WO2019244902A1 - 評価装置及び評価方法 - Google Patents
評価装置及び評価方法 Download PDFInfo
- Publication number
- WO2019244902A1 WO2019244902A1 PCT/JP2019/024167 JP2019024167W WO2019244902A1 WO 2019244902 A1 WO2019244902 A1 WO 2019244902A1 JP 2019024167 W JP2019024167 W JP 2019024167W WO 2019244902 A1 WO2019244902 A1 WO 2019244902A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- data
- evaluation
- model
- probability density
- vae
- Prior art date
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N5/00—Computing arrangements using knowledge-based models
- G06N5/04—Inference or reasoning models
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/047—Probabilistic or stochastic networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/0895—Weakly supervised learning, e.g. semi-supervised or self-supervised learning
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L41/00—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks
- H04L41/14—Network analysis or design
- H04L41/145—Network analysis or design involving simulating, designing, planning or modelling of a network
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L43/00—Arrangements for monitoring or testing data switching networks
- H04L43/08—Monitoring or testing based on specific metrics, e.g. QoS, energy consumption or environmental parameters
- H04L43/0823—Errors, e.g. transmission errors
Definitions
- the present invention relates to an evaluation device and an evaluation method.
- IoT devices With the arrival of the Internet of Things (IoT) era, various devices (IoT devices) have been connected to the Internet and are being used in various ways. Accordingly, security measures for IoT devices such as a traffic session abnormality detection system and an intrusion detection system (IDS: Intrusion Detection System) for IoT devices are expected.
- IDS Intrusion Detection System
- a technique using a probability density estimator based on unsupervised learning such as VAE (Variational Auto Encoder).
- VAE Variational Auto Encoder
- this technique after learning the probability density of normal communication data, communication with a low probability density is detected as abnormal. For this reason, in this technique, only normal communication data needs to be known, and abnormality detection can be performed without learning all malicious data. Therefore, this technology is effective for detecting threats to IoT devices that are still in a transition period and do not know all threat information.
- erroneous detection includes overdetection in which normal communication is erroneously determined to be abnormal.
- Data that can be overdetected include maintenance communications that occur only several times a year, and abnormal amounts of traffic data during the Olympics.
- a function for improving the detection accuracy by feeding back the overdetection data when the occurrence of overdetection is noticed is required.
- the conventional method has the following two problems.
- First as a first problem, there is a problem that the initial learning data set used for the initial learning needs to be saved after the model is generated.
- As a second problem when the overdetection data set is extremely small compared to the initial learning data set, there is a problem that overdetection data cannot be learned with high accuracy.
- the present invention has been made in view of the above, and an object of the present invention is to provide an evaluation device and an evaluation method for performing highly accurate evaluation of the presence / absence of communication data.
- an evaluation device includes a receiving unit that receives an input of communication data to be evaluated, and a first unit that learns a characteristic of a probability density of normal initial learning data. And the second model that has learned the characteristics of the probability density of the normal over-detection data detected as abnormal in the course of the evaluation process, and estimates the probability density of the communication data to be evaluated.
- the evaluation of the presence / absence of abnormality in communication data is executed with high accuracy.
- FIG. 1 is a diagram illustrating an example of a configuration of an evaluation device according to an embodiment.
- FIG. 2 is a diagram illustrating a process of the model generation unit illustrated in FIG.
- FIG. 3 is a diagram illustrating a process of the model generation unit illustrated in FIG.
- FIG. 4 is a diagram illustrating feedback learning in the evaluation device shown in FIG.
- FIG. 5 is a diagram illustrating a model generated by the model generation unit illustrated in FIG.
- FIG. 6 is a diagram illustrating a model generated by the model generation unit illustrated in FIG.
- FIG. 7 is a diagram illustrating a process of the evaluation unit illustrated in FIG. 1.
- FIG. 8 is a flowchart illustrating a processing procedure of a learning process performed by the evaluation device illustrated in FIG. 1 in an initial stage.
- FIG. 8 is a flowchart illustrating a processing procedure of a learning process performed by the evaluation device illustrated in FIG. 1 in an initial stage.
- FIG. 8 is a flowchart illustrating a processing procedure of
- FIG. 9 is a flowchart illustrating a processing procedure of an evaluation process performed by the evaluation device 1 illustrated in FIG.
- FIG. 10 is a diagram illustrating an application example of the evaluation device according to the embodiment.
- FIG. 11 is a diagram illustrating another example of the processing of the evaluation unit illustrated in FIG. 1.
- FIG. 12 is a diagram illustrating feedback learning of a conventional evaluation method.
- FIG. 13 is a diagram illustrating a model used in a conventional evaluation method.
- FIG. 14 is a diagram illustrating a model used in a conventional evaluation method.
- FIG. 15 is a diagram illustrating an example of a computer on which an evaluation device is realized by executing a program.
