Power key information asset identification method based on network flow characterization learning

Network traffic characterization learning with CNN models addresses the limitations of existing fingerprint recognition methods by accurately identifying power key information assets, enhancing security through non-intrusive classification and updating models.

CN119071027AInactive Publication Date: 2024-12-03ZHENGZHOU UNIVERSITY OF AERONAUTICS
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411076589.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-07
Publication Date
2024-12-03
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

Existing methods for fingerprint recognition of power key information assets in power systems face limitations due to the use of TCP protocol features, which many power information assets do not support, and physical fingerprints are unreliable due to environmental factors and lack of standardization, making it difficult to identify these assets without disrupting normal business communication.

Method used

A method using network traffic characterization learning, specifically through CNN models, to identify power key information assets by capturing, preprocessing, and classifying network traffic data, integrating and quantifying equipment features, and storing them in a fingerprint database, while updating models with new information.

Benefits of technology

Effectively identifies power equipment fingerprints without interfering with normal operations, enhancing security mechanisms like identity authentication and intrusion detection, and improving security protection capabilities.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119071027A_ABST
    Figure CN119071027A_ABST
Patent Text Reader

Abstract

The invention is applicable to the technical field of power grid security, and provides a power key information asset identification method based on network traffic representation learning, which comprises the following steps of: training a classification model by capturing and preprocessing network traffic data of known equipment, classifying unknown equipment traffic by utilizing the model and combining priori knowledge so as to identify equipment individuals, and integrating and quantifying the equipment characteristics to generate an equipment unique fingerprint value for maintaining the equipment unique identifier and information updating in the equipment information base. According to the method, fingerprint identification can be effectively carried out on power key information assets while normal service communication of the power equipment is not interfered, and basic information of the power equipment can be acquired as much as possible. The method not only helps researchers to develop security mechanisms such as identity authentication, security access, asset management, intrusion detection and the like on the electric power information assets, but also helps to improve a penetration test technology in the field of electric power industry, thereby achieving the purpose of improving the security protection capability of the electric power information assets in multiple aspects.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of power grid security, and in particular relates to a method for identifying key power information assets based on network traffic characterization learning. Background Art

[0002] Power information assets are an important part of the power information system, including human resources, services, business processes, documents and data, hardware and facilities, software and systems in the power system. They are large in scale, diverse in types and widely distributed. Assets responsible for the main business of the power grid are called power key information assets, such as intelligent electronic devices (IEDs) in power information systems such as substation automation systems. These devices undertake important functions such as data processing, control and debugging, network communication, and human-computer interaction in the power grid. Threats, assets and vulnerabilities are the three main factors affecting the network security risk assessment of power information systems. Power information assets are both the target of threats and the source of vulnerabilities. They can easily become the main target of attackers when they launch network security attacks on the power grid. The key to improving the security and stability of power information systems is to enhance the security protection capabilities of power information assets.

[0003] There is a lack of research on fingerprint recognition of power information assets in existing research results at home and abroad. In recent years, relevant scholars have studied the fingerprint recognition technology of industrial control system (ICS) equipment from a general perspective. Compared with general Internet systems, power information systems and ICS have higher levels of hierarchy and structure, more stable system topology and lower system component volatility. At the same time, ICS equipment is often resource-constrained, with strict bandwidth restrictions and low latency requirements. This also means that if an active detection method similar to that used by Nmap is implemented on ICS equipment to obtain device fingerprint information, it may block the normal business communication of the device, and in severe cases may even lead to a denial of service (DoS) attack. Therefore, existing research mostly uses passive methods to obtain device fingerprint information, mainly based on the network characteristics and physical characteristics of the device. However, existing research results are mainly aimed at the problem of device fingerprint recognition in ICS systems, and cannot be fully applied to the fingerprint recognition of power key information assets; most of the fingerprint recognition research based on network features uses TCP protocol features, but many power information assets do not support TCP protocol, but use specific power industrial control protocols or even private protocols for communication, and need to extract corresponding protocol features in a targeted manner; physical fingerprints also have certain limitations: radio frequency fingerprints are easily affected by environmental factors and change, noise fingerprints are only applicable to sensors, and the physical operation time of different devices is for different processes, and it is impossible to unify the standards, which is difficult to achieve when there are many types of equipment. For this reason, the present invention proposes a method for identifying power key information assets based on network traffic characterization learning. Summary of the invention

