Detection System and Method for Unauthorized Devices
The system addresses the computational and privacy issues in AMI networks by using Fourier transform and statistical fusion analysis on firewall logs to detect unauthorized devices efficiently and effectively, preserving privacy and improving detection performance.
Patent Information
- Application Number
- KR1020250058345
- Authority / Receiving Office
- KR · KR
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2025-05-02
- Publication Date
- 2026-11-10
AI Technical Summary
Existing intrusion detection methods for Advanced Metering Infrastructure (AMI) networks are computationally intensive and require large data volumes, and deep packet analysis is not feasible due to security policies and privacy concerns, making it difficult to detect unauthorized devices and malicious traffic effectively.
A detection system utilizing an IoT unit, data collection unit, and anomaly detection unit that analyzes firewall logs using Fourier transform and statistical fusion analysis to identify unauthorized equipment and abnormal operations, reducing computational load and data requirements.
The system efficiently detects unauthorized devices with minimal data volume, preserving privacy and being applicable to systems where deep packet analysis is challenging, enhancing detection performance by incorporating security operator experience.
Smart Images

Figure PAT00004_ABST
Abstract
Description
Technology Field
[0001] The present invention relates to a detection system and method for unauthorized devices that perform artificial intelligence-statistical fusion analysis using Fourier transform with characteristics of AMI data traffic as information collected from outside a firewall.
[0002] More specifically, the present invention relates to a method for checking for abnormal operation of a data concentrator to prepare for communication by unauthorized equipment and unknown cyber attacks by analyzing the periodicity of the data communication pattern of the data concentrator when collecting and transmitting data from the data concentrator, extracting and learning frequency components. Background Technology
[0004] Advanced Metering Infrastructure (AMI) is an intelligent power metering system that enables power companies to meter customers' electricity consumption using power communication networks and provide usage information in real time, thereby inducing customers to save electricity and respond to demand on their own.
[0005] The AMI system consists of an electronic power meter, a modem, a data concentrator, a firewall, and a Front-End Processor (FEP), and data collection and communication occur from the outside to the inside in this order. The data concentrator is a device designed to manage various electronic meters and acts as an IoT gateway that communicates with a certain number of electronic meters to collect data.
[0006] In the AMI system, meter measurements are collected by the Data Concentrator Unit (DCU) and transmitted to the Data Collection Server. Although the Data Collection Server is located within KEPCO's internal network, the Data Concentrator is external, making it vulnerable to direct access by malicious attackers. An attacker could install unauthorized equipment on the Data Concentrator and disrupt the AMI network.
[0007] Deep Packet Inspection, a technology utilized in existing intrusion detection methods, is a technique that separates communication traffic packets to examine the payload.
[0008] Figure 1 is a block diagram illustrating the system structure for deep packet analysis.
[0009] Existing deep packet analysis technology requires equipment for switch port mirroring to detect intrusions, and storing the collected data necessitates a larger capacity than the firewall log data proposed subsequently. Furthermore, it cannot be applied to systems where deep packet analysis is not feasible due to security and operational policies. In addition, since power consumption is treated as personal information, privacy infringement prevention technology must be implemented to utilize deep packet analysis. Prior art literature
[0011] Republic of Korea Public Notice No. 10-2015-0004220 The problem to be solved
[0012] The present invention aims to provide a detection system and method for unauthorized devices that can reduce computational load when detecting unauthorized devices or malicious traffic on an AMI network. means of solving the problem
[0014] A detection system for unauthorized devices according to one aspect of the present invention may include: an IoT unit located outside a protected organization (area) and measuring data; a collection unit that collects data generated by the IoT unit and transmits the collected data to the inside of the organization along with a firewall function; and an anomaly detection unit that detects unauthorized equipment and abnormal operation using the collected data, including communication logs passing through the firewall, received from the collection unit.
[0015] Here, it may further include a notification unit that checks whether an anomaly has been detected and transmits it to the security controller.