- the evaluation device generates an over-detection VAE model in which only the over-detection data is learned, in addition to the learning data VAE model in which normal learning data is learned.
- Over-detection data is normal communication data evaluated as abnormal in the course of the evaluation process, and occurs only in a small amount. Since the evaluation device according to the present embodiment performs evaluation based on the probability density obtained by combining the two generated VAE models at the model level, it realizes feedback of over-detection data and high accuracy of detection. I do.
- VAE accepts an input of a data point x i, and outputs the anomaly score corresponding to the data (score) (degree of abnormality). Assuming that the estimated value of the probability density is p (x i ), the anomaly score is an approximate value of ⁇ logp (x i ). Therefore, the higher the value of the anomaly score output by the VAE, the higher the degree of abnormality of the communication data.
- FIG. 1 is a diagram illustrating an example of a configuration of an evaluation device according to an embodiment. As shown in FIG. 1, the evaluation device 1 has a communication unit 10, a storage unit 11, and a control unit 12.
- the communication unit 10 is a communication interface for transmitting and receiving various information to and from other devices connected via a network or the like.
- the communication unit 10 is realized by a NIC (Network Interface Card) or the like, and performs communication between another device and a control unit 12 (described later) via an electric communication line such as a LAN (Local Area Network) or the Internet.
- the communication unit 10 is connected to, for example, an external device via a network or the like, and receives input of communication data to be evaluated.
- the storage unit 11 is realized by a semiconductor memory element such as a RAM (Random Access Memory) or a flash memory (Flash Memory), or a storage device such as a hard disk or an optical disk, and stores a processing program for operating the evaluation device 1 and a processing program for the processing program. Data and the like used during execution are stored.
- the storage unit 11 has a VAE model 111 for learning data and a VAE model 112 for over-detection.
- the learning data VAE model 111 is a learning data VAE model (first model) that has learned normal learning data, and is a model that has learned the characteristics of the probability density of normal initial learning data.
- the over-detection VAE model 112 is an over-detection VAE model (second model) that has learned only over-detection data.
- the over-detection VAE model 112 has a characteristic of the probability density of normal over-detection data evaluated as abnormal during the evaluation process. It is a learned model. Each model has model parameters of the trained VAE.
- the control unit 12 has an internal memory for storing programs and required data defining various processing procedures and the like, and executes various processes by using these.
- the control unit 12 is an electronic circuit such as a CPU (Central Processing Unit) or an MPU (Micro Processing Unit).
- the control unit 12 includes a reception unit 120, a model generation unit 121 (generation unit), and an evaluation unit 123.
- the model generation unit 121 has the VAE 122 as a probability density estimator, learns input data, and generates a VAE model or updates VAE model parameters.
- the model generation unit 121 stores the model parameters of the generated VAE model or the updated VAE model in the storage unit 11.
- FIGS. 2 and 3 are views for explaining the processing of the model generation unit 121 shown in FIG.
- the model generation unit 121 learns a large amount of normal learning data Ds (for example, HTTP communication) as initial learning data, and generates a VAE model 111 for learning data. .
- Ds normal learning data
- VAE model 111 VAE model 111 for learning data.
- the model generation unit 121 learns a small amount of over-detection data De (for example, FTP communication) collected in the course of the evaluation process, and newly generates the over-detection VAE model 112. I do.
- the model generation unit 121 learns the over-detection data fed back.
- the model generation unit 121 learns the input over-detection data and generates the over-detection VAE model 112 or updates the parameters of the over-detection VAE model 112. Thereby, the over-detection data is fed back to the evaluation device 1.
- FIG. 4 is a diagram illustrating feedback learning in the evaluation device shown in FIG.
- FIGS. 5 and 6 are diagrams illustrating the model generated by the model generation unit 121 shown in FIG.
- the model generation unit 121 uses the number of initial learning data Ds and a small amount of over-detection data De fed back to perform over-detection.
- the detection data De is learned with high accuracy.
- the model generation unit 121 generates the over-detection VAE model 112 or updates the model parameters of the over-detection VAE model 112.
- the evaluation device 1 needs to store only the number of initial learning data Ds for feedback learning of overdetection data. Further, since the evaluation device 1 learns only a small amount of over-detection data, the learning time can be shorter than learning a large amount of initial learning data. Further, since the evaluation device 1 learns only the over-detection data, it can execute the learning with high accuracy.
- the learning data VAE model 111 is obtained by accurately learning normal learning data in the initial stage (see (1a) in FIG. 4), and has been created from the initial learning data Ds in the past. (See (1b) in FIG. 4).
- the VAE model for learning data 111 shows a low anomaly score with respect to normal communication data in normal time (see FIG. 5).
- the over-detection VAE model 112 is obtained by learning the over-detection data with high accuracy, and shows a low anomaly score with respect to the over-detection data (see FIG. 6).
