An industrial control honeypot interactive system based on time series prediction

By combining neural network model and time series prediction technology, the status of industrial control equipment is simulated and prediction data is generated, the problem of insufficient interactive capabilities of industrial control honeypots is solved, effective deception and information collection is achieved against attackers, and the defense capabilities of industrial control networks are improved .

CN112182564BActive Publication Date: 2024-11-26NORTHEASTERN UNIV CHINA
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202010844667.5
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2020-08-20
Publication Date
2024-11-26
Estimated Expiration
2040-08-20

AI Technical Summary

Technical Problem

The existing industrial honeypots are insufficient in terms of interaction capabilities and cannot effectively respond to attackers' requests. They are easily visible and lack targeted to industrial control systems. It is difficult to make immediate responses when attackers monitor equipment status.

Method used

Combined with neural network model prediction technology, time series prediction method is used to simulate the state changes of industrial control equipment, generate prediction data, improve the interaction ability and real-time nature of honeypots, and predict data through two-way LSTM network to generate reasonable response data, deceive attackers and Capture attack information.

Benefits of technology

It improves the deceptiveness and interaction ability of honeypots, can effectively respond to attackers without exposing the real device status, collect attack information, and improves the active defense capabilities of the industrial control network.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN112182564B_ABST
    Figure CN112182564B_ABST
Patent Text Reader

Abstract

The present invention belongs to the field of network security technology, and discloses an industrial control honeypot interaction system based on time series prediction. The present invention uses prediction data to make predictions, thereby improving the security of data. Industrial control data is mostly quasi-periodic and has certain rules. Using prediction data for prediction avoids re-entering real data, thereby ensuring the real-time nature of honeypot interaction. The present invention uses a time series prediction method to make long-term predictions on the state changes of real equipment in industrial control scenarios, and combines honeypot technology to complete deep simulation of industrial control equipment, and makes instant response information to attackers that conforms to the industrial control scenario, thereby improving the deceptiveness of the honeypot, and can lure attackers to attack multiple times while also collecting attack information, which is conducive to profiling attackers, turning passivity into initiative, and better maintaining the security of industrial control networks.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of network security, and in particular relates to an industrial control honeypot interaction system based on a time series prediction algorithm. Background Art

[0002] In recent years, as the industrial control system network and the Internet of Things environment have become more open and changeable, industrial control systems have become relatively more vulnerable, and various network attacks on industrial control systems have increased, exposing the serious deficiencies in the security protection of industrial control systems. The security of industrial control systems faces huge challenges.

[0003] However, there are significant differences between industrial control networks and traditional IT information networks in terms of network edge, architecture, and transmission content. In addition, industrial control networks themselves also have many vulnerabilities and industry-specific requirements in terms of security defense. Honeypot technology, as an active defense technology, can detect attacks from outside and inside the network system without changing the network structure. Although experts and scholars at home and abroad have tried to use different honeypot technologies to solve network security problems, most of them are limited to the fields of malware, malicious code sample capture, botnet tracing, and intrusion detection, and most of the research objects are traditional IT networks. There are relatively few studies on honeypots for industrial control networks. In addition, the existing industrial control honeypots are mostly low-interaction in terms of interaction capabilities, and cannot reasonably and effectively respond to attackers' requests in combination with actual industrial control scenarios. They are easily discovered by attackers and lose their actual value. Therefore, improving the interaction capabilities of industrial control honeypot systems is of great significance to improving the level of active defense systems in industrial control networks. Since attacks on industrial control systems are highly purposeful, attackers need to monitor the status of industrial control equipment for a long time before carrying out the actual attack to determine whether the device is their target. However, there is currently no effective method to make instant responses that are consistent with the actual status of industrial control equipment when attackers monitor the device status. Summary of the invention

[0004] In view of the problem of insufficient interaction capability in the above-mentioned existing honeypot technology, the present invention combines the neural network model prediction technology and the industrial control network honeypot interaction technology for the first time, and proposes an industrial control honeypot interaction optimization scheme based on time series prediction. Ordinary time series prediction uses real data for prediction, and too long a delay will lose the connection with the attacker. The present invention uses predicted data for prediction, which improves the security of the data. Most industrial control data is quasi-periodic and has certain rules. Using predicted data for prediction avoids the need to input real data again, ensuring the real-time nature of honeypot interaction.

[0005] The honeypot uses time series prediction methods to simulate the state changes of real industrial control equipment and generate reasonable prediction data. When an attacker invades and the honeypot receives the attacker's request information, it responds to the attacker based on the estimated state of the real device at the current time node, which is both convincing and does not expose the accurate working state of the real industrial control equipment. The success of the honeypot interaction attracts the attacker to attack again. At the same time, it can also capture the attacker's traffic information and parse the corresponding industrial control protocol data packets to obtain the attacker's attack type, attack method, attack address and geographic location information, generate threat warning information, and provide effective protection strategies for industrial control network security.