[0016] Here, the communication log may include two or more of the time the data packet passed through the firewall, the protocol, the number of packet transmissions, the source IP, and the destination port.
[0017] Here, the anomaly detection unit may include: a data preprocessing module that processes collected data to input the data into the learning module below; a learning model that learns the ratio of frequencies as a criterion for labeling (classifying) as abnormal operation as a method for estimating unauthorized equipment of the preprocessed data; and a statistical module that reflects the technical experience of a security controller using statistical techniques on the data output from the learning model.
[0018] Here, the data preprocessing module can extract the arrival time, IP address, destination port, and packet transmission count (noise) from firewall logs collected over a day (necessary for performance improvement when generating test data) and sort them in order of transmission time.
[0019] Here, the data preprocessing module can convert the periodicity of the communication pattern of the corresponding data concentrator into frequency components by performing a Fourier transform of the number of packet transmissions for arrival times in minutes based on a predetermined period (e.g., 1440 minutes) for each IP of the sorted extracted data.
[0020] Here, the data preprocessing module can extract training data using only integer frequency components excluding decimal points based on the number of divisors of the predetermined period in order to remove harmonics generated during the discrete Fourier transform.
[0021] Here, the statistical module can specify a statistical parameter (sigma, threshold value, standard deviation) that serves as a sorting criterion designated by an external security control officer, assign a rank to abnormal operation signals that deviate significantly from the average in terms of the specified statistical parameter, and output signals that are subject to a direct review request in a sorted form according to the assigned rank.
[0023] A method for detecting an unauthorized device according to another aspect of the present invention may include: collecting data transmitted through a firewall into the interior of a protected organization (area) equipped with an external IoT unit; performing a Fourier transform on the collected data to replace the periodicity of the communication pattern with frequency component components; detecting unauthorized equipment and abnormal operation using the Fourier transformed data; and providing the abnormal detection result to an external security controller.
[0024] Here, prior to the step of performing a Fourier transform on the collected data, a data preprocessing step may be performed to extract the arrival time, IP address, destination port, and number of packet transmissions from firewall communication logs collected over a predetermined period and sort them in order of transmission time.
[0025] Here, the communication log may include two or more of the time the packet passed through the firewall, the protocol, the number of packet transmissions, the source IP, and the destination port.
[0026] Here, the step of detecting the unauthorized equipment and abnormal operation can be performed by a pre-machine-learned learning model.
[0027] Herein, a training data preprocessing step for processing data to input collected data into the training of the training model; and a step of training the training model using a frequency ratio as a criterion for labeling (classifying) as abnormal operation as a method for inferring unauthorized equipment of the preprocessed data may further be included.
[0028] Here, a statistical supplementation step may be further included to reflect the technical experience of the security controller using statistical techniques on the data output from the above learning model.
[0029] Here, in the statistical supplementation step, a statistical parameter (sigma, threshold value, standard deviation) that serves as a sorting criterion designated by an external security control operator is designated, and a rank is assigned to abnormal operation signals that deviate significantly from the average in terms of the designated statistical parameter, and signals subject to direct review requests can be output in a sorted form according to the assigned rank.
[0030] Here, in the above training data preprocessing step, in order to remove harmonics generated during the Discrete Fourier Transform, training data can be extracted using only integer frequency components excluding decimal points based on a predetermined number of period factors. Effects of the invention
[0032] When implementing the method for detecting unauthorized devices according to the concept of the present invention with the above-described configuration, there is an advantage in that unauthorized devices can be detected with a small amount of data without infringing on personal information by utilizing firewall logs in the intrusion detection method.
[0033] Specifically, to examine AMI data in detail, the switch port mirroring method requires approximately 7.8 times more data volume and 9.2 times more rows than the firewall log method; therefore, using firewall logs is more efficient in terms of training time. This assumes that a single server is connected to the switch. Since multiple servers are typically connected to a switch, extracting only necessary packets from unnecessary ones results in greater overhead. By utilizing firewall logs for intrusion detection, unauthorized devices can be detected without infringing upon personal information such as power consumption.