- the evaluation unit 123 estimates the probability density of the communication data to be evaluated using the VAE model 111 for learning data and the VAE model 112 for overdetection, and determines whether there is any abnormality in the communication data to be evaluated based on the estimated probability density. To evaluate. The evaluation unit 123 evaluates the communication data to be evaluated based on a probability density obtained by combining the probability density estimated by applying the VAE model 111 for learning data and the probability density estimated by applying the VAE model 112 for overdetection. Evaluate for abnormalities. When the combined probability density is lower than the predetermined value, the evaluation unit 123 detects that the communication data to be evaluated is abnormal, and notifies an external coping device or the like of the occurrence of the communication data abnormality. The evaluation unit 123 has a combining unit 124 and an abnormality presence / absence evaluation unit 126.
- the combining unit 124 includes, for example, a first VAE 1251 to which the model parameters of the learning data VAE model 111 are applied, and a second VAE 1252 to which the model parameters of the VAE model 112 for over-detection are applied.
- the combining unit 124 combines the probability density estimated by applying the VAE model 111 for learning data and the probability density estimated by applying the VAE model 112 for overdetection.
- the combining unit 124 When the overdetection VAE model 112 is generated or updated by feedback of the overdetection data, the combining unit 124 combines the overdetection VAE model 112 and the learning data VAE model 111 at a model level.
- the combination at the model level indicates that the scores, which are the outputs of the respective VAE models, are combined based on the following equation (1).
- the combining unit 124 calculates the anomaly score estimated by the first VAE 1251 to which the learning data VAE model 111 has been applied and the anomaly score estimated by the second VAE 1252 to which the over-detection VAE model 112 has been applied, by the following equation (1). To calculate the combined anomaly score.
- score n is an anomaly score output by the first VAE 1251 to which the VAE model 111 for learning data obtained by learning the initial learning data Ds is applied.
- the score od is an anomaly score output by the second VAE 1252 to which the VAE model 112 for over detection that has learned the over-detection data De is applied.
- the score concat is the binding anomaly score.
- N n is the number of learning data.
- N od is the number of over-detection data.
- the abnormality presence / absence evaluation unit 126 evaluates the presence / absence of abnormality in the communication data to be evaluated based on the probability density combined by the combining unit 124.
- the abnormality presence / absence evaluation unit 126 detects the presence / absence of abnormality in the communication data to be evaluated based on the combined anomaly score calculated by the combining unit 124. Specifically, the abnormality presence / absence evaluating unit 126 evaluates that the communication data to be evaluated is abnormal when the combined anomaly score is higher than a predetermined value. When the combined anomaly score is equal to or smaller than a predetermined value, the abnormality presence / absence evaluation unit 126 evaluates that the communication data to be evaluated is normal.
- FIG. 7 is a diagram for explaining the processing of the evaluation unit 123 shown in FIG.
- the evaluation unit 123 inputs the learned learning data VAE model 111 and the over-detection VAE model 112 (see arrows Y1 and Y2), and converts the evaluation communication data (evaluation data) Dt obtained from the network. evaluate.
- the evaluation unit 123 applies the anomaly score output by the first VAE 1251 to the evaluation data Dt and the anomaly score output by the second VAE 1252 to the evaluation data Dt in Expression (1), thereby obtaining the combined anomaly. Get score.
- the evaluation unit 123 evaluates that the communication data to be evaluated is abnormal, and outputs the evaluation result Dr to the coping device or the like.
- FIG. 8 is a flowchart illustrating a processing procedure of a learning process performed by the evaluation device 1 illustrated in FIG. 1 in an initial stage.
- the model generation unit 121 when the model generation unit 121 receives an instruction to generate the VAE model 111 for learning data, which is an initial model, in an initial stage (step S1), the model generation unit 121 receives input of initial learning data (step S2). . Then, the model generation unit 121 learns the initial learning data, and generates the learning data VAE model 111 (step S3). The model generation unit 121 stores the generated model parameters of the learning data VAE model 111 in the storage unit 11.
- FIG. 9 is a flowchart illustrating a processing procedure of an evaluation process performed by the evaluation device 1 illustrated in FIG.
- the evaluating unit 123 estimates the probability density of the evaluation target data by applying the learned model (Step S12). (Step S13).
- the storage unit 11 stores only the learning data VAE model 111.
- the evaluation unit 123 estimates the probability density of the evaluation data by applying the learning data VAE model 111 to the first VAE. If the overdetection data has been fed back, the storage unit 11 stores both the learning data VAE model 111 and the overdetection VAE model 112. In this case, the evaluation unit 123 applies the learning data VAE model 111 to the first VAE 1251, applies the over-detection VAE model 112 to the second VAE 1252, and estimates the probability density of the evaluation data in each VAE.