[0004] The purpose of the present invention is to provide a method for identifying key power information assets based on network traffic characterization learning, aiming to solve the problems raised in the above-mentioned background technology.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] A method for identifying power key information assets based on network traffic characterization learning includes the following steps:

[0007] Step 1, preparation phase: First, capture the original network traffic generated by the known power key information assets during communication; then split the original traffic data according to the logical hierarchy, pre-process it into the format of the representation learning model input, and automatically annotate it at the granularity of individual devices; finally, based on the representation learning, use the traffic data of each major category of equipment to train a classification model to realize the identification of individual devices;

[0008] Step 2, Identification phase: First, confirm the source of the network traffic of the device to be identified based on prior knowledge, including the system and type to which the device belongs; then select the corresponding classification model according to the identified device type to classify the traffic, and further identify the individual device; finally, update and retrain the device fingerprint information library and classification model respectively;

[0009] Step 3, equipment fingerprint information processing: the power key information asset equipment features extracted in the preparation stage and the identification stage are integrated, quantified, structured and matched, and stored in the power key information asset fingerprint information database, and the final equipment unique fingerprint value is generated as the unique identifier of the power key information asset related to its own information; if part of the equipment information is found to exist in the fingerprint information database through matching when storing new information, the information database is updated, and the update rule is to use the equipment information obtained by this identification to overwrite the existing information of the same equipment in the information database, and regenerate the equipment fingerprint value.

[0010] Furthermore, the method for selecting the flow characteristics is:

[0011] The network traffic data of substation automation equipment and distribution automation equipment captured on site are parsed and analyzed, and fields associated with the equipment's own information and its business functions are selected from the messages of the protocols used by the two types of equipment as traffic features for classification.

[0012] Furthermore, the representation learning model adopts CNN.

[0013] Furthermore, the specific process of the pre-processing step is:

[0014] a. Traffic splitting: Split the original traffic file based on the unique and unchangeable device source MAC address in the entire network to obtain multiple traffic sub-files with the same format. Then sort the messages in each sub-file in chronological order to obtain the message flow sent by each device.

[0015] b. Traffic cleaning, including:

[0016] Message screening: Perform compliance checks on the captured traffic data according to the frame format of the power industrial control protocol message specified in the standard, discard malformed messages with abnormal content or format, and screen out messages that meet the standards for characterization learning;

[0017] Message structure adjustment: During cleaning, the selected feature fields in the message are retained, the other parts are discarded, and the retained fields are sequentially spliced ​​to obtain a new message; during training, the part of each message that exceeds 784 bytes is discarded, and the message that is less than 784 bytes is filled with 0x00;

[0018] c. Data type conversion: convert the cleaned message data with a uniform length of 784 bytes into a 28pix*28pix grayscale image, and then convert the generated grayscale image into an IDX file.

[0019] Furthermore, in step 2, if the device cannot be successfully identified, the device information is obtained manually.

[0020] Furthermore, the following specific processes are included:

[0021] The original traffic files of the devices that have been identified by the topology level and whose type range has been preliminarily determined are preprocessed. The preprocessing process includes three stages: splitting, cleaning, and type conversion. The IDX files obtained after preprocessing are used as the input of the traffic classification model. The topology level code of the corresponding device is obtained from the device topology level list generated after the topology level identification according to the source MAC address, the device type range is determined, and the characterization learning classification model trained for the corresponding type of device is selected to further realize the fine-grained identification of the device.

[0022] When training the representation learning model, the key information asset flow data of electric power collected in real time during communication is used; when new equipment is added to the network, the network traffic is recollected to retrain and partially update the model.

[0023] Compared with the prior art, the present invention has the following beneficial effects:

[0024] This method of identifying power key information assets based on network traffic characterization learning can effectively identify the fingerprints of power equipment (i.e., power key information assets) without interfering with the normal business communications of power equipment, and obtain as much basic information of power equipment as possible. This not only helps researchers to conduct research on security mechanisms such as identity authentication, secure access, asset management, and intrusion detection for power information assets, but also helps to improve penetration testing technology in the power industry, thereby achieving the goal of improving the security protection capabilities of power information assets in many aspects. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 This is a typical substation automation system architecture diagram.

[0026] Figure 2 This is a typical schematic diagram of the distribution automation system architecture.

[0027] Figure 3 It is an implementation flow chart of the present invention.

[0028] Figure 4 This is a flow chart for preprocessing the power key information asset flow data in the present invention.

[0029] Figure 5 This is a flow chart for cleaning the flow of key power information assets in the present invention.