[0034] In addition, since AMI data is treated as personal information, using packet deep analysis methods requires technology or policies (such as masking, anonymization, and encryption) to prevent personal information infringement, but using firewall logs is more efficient because it does not require such measures.
[0035] The method for detecting unauthorized devices according to the present invention has the advantage of being easily applicable to systems where deep packet analysis is difficult or systems where operational interruption is difficult, by utilizing firewall logs in the intrusion detection method.
[0036] Specifically, while other intrusion detection methods require the installation of packet mirroring devices, firewalls are easily applicable and portable to any system because they are a fundamental security technology found in all servers.
[0037] The detection method for unauthorized devices according to the present invention has the advantage of enhancing detection performance by utilizing a statistical fusion analysis method and reflecting the experiential skills of a security control operator.
[0038] Specifically, classification methods based on machine learning vary depending on the quality of the data being trained. If communication failures occur frequently and the number of authorized devices that do not exhibit a consistent pattern increases, the ability to detect unauthorized devices may decrease. However, detection performance can be improved because the threshold value for classification can be adjusted based on the experience of security operators by changing the sigma value of statistical techniques.
[0039] The method for detecting unauthorized devices according to the present invention can reduce the amount of computation required to detect unauthorized devices or malicious traffic on an AMI network, and thus has the advantage of being usable for providing unauthorized device detection solutions (facilities) for third-party vendors, microgrids, or customer networks.
[0040] In addition, other features and advantages of the present invention may be newly identified through the embodiments of the present invention. Brief explanation of the drawing
[0042] FIG. 1 is a block diagram illustrating the system structure for deep packet analysis. Figure 2 is a graph illustrating time packets of communication between a data concentrator and a data collection server. FIG. 3 is a graph illustrating the frequency-amplitude obtained by applying a Fourier transform to the time-packet data of FIG. 2. FIG. 4 is a conceptual diagram illustrating the configuration of an unauthorized device detection system and the functional flow of an unauthorized device detection method utilizing the characteristics of AMI data traffic and an artificial intelligence-statistical fusion analysis method according to the concept of the present invention. Figure 5 shows graphs illustrating the operation of a program for supervised learning by data concentrator IP. FIGS. 6 to 9 are graphs illustrating data patterns in cases where it is difficult to determine normal equipment. FIG. 10 is a flowchart illustrating an embodiment of a method for detecting unauthorized devices that performs artificial intelligence-statistical fusion analysis of AMI data traffic characteristics and Fourier transform according to the concept of the present invention. Figure 11 is an example of a comparison of deep packet analysis and firewall log data sizes. Specific details for implementing the invention
[0043] In describing the present invention, terms such as "first," "second," etc., may be used to describe various components, but the components may not be limited by these terms. The terms are intended solely to distinguish one component from another. For example, without departing from the scope of the present invention, the first component may be named the second component, and similarly, the second component may be named the first component.
[0044] When it is mentioned that a component is connected to or coupled with another component, it can be understood that it may be directly connected to or coupled with that other component, or that there may be other components in between.
[0045] The terms used in this specification are used merely to describe specific embodiments and are not intended to limit the invention. Singular expressions may include plural expressions unless the context clearly indicates otherwise.
[0046] In this specification, terms such as "comprising" or "having" are intended to indicate the existence of the features, numbers, steps, actions, components, parts, or combinations thereof described in the specification, and should be understood as not excluding in advance the existence or addition of one or more other features, numbers, steps, actions, components, parts, or combinations thereof.
[0047] In addition, the shapes and sizes of elements in the drawings may be exaggerated for clearer explanation.
[0049] The basic idea of the present invention is to detect intrusion into the internal network of unauthorized equipment by utilizing the communication characteristics (data packet patterns) of a data concentrator in an environment where deep packeting is impossible.