- the evaluation unit 123 calculates a probability density obtained by combining the probability density estimated by applying the VAE model 111 for learning data and the probability density estimated by applying the VAE model 112 for overdetection (step S14). ). Specifically, in the evaluation unit 123, the combining unit 124 sets the anomaly score estimated by the first VAE 1251 using the VAE model 111 for learning data and the anomaly score estimated by the second VAE 1252 using the VAE model 112 for overdetection. Is applied to equation (1) to calculate the combined anomaly score.
- the abnormality presence / absence evaluation unit 126 evaluates the presence / absence of abnormality in the communication data to be evaluated based on the probability density calculated in step S14, and outputs an evaluation result (step S15).
- the abnormality presence / absence evaluating unit 126 evaluates that the communication data to be evaluated is abnormal when the combined anomaly score calculated by the combining unit 124 is higher than a predetermined value.
- the control unit 12 determines whether or not an instruction to learn overdetection data has been received (step S16). For example, the administrator analyzes the detection result output from the evaluation unit 123, and if there is communication data that is detected as abnormal but is actually normal, classifies the communication data as overdetection data. I do. Then, when a predetermined number of overdetection data is collected, the administrator feeds back the collected overdetection data to the evaluation device 1 and instructs the evaluation device 1 to learn the overdetection data.
- the detection result output from the evaluation unit 123 is analyzed, and when a predetermined number of communication data classified as overdetection data is accumulated, the overdetection data to be learned is fed back from the external device. At the same time, a learning instruction for the over-detection data is input.
- the reception unit 120 receives input of overdetection data to be learned (step S17). Subsequently, the model generation unit 121 learns the input overdetection data, and newly generates the overdetection VAE model 112 (Step S18). Alternatively, the model generation unit 121 learns the over-detection data fed back and updates the model parameters of the VAE model 112 for over-detection (step S18).
- control unit 12 determines whether an instruction for ending the evaluation processing has been received (step S19) ). If the control unit 12 determines that the instruction to terminate the evaluation processing has not been received (step S19: No), the process returns to step S11 and accepts the input of the next evaluation data. When determining that the control unit 12 has received the end instruction of the evaluation process (Step S19: Yes), the control unit 12 ends the evaluation process.
- the evaluation device 1 according to the present embodiment can be applied to abnormality detection of an IoT device.
- FIG. 10 is a diagram illustrating an application example of the evaluation device 1 according to the embodiment. As shown in FIG. 10, an evaluation device 1 is provided on a network 3 to which a plurality of IoT devices 2 are connected. In this case, the evaluation device 1 collects the traffic session information transmitted and received by the IoT device 2, learns the probability density of a normal traffic session, and detects an abnormal traffic session.
- the model generation unit 121 receives the initial learning data set and the over-detection data set to be learned, and stores the learned model in which the received data set has been learned in the storage unit 11.
- FIG. 11 is a diagram illustrating another example of the processing of the evaluation unit 123 illustrated in FIG.
- the combining unit 124 receives the model parameters of one or more trained models, and combines the anomaly scores estimated by each VAE to which each trained model is applied.
- the VAE of the combining unit 124 has a function of outputting an estimation result for each input evaluation data.
- FIG. 1 illustrates an example in which the coupling unit 124 includes two VAEs, the configuration is not limited thereto.
- the coupling unit 124 may have a configuration having the same number of VAEs as the number of models to be applied.
- the combining unit 124 may sequentially apply the learned models to one VAE and acquire each anomaly score estimated using each learned model.
- the learned model applied to the combining unit 124 may be the learning data VAE model 111 that has learned the initial learning data or the overdetection VAE model 112 that has learned the overdetection data. Further, a plurality of learning data VAE models 111-1 and 111-2 obtained by learning different learning data may be applied to the combining unit 124 (see arrow Y11). Of course, only one VAE model for learning data may be applied to the combining unit 124.
- the combining unit 124 may employ a plurality of over-detection VAE models 112-1 and 112-2 that have learned different over-detection data (see arrow Y12).
- VAE models 112-1 and 112-2 that have learned different over-detection data (see arrow Y12).
- the VAE model for over-detection has not been generated, so that the VAE model for over-detection need not be applied to the combining unit 124.
- only one VAE model for over-detection may be applied to the combining unit 124.
- the combining unit 124 When a plurality of models are applied, the combining unit 124 combines the anomaly scores of the applied models based on the following equation (2).
- score k is the score output by the k-th model
- N k is the number of data learned by the k-th model.
- the combining unit 124 can combine two or more models at the model level.
- the evaluation device 1 inputs the initial learning data to the model generation unit 121 and obtains the learning data VAE model 111. Then, in the course of the evaluation process, the evaluation device 1 inputs only the learning data VAE model 111 to the combining unit 124 and sequentially evaluates traffic information obtained from the network until some overdetection is detected. Continue.
