment coefficient of the previous detection cycle, T T-1 is the sliding window size of the previous detection cycle, α is a parameter between 0 and 1, e T-1 is the error of the global model in the previous detection cycle, and e0 is the preset error threshold.

[0046] In an embodiment of the present invention, when the error of the global model in the previous detection cycle is greater than the error threshold, it is necessary to increase the sliding window and increase the amount of data input to the global model. Conversely, it is necessary to reduce the sliding window and reduce the amount of data of the global model data, so as to adaptively adjust the data size according to the error situation of the global model and reduce the pressure on network transmission.

[0047] In an optional implementation, for each edge node, the edge node performs multiple rounds of training on a pre-built edge intrusion detection model based on the collected device feature data, obtains the input and output of each round of the edge intrusion detection model and uses it as a global training sample, specifically including:

[0048] For the ith edge node, the ith edge node collects device feature data according to the sampling interval corresponding to the ith edge node, and at the same time obtains edge sampling data of a corresponding sampling period as edge samples of each training round at the beginning of each round of edge intrusion detection model training, and inputs the edge samples into the edge intrusion detection model for training;

[0049] When the preset number of training rounds is reached, the input and output of each round of edge intrusion detection model are obtained and used as global training samples.

[0050] It is worth noting that the network environment of edge computing is heterogeneous, geographically distributed, and has great differences in resource processing capabilities. In order to ensure that edge computing intrusion detection has the characteristics of data collection, real-time response, rational use of resources, safe and efficient system services, etc., the embodiment of the present invention takes into account the processing capabilities of edge nodes and adopts a lightweight data processing method to realize the construction of a distributed model. That is, for different edge nodes, training samples are selected based on the different environments of the edge nodes (each edge node has a different data distribution and different computing capabilities), thereby realizing lightweight measurement of intrusion detection in the edge computing environment.

[0051] Furthermore, the edge samples of each training round of the i-th edge node are selected according to the following rules:

[0052] For the i-th edge node, calculate the ratio of the global intrusion detection interval to the sampling interval of the i-th edge node;

[0053] If the ratio is equal to the number of training rounds, the edge node uses the edge sampling data of the j-th sampling period as the edge sample for the j-th round of edge intrusion detection model training;

[0054] If the ratio is not equal to the number of training rounds, the edge node uses the edge sampling data of any sampling period between the b*(j-1)+1 to b*j sampling periods as the edge sample for the j-th round of edge intrusion detection model training; wherein b=ratio / number of training rounds.

[0055] For example, suppose the detection cycle of cloud nodes is once every 2 hours. However, the sampling interval of edge nodes is once every 5 minutes for some and once every half an hour for others. If 4 rounds of training are set to estimate the parameters of the global model, then the edge nodes that collect device feature data once every half an hour just use the data of 4 sampling cycles. However, for edge nodes that collect device feature data once every 5 minutes, the device feature data of a certain sampling cycle can be randomly selected every half an hour for training.

[0056] Furthermore, the sampling interval corresponding to the i-th edge node is specifically:

[0057]

[0058] Among them, INT i is the sampling interval corresponding to the i-th edge node, k is a constant, and f i is the sampling parameter of the device corresponding to the i-th edge node, INT a is the preset standard sampling interval. k is selected based on the actual network conditions. The actual network conditions need to take into account the network protocols between edge nodes and IoT terminals. Current network protocols include 5G, 4G, 3G, LoRa, WiFi, Zigbee, etc. f, as a sampling parameter of IoT devices, is closely related to the data transmission frequency of the device.

[0059] Based on the sampling interval set by the above rules, each edge node builds its own edge intrusion detection model. These edge intrusion detection models are geographically dispersed, and the data distribution processed by each edge node is different. The access of its equipment is heterogeneous, so the parameter estimation of the global model is different.