[0006] The present invention is an industrial control honeypot interaction system based on time series prediction, comprising the following modules:

[0007] 101 interface module: used to implement the standard interface of the protocol used by the current industrial control network to connect to the Internet; when receiving communication data, the honeypot filters out the data packets with industrial protocols at the application layer and discards other invalid data packets;

[0008] 102 Industrial control protocol parsing module: used to parse industrial protocol data packets; when the honeypot receives a data packet request containing an industrial protocol, the honeypot will disassemble the data packet by field according to the protocol specification of the protocol, and store the original data flow and parsing results in the database;

[0009] 103 Data acquisition module: responsible for completing the collection of data on the status changes of on-site industrial control equipment; and saving the collected results in a file in timestamp order to provide data support for subsequent prediction modules; since this type of industrial control data is mostly periodic and has certain regular characteristics, and in order to ensure the safety of the working status of real industrial equipment and avoid the operational burden of re-entering real data, the present invention uses the data results of the previous prediction as training data for the next prediction.

[0010] 104 Prediction module: responsible for using the time series prediction model to train the data generated by the data acquisition module and generate long-term prediction results; the time series prediction model consists of two parts: encoding and decoding. In the encoding stage, a bidirectional LSTM network is used to convert the input data into an intermediate vector; in the decoding stage, the LSTM network is used to use the intermediate vector and the output of the previous moment to generate new prediction data;

[0011] 105 Interaction module: used to respond to the attacker's request information; determine whether the attacker attempts to read the device's status information through the protocol analysis result; if it does not involve reading the status information, respond to the request information according to the protocol specification; if it involves reading the status information, write the device status prediction result of the prediction module into the response data packet to complete the deep interaction with the attacker;

[0012] 106 Data storage module: responsible for recording the request information sent to the honeypot. This data will be used for threat intelligence analysis and serve as a threat warning. The attacker's geographic location is obtained and recorded based on the IP address in the attacker's request information and the captured network flow data. In addition, the time of the attack, the attacker's organization, and the attack behavior information will also be recorded and stored.

[0013] Further, the above-mentioned 104 prediction module: generating new prediction data has the following specific steps:

[0014] 1) Use sliding window method for data input, the model input is {x t1 ,x t2 ,...,x tn}, tn represents the nth recorded time node of the industrial control equipment status change, x tn Represents the data value of the actual industrial control equipment status change at time tn, and the output is {x tn+1 ,x tn+2 ,...,x tn+m}, m is the time step;

[0015] 2) The input data is sent to the encoder, and its hidden layer state is calculated through LSTM forward and backward propagation, and the hidden layer state is weighted averaged to obtain the intermediate vector. The formula is as follows:

[0016]

[0017] Where C is the intermediate vector, h t is the hidden layer state of the tth layer in the encoding stage, α j It is a 0-1 parameter generated according to the importance of the hidden layer state of the jth layer. enc-for (x,h) is the LSTM forward propagation calculation function, LSTM enc-back (x,h) is the LSTM back-propagation calculation function, and the intermediate vector C is calculated by the sum of the weighted hidden layer states at each time step using the tanh activation function.

[0018] 3) In the decoding stage, the intermediate vector and the output of the previous moment are used to obtain the current prediction data; the formula is as follows

[0019]

[0020] Where V is the initial vector matrix, h 0 is the initial hidden layer state in the decoding stage, h t is the hidden layer state of the tth layer in the decoding stage, P is the model output, and the value with the largest probability value is selected for output through the softmax function;

[0021] 4) Add the output data to the original data to determine whether the set prediction time step is reached. If not, repeat the above process until the set prediction time step is reached to complete the long-term prediction of the device status.

[0022] The above-mentioned data on the change of the state of the industrial control equipment includes the change of the register value and the change of the coil value.

[0023] The present invention has the following advantages and beneficial technical effects:

[0024] The present invention uses a time series prediction method to make long-term predictions on the real device status changes in industrial control scenarios, and combines honeypot technology to complete deep simulation of industrial control equipment, and make instant response information to attackers that conforms to the industrial control scenario, thereby improving the deceptiveness of the honeypot, luring attackers to attack multiple times while also collecting attack information, which is conducive to profiling attackers, turning passivity into initiative, and better maintaining the security of industrial control networks. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 It is a structural schematic diagram of the present invention.

[0026] Figure 2 It is a line chart of the test data set and the corresponding time series prediction results DETAILED DESCRIPTION

[0027] The present invention aims to improve the honeypot interaction capability and provides an implementation scheme of industrial control honeypot interaction technology using time series prediction. For ease of understanding, the technical implementation scheme of the present invention and its features are described as follows.