- the evaluation apparatus 1 inputs a data set of overdetection data to the model generation unit 121, and generates an overdetection VAE model 112 that has learned overdetection data. After that, the evaluation device 1 inputs the learning data VAE model 111 and the over-detection VAE model 112 to the combining unit 124, and continuously evaluates the traffic information similarly obtained from the network.
- the evaluation device 1 continuously improves the detection accuracy by sequentially repeating the processes of overdetection detection, overdetection data learning, and model combination.
- FIG. 12 is a diagram illustrating feedback learning of a conventional evaluation method.
- FIG. 13 and FIG. 14 are diagrams illustrating a model used in a conventional evaluation method.
- the conventional VAE model shows a low anomaly score for communication data corresponding to a large amount of learning data at the time of evaluation (see FIG. 13), but still shows a high anomaly score for overdetected data. (See FIG. 14).
- the conventional evaluation method since the number of data is uneven, the overdetected data cannot be learned with high accuracy. Further, in the conventional evaluation method, it is necessary to store a large amount of initial learning data for feedback learning of over-detection data. In addition, since a new VAE model is generated again, it takes more time than the initial learning time. Was needed.
- the learning data is camera communication (369 data)
- the over-detection data is SSH communication (10 data).
- the average score of the learning data is -16.3808.
- the average score of the over-detection data is slightly improved compared to before the over-detection data feedback, but still shows a high score of 44.6441, and the accuracy remains low.
- the time required for the re-learning is 14.157 (sec), which is longer than that at the time of the initial learning.
- the average score of the learning data is -25.2625.
- the average score of the over-detection data is significantly improved to -24.0182 as compared with the conventional evaluation method.
- the time required for re-learning is greatly reduced to 3.937 (sec) as compared with the conventional evaluation method.
- the probability density of the evaluation data is estimated using the learning data VAE model that has learned the normal learning data and the overdetection VAE model that has learned the overdetection data.
- the presence or absence of an abnormality in the evaluation data is evaluated based on the probability density obtained. That is, in the present embodiment, an overdetection VAE model in which only overdetection data is feedback-learned is generated separately from the learning data VAE model in which normal learning data is learned, and the two generated VAE models are estimated. Evaluation is performed based on the probability density obtained by combining the obtained probability densities.
- over-detection data cannot be learned with high accuracy, and a large amount of initial learning data needs to be stored for feedback learning of over-detection data.
- More time was required than at the time of the initial learning.
- the evaluation device 1 in the evaluation device 1 according to the present embodiment, only the number of initial learning data Ds need be stored for feedback learning of overdetection data. In the evaluation process, the evaluation device 1 needs to learn only a small amount of over-detection data in the course of the evaluation process, as shown in the above-described evaluation experiment result. Can be significantly shortened. Further, as shown in the above-described evaluation experiment results, the evaluation device 1 can evaluate the overdetection data with high accuracy even if there is a deviation in the number between the overdetection data and the learning data.
- each component of each device illustrated is a functional concept and does not necessarily need to be physically configured as illustrated.
- the specific form of distribution / integration of each device is not limited to that shown in the figure, and all or a part thereof may be functionally or physically distributed / arbitrarily divided into arbitrary units according to various loads and usage conditions. Can be integrated and configured.
- all or any part of each processing function performed by each device can be realized by a CPU and a program analyzed and executed by the CPU, or can be realized as hardware by wired logic.
- FIG. 15 is a diagram illustrating an example of a computer on which the evaluation device 1 is realized by executing a program.
- the computer 1000 has, for example, a memory 1010 and a CPU 1020. Further, the computer 1000 has a hard disk drive interface 1030, a disk drive interface 1040, a serial port interface 1050, a video adapter 1060, and a network interface 1070. These units are connected by a bus 1080.
- the memory 1010 includes a ROM (Read Only Memory) 1011 and a RAM 1012.
- the ROM 1011 stores, for example, a boot program such as a BIOS (Basic Input Output System).
- BIOS Basic Input Output System
- the hard disk drive interface 1030 is connected to the hard disk drive 1090.
- the disk drive interface 1040 is connected to the disk drive 1100.
- a removable storage medium such as a magnetic disk or an optical disk is inserted into the disk drive 1100.
- the serial port interface 1050 is connected to, for example, a mouse 1110 and a keyboard 1120.
- the video adapter 1060 is connected to the display 1130, for example.
- the hard disk drive 1090 stores, for example, an operating system (OS) 1091, an application program 1092, a program module 1093, and program data 1094. That is, a program that defines each process of the evaluation device 1 is implemented as a program module 1093 in which codes executable by a computer are described.
- the program module 1093 is stored in, for example, the hard disk drive 1090.
- a program module 1093 for executing the same processing as the functional configuration in the evaluation device 1 is stored in the hard disk drive 1090.