[0050] Due to standards set by Korea Electric Power Corporation (KEPCO), communication in data concentrators exhibits a pattern. To analyze this communication pattern, the periodicity of the pattern is converted into frequency using a Fourier transform, based on the arrival times and packet volumes (based on a predetermined time unit) recorded in the firewall logs.
[0051] Figure 2 is a graph illustrating time-packets of communication between a data concentrator and a data collection server.
[0052] Figure 3 is a graph illustrating the frequency-amplitude obtained by applying a Fourier transform to the time-packet data of Figure 2.
[0054] Figure 2 is an example graph of packet traffic over a 24-hour period of a data concentrator. Traffic higher than the average around midnight is control communication for device operation. Communication is conducted in one-hour intervals.
[0055] Applying a Fourier transform to this yields a frequency-amplitude graph as shown in Fig. 3. Here, for computational convenience, one day was substituted with 1 for the calculation. In other words, if a signal occurs every hour in a 24-hour period, it occurs a total of 24 times, allowing us to identify a frequency that vibrates 24 times during the day. In the graph shown, the frequencies of 24 Hz and its multiples are high; this is because multiples of 24 Hz appear as harmonics when the discrete communication is Fourier transformed. To exclude harmonics, the interval values of the actual data are retained, and frequencies that are not intervals are excluded during subsequent machine learning.
[0057] FIG. 4 is a conceptual diagram illustrating the configuration of an unauthorized device detection system and the functional flow of an unauthorized device detection method utilizing the characteristics of AMI data traffic and an artificial intelligence-statistical fusion analysis method according to the concept of the present invention.
[0058] A detection system for unauthorized devices in a city may include: an IoT unit (100, 200) located outside the organization (area) to be protected and measuring data; a collection unit (300) that collects data generated by the IoT unit (100, 200) and transmits the collected data to the inside of the organization along with a firewall (310) function; an anomaly detection unit (400) that detects unauthorized equipment and abnormal operation using the collected data, including communication logs passing through the firewall (310), received from the collection unit (300); and a notification unit (500) that confirms whether an anomaly has been detected and transmits (outputs) it to an external security control officer.
[0059] The urban IoT unit (100, 200) consists of an electronic power meter (100) belonging to the Advanced Metering Infrastructure (AMI) and a data concentrator unit (DCU) (200) outside the organization that collects sensing signals from IoT sensors equipped in the electronic power meter (100), etc.
[0060] Here, the collection unit (300) stores communication logs passing through a firewall (310) transmitted from the outside to the inside based on the boundary of the organization and transmits them to an anomaly detection unit (400). The firewall log may include the time of passing through the firewall, protocol, number of packet transmissions, source IP, destination port, etc.
[0061] Here, the anomaly detection unit (400) may include: a data preprocessing module (410) that processes collected data to input the data into the learning model (420); a learning model (420) that learns the ratio of frequencies as a criterion for labeling (classifying) as abnormal operation as a method for estimating unauthorized equipment of the preprocessed data; and a statistical module (430) that reflects the technical experience of a security controller using statistical techniques on the data output from the learning model (420).
[0062] Here, the data preprocessing module (410) can extract the arrival time, IP address, destination port, number of packet transmissions, and noise necessary for performance improvement when generating test data from firewall logs collected over a predetermined period of one day, and sort them in order of transmission time to improve processing speed. Here, the above-mentioned values (arrival time, IP address, destination port, number of packet transmissions, noise) of data packets are collected at 1-minute intervals as a predetermined unit period. Here, the noise may be noise resulting from a determination by the firewall.
[0063] The above data preprocessing module (410) is a module that preprocesses data so that firewall logs can utilize Fourier transform formulas. The learning model (420) can classify the presence or absence of normal equipment using a previously learned model. The statistics module (430) reflects the classification results of abnormal operation and unauthorized equipment through the security manager's settings / direct input.