[0028] When using the present invention, users who have more frequent debugging and modification needs can directly deploy the present invention on their personal computers, and users who do not have such needs can also choose to deploy it in a cloud server environment. When testing this product, a personal computer is selected as the honeypot host, and the client of the interactive behavior module, prediction module, data acquisition module and industrial control protocol parsing module in the honeypot system is deployed on the honeypot device, and the server side of the data storage module is deployed on the data processing server. In order to display the obtained data, the data visualization module is additionally used and deployed on the visualization display host. After deployment, the present invention starts working and waits to capture threat attack information data on the Internet. The collected data intelligence can also be used for subsequent analysis. To complete the above work, the present invention includes the following modules:

[0029] 101 interface module:

[0030] The first thing to configure in the whole system is the interface module. Use the socket programming technology under the Linux system. Create a process and apply for a socket as the server in the Client-Server structure. Then initialize the socket address structure, select TCP / IP as the protocol, and then configure the IP and port number to wait for the request to connect.

[0031] 102 Industrial control protocol parsing module: After the attacker establishes a connection with the honeypot, the honeypot receives the request message sent by the attacker as the server, saves the application layer data of the message, records the attacker's IP address and uses GeoLite2-City.mmdb to locate its geographical location, and then processes the request message. For messages that meet the protocol specifications used by the honeypot, extract its function code field and analyze the attack intention. Repeat the above steps until the connection between the attacker and the honeypot system is closed. For the captured attacker information, the data capture module will also write this information into the log file for users to view.

[0032] 103 Data acquisition module: The test uses the flow data of the SwaT water treatment scenario as the initial data. The data is from December 2015, and the prediction for the current moment is

[0033] 104 prediction module:

[0034] (1) Data processing

[0035] For the original industrial control network traffic data, we first extract traffic features based on the specific industrial scenarios where the traffic is generated. During the test, we extract traffic information feature values ​​from the data that can be used to predict the interactive response behavior of industrial control equipment in the scenario. Then, we first need to filter the obtained raw data information to remove unreasonable data information such as blank data and erroneous data to ensure the validity of the data. Secondly, we need to normalize the data to facilitate the training of the neural network model.

[0036] (2) Model training and prediction

[0037] In the encoding stage, two LSTM neural networks with the same time step are stacked to obtain a bidirectional LSTM (Bi-LSTM), so that the network model can learn the connection between the data feature time in adjacent time steps and their historical time steps. In the decoding stage, a unidirectional LSTM neural network is used to predict the value. The network model training uses a batch size of 64, an epoch of 20, and an initial learning rate of 0.002. To prevent the gradient explosion of the LSTM neural network, the gradient is clipped during the training process, and the clipping threshold is set to 500. The stochastic gradient descent algorithm is used in the network model training process. After 500 and 1500 batches, the learning rate is reduced to 0.2 times and 0.1 times the original. In the training process of the neural network, the original training data set needs to be divided into two parts, used for model training and model verification respectively. In order to ensure the time series of the training data, the data of the first 60% of the time steps are taken as the training set, and the data of the first 60%-80% of the time steps are taken as the test set. The LSTM-based time series prediction model is used to compare the output of the test set with the last 20% of the verification data sorted by time to ensure the rationality of the output prediction sequence of the selected model.

[0038] In addition, in order to prove that LSTM enc-dec To compare the superiority of the algorithm, the experiment selected the RNN model as the experimental result, and used the Mean Absolute Percentage Error (MAPE) as the error value measurement standard. Compared with the Mean Square Error (MSE), it not only considers the error between the predicted value and the true value, but also considers the ratio of the error to the true value. The calculation formula is as follows.

[0039]

[0040] The comparison results are shown in the following table

[0041] Table 1 MAPE error of two models at different time steps

[0042]

[0043]

[0044] As shown in the table, in the prediction data of different step sizes of most sensors, LSTM enc-dec The predicted results are mostly better.

[0045] 105 interactive modules:

[0046] The main function of the interaction module is to complete the response to the visitor. The interaction module generates an interaction data frame based on the analysis of the industrial control protocol communication data frame according to the attack intention obtained by the industrial control protocol analysis module, and then uses the prediction module to predict the state value of the target sensor at the time of the attack to fill the fixed data field. Then, the socket process of the interface module is also used to send the response data frame back to the attacker using a function.

[0047] 106 Data Storage Module:

[0048] The present invention uses the table structure of the MySQL database to store the relevant information of the attacker's IP, including the attack time and the geographical location of the attacking IP. At the same time, in order to make these data more intuitive and vivid, the present invention also has a visual interface to display the data obtained in the database, and uses a variety of charts to analyze the attacker's information from multiple angles, which is helpful for the next step of threat intelligence analysis.