- the hard disk drive 1090 may be replaced by an SSD (Solid State Drive).
- the setting data used in the processing of the above-described embodiment is stored as the program data 1094 in the memory 1010 or the hard disk drive 1090, for example. Then, the CPU 1020 reads out the program module 1093 and the program data 1094 stored in the memory 1010 and the hard disk drive 1090 to the RAM 1012 as needed, and executes them.
- the program module 1093 and the program data 1094 are not limited to being stored in the hard disk drive 1090, but may be stored in, for example, a removable storage medium and read out by the CPU 1020 via the disk drive 1100 or the like. Alternatively, the program module 1093 and the program data 1094 may be stored in another computer connected via a network (LAN, Wide Area Network (WAN), or the like). Then, the program module 1093 and the program data 1094 may be read from another computer by the CPU 1020 via the network interface 1070.
- LAN Local Area Network
- WAN Wide Area Network
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Software Systems (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- Computational Linguistics (AREA)
- Biomedical Technology (AREA)
- General Health & Medical Sciences (AREA)
- Biophysics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computer Networks & Wireless Communication (AREA)
- Signal Processing (AREA)
- Environmental & Geological Engineering (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Medical Informatics (AREA)
- Probability & Statistics with Applications (AREA)
- Data Exchanges In Wide-Area Networks (AREA)
- Debugging And Monitoring (AREA)
- Maintenance And Management Of Digital Transmission (AREA)
Abstract
Description
本発明の実施の形態について説明する。実施の形態に係る評価装置は、正常な学習データを学習した学習データ用VAEモデルに加え、過検知データのみを学習した過検知用VAEモデルを生成する。過検知データは、評価処理の過程において異常と評価された正常な通信データであり、少量しか発生しない。本実施の形態に係る評価装置は、生成した2つのVAEモデルをモデルレベルで結合して得られた確率密度を基に評価を行うため、過検知データのフィードバックと検知の高精度化とを実現する。
そこで、実施の形態に係る評価装置の構成について具体的に説明する。図1は、実施の形態に係る評価装置の構成の一例を示す図である。図1に示すように、評価装置1は、通信部10、記憶部11及び制御部12を有する。
次に、評価装置1が初期段階に行う学習処理について説明する。図8は、図1に示す評価装置1が初期段階に行う学習処理の処理手順を示すフローチャートである。
次に、評価装置1の評価処理について説明する。図9は、図1に示す評価装置1が行う評価処理の処理手順を示すフローチャートである。
例えば、本実施の形態に係る評価装置1は、IoT機器の異常検知に適用することができる。図10は、実施の形態に係る評価装置1の適用例を説明する図である。図10に示すように、複数のIoT機器2が接続されたネットワーク3上に、評価装置1を設ける。この場合、評価装置1は、IoT機器2が送受信するトラフィックセッション情報を収集し、正常トラフィックセッションの確率密度の学習、及び、異常トラフィックセッションの検知を行う。
次に、従来の評価方法について説明する。図12は、従来の評価方法のフィードバック学習を説明する図である。図13及び図14は、従来の評価方法において用いられるモデルを説明する図である。
そこで、実際のIoT機器間のトラフィックセッションデータに対し、従来の評価方法と、本実施の形態に係る評価方法とを用いてそれぞれ評価を行った結果を示す。学習データは、カメラ通信(369データ)であり、過検知データは、SSH通信(10データ)である。
このように、本実施の形態では、正常な学習データを学習した学習データ用VAEモデルと、過検知データを学習した過検知用VAEモデルとを用いて、評価データの確率密度を推定し、推定した確率密度を基に評価データの異常の有無を評価する。すなわち、本実施の形態では、正常な学習データを学習した学習データ用VAEモデルとは別に、過検知データのみをフィードバック学習した過検知用VAEモデルとを生成し、生成した2つのVAEモデルが推定した確率密度を結合して得られた確率密度を基に評価を行う。
図示した各装置の各構成要素は機能概念的なものであり、必ずしも物理的に図示の如く構成されていることを要しない。すなわち、各装置の分散・統合の具体的形態は図示のものに限られず、その全部又は一部を、各種の負荷や使用状況等に応じて、任意の単位で機能的又は物理的に分散・統合して構成することができる。さらに、各装置にて行なわれる各処理機能は、その全部又は任意の一部が、CPU及び当該CPUにて解析実行されるプログラムにて実現され、あるいは、ワイヤードロジックによるハードウェアとして実現され得る。
図15は、プログラムが実行されることにより、評価装置1が実現されるコンピュータの一例を示す図である。コンピュータ1000は、例えば、メモリ1010、CPU1020を有する。また、コンピュータ1000は、ハードディスクドライブインタフェース1030、ディスクドライブインタフェース1040、シリアルポートインタフェース1050、ビデオアダプタ1060、ネットワークインタフェース1070を有する。これらの各部は、バス1080によって接続される。