[0030] Figure 6 It is a schematic diagram of message structure adjustment in the present invention.

[0031] Figure 7 It is a grayscale image generated from the flow data of the process layer equipment in the present invention.

[0032] Figure 8 This is the IDX file structure generated by the traffic data conversion in the present invention.

[0033] Fig. 9 Schematic diagram of the CNN representation learning model architecture in the present invention.

[0034] Fig.10 This is a diagram of the device traffic identification process based on representation learning in the present invention. DETAILED DESCRIPTION

[0035] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0036] The specific implementation of the present invention is described in detail below in conjunction with specific embodiments.

[0037] The power grid is mainly composed of two parts: the primary system and the secondary system. Among them, the primary power system is directly related to power generation, transmission, substation and other businesses. It is the main body of the power supply load and the carrier of the power load. It can be regarded as the physical system part of the power grid; the secondary power system is used to provide the data collection, information transmission, information processing, device control and other functions required for the safe operation of the primary system. It is composed of some equipment, software and the interactive information therein, and can be regarded as the information system of the power grid. The secondary power system is divided into two parts, namely the production control area and the management information area. Among them, the production control area can be divided into safety zone I and safety zone II. Among them, safety zone I is directly connected to the primary power equipment and is responsible for real-time monitoring of the operation status of the power grid. Its safety is directly related to whether the power production link can proceed smoothly. It is the most important part of the secondary power system. Under this premise, we comprehensively considered the typicality of the business systems in security zone I and the feasibility of information collection and experiments in the research, and selected the main equipment of the substation automation system (SAS) and the main equipment of the distribution automation system (DAS) as the research objects, namely, the key power information assets referred to in this invention.

[0038] like Figure 1 The substation automation system shown in the figure realizes the three major functions of control, monitoring and protection of the substation by utilizing advanced information and communication technology to functionally reorganize and optimize the information assets in the substation. The system mainly includes merging units, intelligent terminals, measurement and control devices, protection devices and the host of the monitoring master station.

[0039] like Figure 2 The figure shows that the distribution automation system uses modern electronic technology, computer network technology, and communication technology to manage the distribution network. Its functions mainly include substation automation, distribution management automation, feeder automation, and user automation to achieve the monitoring, protection, control, and modernization of power consumption and distribution management of the distribution network. To a certain extent, it can be said to be an extension of the substation automation system, but compared with the substation automation system, it has a larger scope and contains more types of equipment, mainly including three types of feeder terminals (Feeder Terminal Unit, FTU), distribution terminal (Distribution Terminal Unit, DTU), and distribution transformer terminal (Transformer Terminal Unit, TTU). Common ones include intelligent circuit breakers, intelligent switches, intelligent relays, intelligent current transformers, intelligent voltage transformers, intelligent energy meters, and intelligent controllers.

[0040] like Figure 3 As shown, a method for identifying key power information assets based on network traffic characterization learning is provided in one embodiment of the present invention, comprising the following steps:

[0041] Step 1, preparation phase: First, capture the original network traffic generated by the known power key information assets during communication; then split the original traffic data according to the logical hierarchy, pre-process it into the format of the representation learning model input, and automatically annotate it at the granularity of individual devices; finally, based on the representation learning, use the traffic data of each major category of equipment to train a classification model to realize the identification of individual devices;

[0042] Step 2, identification stage: Since the topological level discrimination of substation automation equipment and distribution automation equipment is based on different communication mode characteristics, the network traffic source of the device to be identified is first confirmed based on prior knowledge, including the system and type to which the device belongs; then, the classification model of the corresponding category is selected according to the identified device type to classify the traffic, and further realize the identification of the individual device; if the device cannot be successfully identified in the above process, it is necessary to obtain the device information manually. Finally, the device fingerprint information library and classification model are updated and retrained respectively;

[0043] Step 3, equipment fingerprint information processing: the power key information asset equipment features extracted in the preparation stage and the identification stage are integrated, quantified, structured and matched, and stored in the power key information asset fingerprint information database, and the final equipment unique fingerprint value is generated as the unique identifier of the power key information asset related to its own information; if part of the equipment information is found to exist in the fingerprint information database through matching when storing new information, the information database is updated, and the update rule is to use the equipment information obtained by this identification to overwrite the existing information of the same equipment in the information database, and regenerate the equipment fingerprint value.