[0064] Here, the Fourier transform is a formula that converts a graph with respect to time into a graph with respect to frequency. The communication of the data concentrator (200) communicates with the data collection server (320) using specific patterns. The data collection server (320) is a component of the AMI system.
[0065] To train the above learning model (420), a detailed communication record (amount of packets over time) for one day is extracted from a data concentrator (200) communicating with a data collection server (320). Frequency components are extracted by performing a Fourier Transform on the pattern periodicity of the communication record, and the learning model (420) determines whether the equipment is normal or unauthorized based on this.
[0067] Figure 5 shows graphs illustrating the operation of a program for supervised learning by data concentrator IP.
[0068] Here, the data preprocessing module can use the sorted extracted data as training and test data by performing a Fourier transform on the number of packet transmissions for arrival times (e.g., applying 1 minute as the unit time) based on 1440 minutes (1 day) for each IP, thereby converting the periodicity of the communication pattern of the corresponding data concentrator into frequency components.
[0069] Here, the frequency components are expressed in minutes and include frequency components corresponding to divisors of 1440 for harmonic elimination.
[0070] Accordingly, in the case of the improved implementation, a method can be applied to extract training data using only integer frequency components excluding decimal points based on the divisors of 1440 in order to remove harmonics generated during the Discrete Fourier Transform.
[0071] For example, to remove harmonics generated during the Discrete Fourier Transform, training data can be extracted using only integer frequency components excluding decimal points based on the divisors of 1440.
[0072] For example, the frequency component of 15 minutes is 96 Hz. Harmonics refer to the continuous occurrence of multiples of 96 Hz. 96 Hz, 192 Hz, 288 Hz... These frequency components represent 15 minutes, 7.5 minutes, 3.75 minutes... respectively in the time-packet volume graph.
[0073] However, since the present invention uses data measured in 1-minute intervals, decimal values cannot appear and must be removed.
[0074] Since a day consists of a total of 1,440 minutes (60 * 24), when frequency values are converted into time components, only the amplitudes of the divisors of 1,440 minutes are recorded. (The wavelength of the shortest signal is 2 minutes, the wavelength of the latest signal is 1,440 minutes; if the wavelength is 1 minute, it is treated as a linear component in the Discrete Fourier Transform.)
[0075] If the method of the present invention is used on a signal with a period of 11 minutes that is not a divisor of 1440, it is difficult to fully extract only the 11th-minute component from the value obtained by Fourier transform due to the harmonics of the 11th-minute frequency component.
[0077] Here, in order to train the learning model (420), the learning module not shown can label the data using automatic labeling technology based on the specification information of the equipment determined by the organization.
[0078] Here, the above automatic labeling technology extracts a sample from a data population and classifies data in which the frequency component corresponding to the specifications of the equipment exceeds a threshold value set by the security manager as authorized equipment with a normal pattern.
[0079] Here, the threshold value mentioned above can be set to a value similar to the result directly classified by the security controller and the result predicted by changing the threshold value of the frequency component corresponding to the equipment specifications for the extracted sample.
[0080] However, due to communication and device failures of the data concentrator (200), normal equipment may exhibit abnormal operation. Cases where it is difficult to determine normal equipment using Fourier transform are as follows.
[0081] 1) The communication pattern cycle differs depending on the type of data concentrator.
[0082] 2) There are cases where no communication packets are transmitted for an arbitrary period.
[0083] 3) There are cases where communication takes a certain amount of time due to transmission delay.
[0084] 4) The amount of packets transmitted varies by equipment, and there are cases where they have complex cycles.
[0085] 5) It includes not only power consumption but also communication for equipment control.
[0086] Therefore, in the machine learning method, the above learning model (420) is trained using frequency values extracted by Fourier transform to classify normal equipment and non-normal equipment.