[0060] Based on the technical solutions provided above, it can be seen that the lightweight of this application is reflected in the sample screening. Training samples are selected based on the different environments of edge nodes (each edge node has a different data distribution and different computing power), which is suitable for lightweight measurement of intrusion detection in edge computing environments. In this way, it is different from the centralized control model in the traditional cloud environment. It not only reduces the pressure of cloud node communication transmission, but also improves the ability of cloud nodes to respond to services. The efficiency is reflected in the division of labor and coordination mechanism between edge nodes and cloud nodes. The edge nodes use specific rules to implement the construction of the model, and the cloud service nodes use sliding windows of adaptive size to process data streams, and control the granularity of data streams according to different security control requirements (mainly reflected in the accuracy of prediction errors). The detection accuracy reflects the online update of the knowledge-driven intrusion detection system, that is, the distribution estimation of the learning parameters of the cloud node intrusion detection model is realized based on the Bayesian deep learning model, and the online update of the cloud node intrusion detection model is realized.

[0061] A second aspect of an embodiment of the present invention provides an intrusion detection system, including a cloud node and a plurality of edge nodes; wherein:

[0062] The edge node is used to perform multiple rounds of training on the pre-built edge intrusion detection model based on the collected device feature data in the current detection cycle, obtain the input and output of each round of the edge intrusion detection model and use it as a global training sample;

[0063] The cloud node is used to train the parameters of the pre-built global model based on the global training sample to obtain the global model with optimized parameters;

[0064] The cloud node is also used to input the sampling data collected in the current detection cycle into the global model after parameter optimization to obtain the intrusion detection result of the current detection cycle; wherein, the data volume of the sampling data collected in the current detection cycle is determined according to the sliding window size of the current detection cycle sampling, and the sliding window size of the current detection cycle is obtained by adjusting the size of the sliding window of the previous detection cycle based on the error of the global model of the previous detection cycle.

[0065] In summary, this application provides a set of efficient, lightweight, high-accuracy intrusion detection models with real-time perception and data granularity adaptive control, with the following specific advantages:

[0066] (1) The constructed intrusion detection model has the characteristics of lightweight and real-time perception, which is reflected in the characteristics of edge nodes, the use of specific rules to implement sample screening, the construction of a lightweight detection model, and the realization of real-time detection of the model;

[0067] (2) Efficient detection is reflected in the cloud node’s use of an adaptive sliding window size to implement the input method of adaptively controlling the granularity of global model data. This not only reduces the pressure on cloud node communication transmission, but also improves the cloud node’s ability to respond to services.

[0068] (3) The detection accuracy is high, which is reflected in driving the online update of the intrusion detection system, that is, based on the Bayesian deep learning model, the distribution estimation of the learning parameters of the cloud node intrusion detection model is realized, and the online update of the cloud node intrusion detection model is realized. The parameter update of the global model will change with the change of the data distribution of the edge node, which further reflects the nature of the online learning of the global model.

[0069] The above is a preferred embodiment of the present invention. It should be pointed out that a person skilled in the art can make several improvements and modifications without departing from the principle of the present invention. These improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. An intrusion detection method, characterized in that: include: In the current detection cycle, for each edge node, the edge node performs multiple rounds of training on the pre-built edge intrusion detection model based on the collected device feature data, obtains the input and output of each round of edge intrusion detection model and uses it as a global training sample; The cloud node trains the parameters of the pre-built global model based on the global training samples to obtain a global model with optimized parameters; wherein the global model is a detection model based on Bayesian deep learning; The cloud node inputs the sampled data collected in the current detection cycle into the global model after parameter optimization to obtain the intrusion detection result of the current detection cycle; wherein the data volume of the sampled data collected in the current detection cycle is determined according to the sliding window size of the current detection cycle sampling, and the sliding window size of the current detection cycle is obtained by adjusting the sliding window size of the previous detection cycle based on the error of the global model of the previous detection cycle; The sliding window size of the current detection cycle is obtained by adjusting the sliding window size of the previous detection cycle based on the error of the global model of the previous detection cycle, specifically: The sliding window size of the current detection cycle is calculated according to the following formula: T T =C T-1 ×T T-1 C T-1 =(1-α×(e T-1 -e0)) Among them, T T is the sliding window size of the current detection cycle, C T-1 is the sliding window adaptive adjustment coefficient of the previous detection cycle, T T-1 is the sliding window size of the previous detection cycle, α is a parameter between 0 and 1, e T-1 is the error of the global model in the previous detection cycle, and e0 is the preset error threshold.