2 IoT機器
3 ネットワーク
10 通信部
11 記憶部
12 制御部
111 学習データ用VAEモデル
112 過検知用VAEモデル
120 受付部
121 モデル生成部
122 VAE
123 評価部
124 結合部
1251 第1VAE
1252 第2VAE
126 異常有無評価部
Claims (3)
- 評価対象の通信データの入力を受け付ける受付部と、
正常な初期学習データの確率密度の特徴を学習した第1のモデルと、評価処理の過程において異常と検知された正常な過検知データの確率密度の特徴を学習した第2のモデルとを用いて前記評価対象の通信データの確率密度を推定し、推定した確率密度を基に前記評価対象の通信データの異常の有無を評価する評価部と、
を有することを特徴とする評価装置。 - 前記正常な初期学習データが入力された場合に前記正常な初期学習データの確率密度の特徴を学習して前記第1のモデルを生成し、前記評価処理の過程にて収集された前記過検知データが入力された場合に前記過検知データの確率密度の特徴を学習して前記第2のモデルを生成する生成部をさらに有し、
前記評価部は、前記第1のモデルを適用して推定した確率密度と、前記第2のモデルを適用して推定した確率密度とを結合した確率密度を基に前記評価対象の通信データの異常の有無を評価することを特徴とする請求項1に記載の評価装置。 - 評価装置によって実行される評価方法であって、
評価対象の通信データの入力を受け付ける工程と、
正常な初期学習データの確率密度の特徴を学習した第1のモデルと、評価処理の過程において異常と検知された正常な過検知データの確率密度の特徴を学習した第2のモデルとを用いて前記評価対象の通信データの確率密度を推定し、推定した確率密度を基に前記評価対象の通信データの異常の有無を評価する工程と、
を含んだことを特徴とする評価方法。
Priority Applications (4)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
US17/252,751 US20210256402A1 (en) | 2018-06-20 | 2019-06-18 | Evaluation device and evaluation method |
AU2019288014A AU2019288014B2 (en) | 2018-06-20 | 2019-06-18 | Evaluation device and evaluation method |
EP19823058.3A EP3796599B1 (en) | 2018-06-20 | 2019-06-18 | Evaluation device and evaluation method |
CN201980040020.2A CN112425123B (zh) | 2018-06-20 | 2019-06-18 | 评价装置和评价方法 |
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
JP2018-117456 | 2018-06-20 | ||
JP2018117456A JP6767434B2 (ja) | 2018-06-20 | 2018-06-20 | 評価装置及び評価方法 |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2019244902A1 true WO2019244902A1 (ja) | 2019-12-26 |
Family
ID=68984033
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/JP2019/024167 WO2019244902A1 (ja) | 2018-06-20 | 2019-06-18 | 評価装置及び評価方法 |
Country Status (6)
Country | Link |
---|---|
US (1) | US20210256402A1 (ja) |
EP (1) | EP3796599B1 (ja) |
JP (1) | JP6767434B2 (ja) |
CN (1) | CN112425123B (ja) |
AU (1) | AU2019288014B2 (ja) |
WO (1) | WO2019244902A1 (ja) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JPWO2021149225A1 (ja) * | 2020-01-23 | 2021-07-29 |
Families Citing this family (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US11928208B2 (en) * | 2018-10-02 | 2024-03-12 | Nippon Telegraph And Telephone Corporation | Calculation device, calculation method, and calculation program |
AU2020468530B2 (en) * | 2020-09-18 | 2024-02-01 | Nippon Telegraph And Telephone Corporation | Assessment device, assessment method, and assessment program |
Family Cites Families (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP4668092B2 (ja) * | 2006-03-03 | 2011-04-13 | 三菱電機株式会社 | 学習能力評価装置、学習能力評価方法及び学習能力評価プログラム |
JP2009070071A (ja) * | 2007-09-12 | 2009-04-02 | Toshiba Corp | 学習型プロセス異常診断装置、およびオペレータ判断推測結果収集装置 |
JP4940220B2 (ja) * | 2008-10-15 | 2012-05-30 | 株式会社東芝 | 異常動作検出装置及びプログラム |
CN102814340B (zh) * | 2011-06-08 | 2014-07-09 | 鞍钢股份有限公司 | 热轧带钢宽度控制模型智能学习系统及自学习方法 |
JP2015026252A (ja) * | 2013-07-26 | 2015-02-05 | 株式会社豊田中央研究所 | 異常検知装置及びプログラム |
CN108431834A (zh) * | 2015-12-01 | 2018-08-21 | 首选网络株式会社 | 异常检测系统、异常检测方法、异常检测程序及学得模型的生成方法 |
CN106790008B (zh) * | 2016-12-13 | 2018-08-24 | 浙江中都信息技术有限公司 | 用于在企业网络中检测异常主机的机器学习系统 |
US10635565B2 (en) * | 2017-10-04 | 2020-04-28 | Servicenow, Inc. | Systems and methods for robust anomaly detection |
-
2018
- 2018-06-20 JP JP2018117456A patent/JP6767434B2/ja active Active
-
2019