[0087] By adding statistical techniques instead of relying solely on machine learning, it enables the application of security operators' experience to the system, thereby allowing for the adjustment of distinction criteria between abnormal behavior and unauthorized equipment.
[0088] Here, the statistical module (430) is used to extract data that deviates significantly from the pattern using the standard deviation of the data of unauthorized and abnormal operation equipment detected by the learning model (420), and can be used to reflect the experiential skills of the security manager by adjusting the sigma value.
[0090] Next, I will explain the cases (and measures) where it is difficult to determine the aforementioned normal equipment, referring to the example graph.
[0091] Figures 6 to 9 are graphs illustrating data patterns in cases where it is difficult to determine normal equipment.
[0092] In the case of Fig. 6, the communication pattern period differs for each data concentrator.
[0093] In the case of Fig. 7, there are instances where there is no transmission of communication packets for an arbitrary period due to a communication failure.
[0094] The graph in Fig. 8 shows a case where communication takes place over a certain period due to transmission delay, and a case where the amount of packets transmitted varies by equipment and has a complex cycle.
[0095] Figure 9 shows not only power consumption but also communication for controlling equipment at any given time.
[0097] Next, the significance and function of the above statistics module (430) will be explained with specific examples.
[0098] When an unauthorized device detection system is established according to the concept of the present invention, data deemed as abnormal operation due to signal status or field operations may increase. Although this is not a practical problem, since the training module determines it as abnormal operation because it has exceeded the threshold value set during learning, it may lead to a situation where security controllers must directly verify more data than usual.
[0099] If an experienced security operator is aware of the current environment and can adjust threshold values, the workload will be reduced unlike the above; the statistics module is a mechanism designed to incorporate the security operator's experience, taking these circumstances into account.
[0100] For example, normally, the "Warning Notification Department" requests the security control operator to review about 20 signals directly, but on one day, if they request the direct review of 100 data points, the experienced security control operator uses the "Statistical Module" to adjust the sigma, threshold value, and standard deviation to rank abnormal operation signals that deviate significantly from the average.
[0101] If the security operator first reviews 20 to 30 results that deviate significantly from the average and finds no major abnormalities, the remaining 70 to 80 are judged to be normal; otherwise, if all 20 to 30 are abnormal, it will be necessary to continue reviewing them in the order suggested by the "Statistics Module".
[0102] In other words, by using a statistics module to assign review priorities to signals subject to direct review requests by security monitors, the actual direct review work performed by security monitors can be reduced.
[0103] When a security controller specifies statistical parameters (sigma, threshold value, standard deviation, etc.) that serve as sorting criteria, the statistical module assigns ranks to abnormal operation signals that deviate significantly from the average based on the specified statistical parameters, and outputs signals subject to direct review requests in a sorted form according to the assigned ranks.
[0105] FIG. 10 is a flowchart illustrating an embodiment of a method for detecting unauthorized devices that performs artificial intelligence-statistical fusion analysis of AMI data traffic characteristics and Fourier transform according to the concept of the present invention.
[0106] A method for detecting unauthorized devices in a city may include the steps of: collecting data transmitted through a firewall into the interior of a protected organization (area) equipped with an IoT unit externally (S110); performing a Fourier transform on the collected data to convert the periodicity of the communication pattern into frequency components (S130); detecting unauthorized equipment and abnormal operation using the Fourier transformed data (S140); and providing the abnormal detection results to an external security controller (S150).
[0107] A method for detecting an unauthorized device in a city may perform a data preprocessing step (120) prior to the step (S130) of performing a Fourier transform on the collected data, in which the arrival time, IP address, destination port, and number of packet transmissions are extracted from firewall communication logs collected over a predetermined period and sorted in order of transmission time.
[0108] Here, the communication log may include two or more of the time the packet passed through the firewall, the protocol, the number of packet transmissions, the source IP, and the destination port.