[0044] As a preferred embodiment of the present invention, the method for selecting the flow characteristics is:

[0045] The network traffic data of substation automation equipment and distribution automation equipment captured on site is parsed and analyzed, and fields associated with the equipment's own information and its business functions are selected from the messages of the protocols used by the two types of equipment as traffic features for classification. As shown in Table 1:

[0046] Table 1 Flow characteristics of key power information assets

[0047]

[0048] As a preferred embodiment of the present invention, the selection of the representation learning model: Convolutional Neural Networks (CNN) has been widely used in many fields. Many scholars have used CNN to achieve traffic recognition and achieved good results. Therefore, we use CNN as the representation learning model used when classifying traffic for power key information assets (other models can also be selected).

[0049] The specific architecture of the two-dimensional CNN representation learning model used is as follows Fig. 9 The detailed design of the model parameters is shown in Table 2:

[0050] Table 2 Key parameter design of CNN model architecture

[0051]

[0052] As a preferred embodiment of the present invention, the specific process of the pretreatment step is as follows (eg Figure 4 shown):

[0053] a. Traffic splitting: Since our goal is to identify individual devices, we choose to split the original traffic file based on the device source MAC address, which is unique in the entire network and not easy to change, to obtain multiple traffic sub-files with the same format (all pcap files). After splitting, the messages in each sub-file are sorted in chronological order to obtain the message flow sent by each device;

[0054] b. Traffic cleaning, the process is as follows Figure 5 As shown, including:

[0055] Message screening: Considering that the traffic data captured in the field may generate malformed messages due to errors in the transmission process, that is, the message content is incomplete or the data is abnormal, direct use for training may affect the accuracy of the model. Therefore, it is necessary to perform compliance checks on the captured traffic data according to the frame format of the power industrial control protocol message specified in the standard, discard malformed messages with abnormal content or format, and filter out messages that meet the standards for representation learning, thereby reducing interference with the model training effect;

[0056] The message structure is adjusted, and the process is as follows Figure 6 As shown: the fields in Table 1 are retained during cleaning, the other parts are directly discarded, and the retained fields are spliced ​​in sequence to obtain a new message, so as to achieve fuzzy processing of irrelevant information to a certain extent, and eliminate the interference of irrelevant information on model training. In addition, the input size of the CNN representation learning model is generally fixed at 784 bytes, which is smaller than the message size of some power industrial control protocols. During training, the part of each message exceeding 784 bytes will be discarded, and the message less than 784 bytes will be filled with 0x00. The message cleaning method proposed in the present invention can advance the relative position of the information associated with the device itself in the message, and retain it as much as possible, so that the trained representation learning model can more accurately classify individual devices.

[0057] c. Data type conversion: Since the grayscale value of a single pixel in a grayscale image ranges from 0 to 255, which corresponds exactly to the value range of an 8-bit binary number, i.e., one byte, the message data with a uniform length of 784 bytes after cleaning can be converted into a grayscale image of 28pix*28pix. Figure 7 The grayscale images generated by randomly selecting three frames of different messages from the flow data of three process-layer devices named a, b, and c are shown. As can be seen from the figure, the grayscale images generated by the messages of different devices have obvious differences visible to the naked eye, that is, the grayscale images generated by the flow of individual devices of different power key information assets are more distinguishable, and using the CNN model that is good at image classification to classify the flow of power key information assets should be able to achieve better results.

[0058] The generated grayscale image is then converted into an IDX file. However, since both the message data and the grayscale image grayscale value data are in the form of byte streams and have no essential difference, the message is directly converted into an IDX file during the actual model training and testing phase. The purpose of generating a grayscale image here is more to facilitate intuitive analysis of the effect of traffic classification during the research phase. The generated IDX file structure is as follows: Figure 8 shown.

[0059] As a preferred embodiment of the present invention, the device traffic identification process based on representation learning follows the principle of first classifying into broad categories and then fine-grained identification. After determining the device type range, a representation learning classification model is trained for each broad category of devices to classify the traffic for further fine-grained identification. The identification process of the traffic identification part is as follows: Fig.10 shown.

[0060] When identifying traffic, the original traffic files of the devices that have been identified at the topological level and whose type ranges have been preliminarily determined are first preprocessed. The preprocessing process is divided into three stages: splitting, cleaning, and type conversion. In the splitting stage, the original traffic files of the devices with a clear category range are split according to the source MAC address to obtain the message flow sent by each device, and finally the original traffic files of each device are converted into corresponding IDX files after cleaning and type conversion as the input of the representation learning classification model.