- 2019-06-18 CN CN201980040020.2A patent/CN112425123B/zh active Active
- 2019-06-18 EP EP19823058.3A patent/EP3796599B1/en active Active
- 2019-06-18 AU AU2019288014A patent/AU2019288014B2/en active Active
- 2019-06-18 US US17/252,751 patent/US20210256402A1/en active Pending
- 2019-06-18 WO PCT/JP2019/024167 patent/WO2019244902A1/ja unknown
Non-Patent Citations (2)
Title |
---|
DIEDERIK P KINGMAMAX WELLING, AUTO-ENCODING VARIATIONAL BAYES, 7 June 2018 (2018-06-07), Retrieved from the Internet <URL:https://arxiv.org/abs/1312.6114> |
MASAYOSHI SHIGETA , HISARNICHI OHTANI : "Implementation and evaluation of malware-infected terminal detection system using unsupervised learning", THE 2018 SYMPOSIUM ON CRYPTOGRAPHY AND INFORMATION SECURITY (SCIS2018), 26 January 2018 (2018-01-26), pages 1 - 7, XP009524793 * |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JPWO2021149225A1 (ja) * | 2020-01-23 | 2021-07-29 | ||
WO2021149225A1 (ja) * | 2020-01-23 | 2021-07-29 | 三菱電機株式会社 | モデル生成装置、モデル生成方法及びモデル生成プログラム |
JP7004479B2 (ja) | 2020-01-23 | 2022-01-21 | 三菱電機株式会社 | モデル生成装置、モデル生成方法及びモデル生成プログラム |
Also Published As
Publication number | Publication date |
---|---|
EP3796599A1 (en) | 2021-03-24 |
US20210256402A1 (en) | 2021-08-19 |
JP6767434B2 (ja) | 2020-10-14 |
EP3796599B1 (en) | 2023-10-04 |
AU2019288014A1 (en) | 2021-01-14 |
AU2019288014B2 (en) | 2022-03-17 |
CN112425123A (zh) | 2021-02-26 |
CN112425123B (zh) | 2023-10-27 |
EP3796599A4 (en) | 2022-03-16 |
JP2019220866A (ja) | 2019-12-26 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
JP6691094B2 (ja) | 学習装置、検知システム、学習方法及び学習プログラム | |
WO2019244902A1 (ja) | 評価装置及び評価方法 | |
US20150334125A1 (en) | Identifying threats based on hierarchical classification | |
JP6823501B2 (ja) | 異常検知装置、異常検知方法及びプログラム | |
JP2014229317A (ja) | 1つ以上の画像処理アルゴリズムの自動選択のための方法およびシステム | |
JP6564799B2 (ja) | 閾値決定装置、閾値決定方法及びプログラム | |
JP6845125B2 (ja) | 学習装置、学習方法及び学習プログラム | |
WO2019240038A1 (ja) | 検知装置及び検知方法 | |
JP7127525B2 (ja) | 検知装置、検知方法、および、検知プログラム | |
US20210344607A1 (en) | System and method for classifying network devices | |
US10275890B2 (en) | Image processing device and method for creating a background image | |
WO2020234977A1 (ja) | 情報処理装置、作成方法および作成プログラム | |
Yu-Ting et al. | Real-time risk assessment based on hidden Markov model and security configuration | |
US20200064460A1 (en) | Decision based re-processing framework for radar technology | |
JP7302660B2 (ja) | 学習装置、検知システム、学習方法及び学習プログラム | |
US20220374780A1 (en) | Training device, training method, and training program | |
US20230334361A1 (en) | Training device, training method, and training program | |
US20230177151A1 (en) | System and method of microcontroller security | |
JP7401747B2 (ja) | 類別プログラム、類別装置及び類別方法 | |
JP7176630B2 (ja) | 検知装置、検知方法および検知プログラム | |
CN115499251A (zh) | 一种边缘IoT设备的异常流量及攻击检测方法及系统 | |
CN111667021A (zh) | 一种基于人工智能的前端性能问题检测方法 |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 19823058 Country of ref document: EP Kind code of ref document: A1 |
|
ENP | Entry into the national phase |
Ref document number: 2019823058 Country of ref document: EP Effective date: 20201214 |
|
NENP | Non-entry into the national phase |
Ref country code: DE |
|
ENP | Entry into the national phase |
Ref document number: 2019288014 Country of ref document: AU Date of ref document: 20190618 Kind code of ref document: A |