[0109] Depending on the implementation, the step (S140) of detecting unauthorized equipment and abnormal operation may be performed by the learning module described above in accordance with the concept of the present invention, and in this case, to train the learning model, the method may further include a learning data preprocessing step (substantially identical to step S120) for processing data to input collected data into the learning of the learning model; and a step (S60) of training the learning model using the ratio of frequencies as a criterion for labeling (classifying) as abnormal operation as a method for estimating unauthorized equipment of the preprocessed data.
[0110] Additionally, a statistical supplementation step (S70) may be further included to reflect the technical experience of the security controller using statistical techniques on the data output from the above learning module.
[0111] In the above statistical supplementation step (S70), as previously explained, a statistical parameter (sigma, threshold value, standard deviation) that serves as a sorting criterion designated by an external security control officer is designated, and a rank is assigned to abnormal operation signals that deviate significantly from the average in terms of the designated statistical parameter, and signals subject to direct review request can be output in a sorted form according to the assigned rank.
[0112] Depending on the implementation, in the training data preprocessing step, in order to remove harmonics generated during the Discrete Fourier Transform, training data can be extracted using only integer frequency components excluding decimal points based on the divisors of the number of periods (e.g., 1440).
[0114] Figure 11 is an example of a comparison of deep packet analysis and firewall log data sizes.
[0115] Utilizing firewall logs for intrusion detection allows for the detection of unauthorized devices with minimal data volume and without compromising personal privacy. Specifically, to examine AMI data in detail, the switch port mirroring method requires approximately 7.8 times more data volume and 9.2 times more rows compared to the firewall log method; therefore, using firewall logs is more efficient in terms of training time. This assumes that a single server is connected to the switch. Since multiple servers are typically connected to a switch, extracting only necessary packets from unnecessary ones results in greater overhead.
[0116] By utilizing firewall logs for intrusion detection, unauthorized devices can be detected without infringing upon personal information, such as power consumption. Specifically, since AMI data is treated as personal information, utilizing deep packet analysis methods requires technologies or policies (such as masking, anonymization, and encryption) to prevent personal information infringement; however, the method using firewall logs is more efficient because it does not require such measures.
[0117] Utilizing firewall logs for intrusion detection allows for easy application even to systems where deep packet analysis is difficult or where operational downtime is challenging. Unlike other intrusion detection methods that require the installation of packet mirroring devices, firewalls are a fundamental security technology found in all servers, making them easy to apply and portable to any system.
[0118] By utilizing statistical fusion analysis methods, detection performance can be enhanced by reflecting the experiential skills of security operators. Specifically, classification methods based on machine learning vary depending on the quality of the data being trained. If communication failures occur frequently and the number of authorized devices exhibiting irregular patterns increases, the ability to detect unauthorized devices may decrease. However, detection performance can be improved because the threshold for classification can be adjusted based on the experience of security operators by modifying the sigma value of statistical techniques.