[0049] The present invention uses a time series prediction method to make long-term predictions on the status of real industrial control equipment, completes deep interaction with intruders, and optimizes the industrial control deception environment. The present invention simulates the data transmission process of the real industrial control environment, and the honeypot disguises itself as industrial equipment in the industrial control environment. It follows the protocol used in the environment, can receive instructions and respond, and achieves the purpose of deceiving attackers and capturing attacker data. At the same time, it can also trace the attacker to the source and analyze their attack methods and attack purposes.

[0050] The present invention simulates industrial equipment in different industrial control environments by constructing honeypots using different protocols, including nearly 20 industrial control protocols such as Modbus / TCP, Ethernet / IP, Siemens S7, and DNP3. The industrial standard interfaces of multiple protocols are implemented to simulate multiple communication processes. In the interactive stage, the time series prediction method is used to respond to the attacker's request in accordance with the real industrial control environment, thereby enhancing the simulation of the honeypot to the industrial control equipment, which not only improves the deception but also does not damage the security.

[0051] It should be noted that the above description is a specific implementation method provided in combination with specific content, and it cannot be determined that the present invention is limited to the above description. Any method similar to the method, structure, or device of the present invention, or a method in which a number of substitutions or deductions are made under the concept of the present invention, should be regarded as the protection scope of the present invention.

Claims

1. An industrial control honeypot interaction system based on time series prediction, characterized in that: Contains the following modules: 101 interface module: used to implement the standard interface of the protocol used by the current industrial control network to connect to the Internet; when receiving communication data, the honeypot filters out the data packets with industrial protocols at the application layer and discards other invalid data packets; 102 Industrial control protocol parsing module: used to parse industrial protocol data packets; when the honeypot receives a data packet request containing an industrial protocol, the honeypot will disassemble the data packet by field according to the protocol specification of the protocol, and store the original data flow and parsing results in the database; 103 Data acquisition module: responsible for collecting data on status changes of on-site industrial control equipment; and The collected results are saved in files in timestamp order to provide data support for subsequent prediction modules; 104 Prediction module: responsible for using the time series prediction model to train the data generated by the data acquisition module and generate long-term prediction results; the time series prediction model consists of two parts: encoding and decoding. In the encoding stage, a bidirectional LSTM network is used to convert the input data into an intermediate vector; in the decoding stage, the LSTM network is used to use the intermediate vector and the output of the previous moment to generate new prediction data; 105 Interaction module: used to respond to the attacker's request information; determine whether the attacker attempts to read the device's status information through the protocol analysis result; if it does not involve reading the status information, respond to the request information according to the protocol specification; if it involves reading the status information, write the device status prediction result of the prediction module into the response data packet to complete the deep interaction with the attacker; 106 Data storage module: responsible for recording the request information sent to the honeypot. These data will be used for threat intelligence analysis and play a threat warning role. The attacker’s geographic location is obtained and recorded based on the IP address in the attacker’s request information and the captured network flow data. In addition, the time of the attack, the attacker’s organization, and the attack behavior information will also be recorded and stored; The 104 prediction module: generating new prediction data has the following specific steps: 1) Use sliding window method for data input, the model input is {x t1 ,x t2 ,...,x tn }, tn represents the nth recorded time node of the industrial control equipment status change, x tn Represents the data value of the actual industrial control equipment status change at time tn, and the output is {x tn+1 ,x tn+2 ,...,x tn+m }, m is the time step; 2) The input data is sent to the encoder, and its hidden layer state is calculated through LSTM forward and backward propagation, and the hidden layer state is weighted averaged to obtain the intermediate vector. The formula is as follows: Where C is the intermediate vector, h t is the hidden layer state of the tth layer in the encoding stage, α j It is a 0-1 parameter generated according to the importance of the hidden layer state of the jth layer. enc-for (x,h) is the LSTM forward propagation calculation function, LSTM enc-back (x,h) is the LSTM back-propagation calculation function, and the intermediate vector C is calculated by the sum of the weighted hidden layer states of each time step using the tanh activation function; 3) In the decoding stage, the intermediate vector and the output of the previous moment are used to obtain the current prediction data; the formula is as follows Where V is the initial vector matrix, h 0 is the initial hidden layer state in the decoding stage, h t is the hidden layer state of the tth layer in the decoding stage, P is the model output, and the value with the largest probability value is selected for output through the softmax function; 4) Add the output data to the original data to determine whether the set prediction time step is reached. If not, repeat the above process until the set prediction time step is reached to complete the long-term prediction of the device status.

2. The industrial control honeypot interactive system based on time series prediction according to claim 1 is characterized in that: The data on the status changes of industrial control equipment include changes in register values ​​and changes in coil values.