2. The intrusion detection method according to claim 1, characterized in that: For each edge node, the edge node performs multiple rounds of training on a pre-built edge intrusion detection model based on the collected device feature data, obtains the input and output of each round of edge intrusion detection model and uses it as a global training sample, specifically including: For the ith edge node, the ith edge node collects device feature data according to the sampling interval corresponding to the ith edge node, and at the same time obtains edge sampling data of a corresponding sampling period as edge samples of each training round at the beginning of each round of edge intrusion detection model training, and inputs the edge samples into the edge intrusion detection model for training; When the preset number of training rounds is reached, the input and output of each round of edge intrusion detection model are obtained and used as global training samples.

3. The intrusion detection method according to claim 2, characterized in that: The edge samples of each training round of the i-th edge node are selected according to the following rules: For the i-th edge node, calculate the ratio of the global intrusion detection interval to the sampling interval of the i-th edge node; If the ratio is equal to the number of training rounds, the edge node uses the edge sampling data of the j-th sampling period as the edge sample for the j-th round of edge intrusion detection model training; If the ratio is not equal to the number of training rounds, the edge node uses the edge sampling data of any sampling period between the b*(j-1)+1 to b*j sampling periods as the edge sample for the j-th round of edge intrusion detection model training; wherein b=ratio / number of training rounds.

4. The intrusion detection method according to claim 2, characterized in that: The sampling interval corresponding to the i-th edge node is specifically: Among them, INT i is the sampling interval corresponding to the i-th edge node, k is a constant, and f i is the sampling parameter of the device corresponding to the i-th edge node, INT a is the preset standard sampling interval.

5. The intrusion detection method according to claim 1, characterized in that: The cloud node trains the parameters of the pre-built global model based on the global training sample to obtain a global model with optimized parameters, specifically: After the cloud node performs multiple rounds of training on the global model based on global training samples, the distribution probability of the parameter under different values ​​is obtained.

6. The intrusion detection method according to claim 5, characterized in that: The intrusion detection result corresponding to the current detection period is the distribution probability corresponding to different intrusion detection results.

7. An intrusion detection system, characterized in that: It includes cloud nodes and multiple edge nodes; among them, The edge node is used to perform multiple rounds of training on the pre-built edge intrusion detection model based on the collected device feature data in the current detection cycle, obtain the input and output of each round of the edge intrusion detection model and use it as a global training sample; The cloud node is used to train the parameters of the pre-built global model based on the global training samples to obtain a global model with optimized parameters; wherein the global model is a detection model based on Bayesian deep learning; The cloud node is further used to input the sampled data collected in the current detection cycle into the global model after parameter optimization to obtain the intrusion detection result of the current detection cycle; wherein the data volume of the sampled data collected in the current detection cycle is determined according to the sliding window size of the current detection cycle sampling, and the sliding window size of the current detection cycle is obtained by adjusting the sliding window size of the previous detection cycle based on the error of the global model of the previous detection cycle; The sliding window size of the current detection cycle is obtained by adjusting the sliding window size of the previous detection cycle based on the error of the global model of the previous detection cycle, specifically: The sliding window size of the current detection cycle is calculated according to the following formula: T T =C T-1 ×T T-1 C T-1 =(1-α×(e T-1 -e0)) Among them, T T is the sliding window size of the current detection cycle, C T-1 is the sliding window adaptive adjustment coefficient of the previous detection cycle, T T-1 is the sliding window size of the previous detection cycle, α is a parameter between 0 and 1, e T-1 is the error of the global model in the previous detection cycle, and e0 is the preset error threshold.

Citation Information

Patent Citations

  • Deep learning computing system and method for cloud terminal edge computing fusion

    CN107766889A

  • Large-scale MIMO channel estimation method and device and storage medium

    CN113890794A

Cited By

  • Real-time intrusion detection system and method based on deep learning

    CN121485965A