[0120] Those skilled in the art to which the present invention pertains should understand that the present invention may be implemented in other specific forms without altering its technical concept or essential features, and therefore the embodiments described above should be understood as illustrative in all respects and not restrictive. The scope of the present invention is defined by the claims set forth below rather than by the detailed description, and all modifications or variations derived from the meaning and scope of the claims and equivalent concepts thereof should be interpreted as being included within the scope of the present invention. Explanation of the symbols
[0122] 100 : Electronic power meter 200: Data Concentrator 300 : Collection Department 310 : Firewall 400 : Anomaly Detection Unit 410 : Data Preprocessing Module 420 : Learning Model 430 : Statistics Module
Claims
Claim 1 An unauthorized device detection system comprising: an IoT unit located outside the organization to be protected and measuring data; a collection unit that collects data generated by the IoT unit and transmits the collected data to the inside of the organization along with a firewall function; and an anomaly detection unit that detects unauthorized equipment and abnormal operation using the collected data, including communication logs passing through the firewall received from the collection unit. Claim 2 A detection system for unauthorized devices according to claim 1, further comprising a notification unit that checks whether an anomaly has been detected and transmits it to a security controller. Claim 3 In claim 1, the communication log is a detection system for unauthorized devices comprising two or more of the time a data packet passed through the firewall, a protocol, the number of packet transmissions, a source IP, and a destination port. Claim 4 In claim 1, the anomaly detection unit comprises: a data preprocessing module that processes collected data to input data into a learning module; a learning model that learns the ratio of frequencies as a criterion for labeling as abnormal operation as a method for estimating unauthorized equipment of the preprocessed data; and a statistical module that reflects the technical experience of a security controller using statistical techniques on the data output from the learning model, thereby establishing a detection system for unauthorized devices. Claim 5 In paragraph 4, the data preprocessing module is a detection system for unauthorized devices that extracts arrival time, IP address, destination port, and number of packet transmissions from firewall logs collected over a day and sorts them in order of transmission time. Claim 6 In paragraph 4, the data preprocessing module is a detection system for unauthorized devices that performs a Fourier transform of the number of packet transmissions for arrival times in minutes based on a predetermined period for each IP, based on sorted extracted data. Claim 7 In claim 6, the data preprocessing module is a detection system for an unauthorized device that extracts training data using only integer frequency components excluding decimal points based on the factors of the number of predetermined periods in order to remove harmonics generated during the Discrete Fourier Transform. Claim 8 In paragraph 4, the statistical module is a detection system for unauthorized devices that specifies a statistical parameter (one of sigma, threshold value, and standard deviation) serving as a sorting criterion designated by an external security controller, assigns a rank to abnormal operation signals that deviate significantly from the average in terms of the specified statistical parameter, and outputs signals subject to a direct review request in a sorted form according to the assigned rank. Claim 9 A method for detecting unauthorized devices, comprising: a step of collecting data transmitted through a firewall into the interior of a protected organization equipped with an external IoT unit; a step of performing a Fourier transform on the collected data to replace the periodicity of the communication pattern with frequency components; a step of detecting unauthorized equipment and abnormal operation using the Fourier transformed data; and a step of providing the abnormality detection results to an external security controller. Claim 10 A method for detecting an unauthorized device according to claim 9, wherein, prior to the step of performing a Fourier transform on the collected data, a data preprocessing step is performed to extract the arrival time, IP address, destination port, and number of packet transmissions from firewall communication logs collected over a predetermined period and sort them in order of transmission time. Claim 11 In claim 9, the communication log comprises two or more of the time a packet passed through the firewall, protocol, number of packet transmissions, source IP, and destination port, a method for detecting an unauthorized device. Claim 12 In claim 9, the step of detecting unauthorized equipment and abnormal operation is a method for detecting unauthorized devices performed by a pre-machine-learned learning model. Claim 13 A method for detecting unauthorized devices according to claim 12, further comprising: a training data preprocessing step for processing data to input collected data into the training of the training model; and a step of training the training model using a frequency ratio as a criterion for labeling as abnormal operation as a method for inferring unauthorized devices from the preprocessed data. Claim 14 A method for detecting unauthorized devices according to claim 13, further comprising a statistical supplementation step that incorporates the technical experience of a security controller into the data output from the learning model using statistical techniques. Claim 15 In claim 14, a method for detecting unauthorized devices, wherein in the statistical supplementation step, a statistical parameter (one of sigma, threshold value, and standard deviation) serving as a sorting criterion designated by an external security controller is designated, a rank is assigned to abnormal operation signals that deviate significantly from the average in terms of the designated statistical parameter, and signals subject to direct review requests are output in a sorted form according to the assigned rank. Claim 16 In claim 13, the method for detecting an unauthorized device involves, in the above-mentioned training data preprocessing step, extracting training data using only integer frequency components excluding decimal points based on a predetermined number of period factors in order to remove harmonics generated during the Discrete Fourier Transform.