[0061] According to the source MAC address, the topology level code of the corresponding device is obtained from the device topology level list generated after topology level identification. Based on this, the device type range is determined, and the representation learning traffic classification model trained for the corresponding type of device is selected to further realize fine-grained identification of the device.

[0062] The training of the representation learning model uses the power key information asset flow data collected in real time during communication. When a new device is added to the network, the network traffic needs to be recollected to retrain and update the model. Since each classifier is only used to classify and identify traffic for each major category of equipment, only the classifier corresponding to the type of new equipment needs to be partially updated, without large-scale updates to all models, thereby reducing the workload caused by model updates while improving the classification accuracy of the model.

[0063] The above are only preferred embodiments of the present invention. It should be pointed out that, for those skilled in the art, several modifications and improvements can be made without departing from the concept of the present invention. These should also be regarded as the protection scope of the present invention. These will not affect the effect of the implementation of the present invention and the practicality of the patent.

Claims

1. A method for identifying power key information assets based on network traffic characterization learning, characterized in that: The following steps are involved: Step 1, preparation phase: First, capture the original network traffic generated by the known power key information assets during communication; then split the original traffic data according to the logical hierarchy, pre-process it into the format of the representation learning model input, and automatically annotate it at the granularity of individual devices; finally, based on the representation learning, use the traffic data of each major category of equipment to train a classification model to realize the identification of individual devices; Step 2, Identification phase: First, confirm the source of the network traffic of the device to be identified based on prior knowledge, including the system and type to which the device belongs; then select the corresponding classification model according to the identified device type to classify the traffic, and further identify the individual device; finally, update and retrain the device fingerprint information library and classification model respectively; Step 3, equipment fingerprint information processing: The power key information asset equipment features extracted in the preparation stage and the identification stage are integrated, quantified, structured and matched, and stored in the power key information asset fingerprint information database, and the final equipment unique fingerprint value is generated as the unique identifier of the power key information asset related to its own information; If part of the device information is found to exist in the fingerprint information library through matching when storing new information, the information library will be updated. The update rule is to overwrite the existing information of the same device in the information library with the device information obtained by this identification, and regenerate the device fingerprint value.

2. The method for identifying key power information assets based on network traffic characterization learning according to claim 1 is characterized in that: The method for selecting the flow characteristics is: The network traffic data of substation automation equipment and distribution automation equipment captured on site are parsed and analyzed, and fields associated with the equipment's own information and its business functions are selected from the messages of the protocols used by the two types of equipment as traffic features for classification.

3. The method for identifying key power information assets based on network traffic characterization learning according to claim 1 is characterized in that: The representation learning model adopts CNN.

4. The method for identifying key power information assets based on network traffic characterization learning according to claim 1 is characterized in that: The specific process of the pre-treatment step is: a. Traffic splitting: Split the original traffic file based on the unique and unchangeable device source MAC address in the entire network to obtain multiple traffic sub-files with the same format. Then sort the messages in each sub-file in chronological order to obtain the message flow sent by each device. b. Traffic cleaning, including: Message screening: Perform compliance checks on the captured traffic data according to the frame format of the power industrial control protocol message specified in the standard, discard malformed messages with abnormal content or format, and screen out messages that meet the standards for characterization learning; Message structure adjustment: During cleaning, the selected feature fields in the message are retained, the other parts are discarded, and the retained fields are sequentially spliced ​​to obtain a new message; during training, the part of each message that exceeds 784 bytes is discarded, and the message that is less than 784 bytes is filled with 0x00; c. Data type conversion: The cleaned message data with a uniform length of 784 bytes is converted into a 28pix*28pix grayscale image, and then the generated grayscale image is converted into an IDX file.

5. The method for identifying power key information assets based on network traffic characterization learning according to claim 1 is characterized in that: In the step 2, if the device cannot be successfully identified, the device information is obtained manually.

6. The method for identifying power key information assets based on network traffic characterization learning according to claim 1 is characterized in that: The specific process includes the following: Preprocess the original traffic files of the devices that have been identified at the topological level and whose types have been preliminarily determined. The preprocessing process includes three stages: splitting, cleaning, and type conversion. The IDX files obtained after preprocessing are used as the input of the traffic classification model. Obtain the topology level code of the corresponding device from the device topology level list generated after topology level identification according to the source MAC address, determine the device type range, and select the representation learning classification model trained for the corresponding type of device to further realize fine-grained identification of the device; When training the representation learning model, the key information asset flow data of electric power collected in real time during communication is used; when new equipment is added to the network, the network traffic is recollected to retrain and partially update the model.