JP2013150083A - Network abnormality detection device and network abnormality detection method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide technology which is adaptable in network abnormality detection to the periodical fluctuation of a network.SOLUTION: In a network abnormality detection device 100, a matrix generation unit 111 generates a matrix composed of traffic volumes input and output by devices operating in a network, and an eigenvector arithmetic unit 112 calculates an eigenvector of the matrix, where an eigenvector at the present point of time is referred to as the current vector. A reference vector arithmetic unit 113 executes time series prediction arithmetic on eigenvectors in a predetermined past period by taking the periodical fluctuation of a network into account to calculate a reference vector. An abnormality determination unit 114 calculates an abnormality score from an angular difference between the current vector and the reference vector. Then the abnormality determination unit 114 compares the abnormality score and a threshold value and, if the abnormality score is larger than the threshold value, determines that the network has abnormality. A threshold value arithmetic unit 115 executes time series prediction arithmetic on the past abnormality scores or past threshold values to calculate a threshold value.

Description

  The present invention relates to a technique for detecting a network abnormality.

  In order to provide a network service stably, a telecommunications carrier can detect the occurrence of an abnormality in a device (node) in the network. As an abnormality detection method, the following two methods I and II are roughly used.

  Method I uses a warning that is spontaneously transmitted from a device in the network by using Syslog, SNMP Trap (Simple Network Management Protocol Trap), or the like. In Method I, an alarm is issued only when the self-device monitoring program in the device detects an abnormality such as a rise in temperature or a broken link. Therefore, Method I cannot detect an abnormality that is not set in the device monitoring program, such as an event in which the traffic volume gradually decreases.

  Method II uses ping (Packet Internet Grouper), SNMP Request, or the like to analyze information obtained by making an inquiry to a device in the network and detect an abnormality. In Method II, various information such as CPU (Central Processing Unit) temperature and traffic information obtained by inquiring to the device can be collected, but normal / abnormal can be determined by separately analyzing the collected information. Therefore, analysis skills are required.

  For example, a phenomenon close to normal behavior is not issued by Method I, and is difficult to detect as an abnormality by Method II. One example will be described with reference to FIGS. FIGS. 10A and 10B show a case where routers A, B, and C are connected in series. FIG. 10A shows a case where the traffic volume between routers A and B is equal to the traffic volume between routers B and C. In the case of FIG. 10A, it is determined to be normal in both methods I and II.

  On the other hand, the example shown in FIG. 10B is a phenomenon in which an abnormality occurs in the router B due to aged deterioration or the like, and the traffic volume between the routers A and B and the traffic volume between the routers B and C are not equal. It represents the case where has occurred. In the case of FIG. 10B, even if the traffic volume between the routers B-C gradually decreases, the traffic is not completely blocked, so that the alarm is not issued in Method I. In Method II, even if the information about the traffic volume collected from each router is analyzed, the traffic volume may naturally decrease depending on the network usage status, or it may increase suddenly due to a disaster or event. Therefore, it is difficult to determine whether or not there is an abnormality only from the increase / decrease trend of the traffic volume.

  Non-Patent Documents 1 and 2 below disclose techniques relating to abnormality detection by Method II. Specifically, as shown in FIG. 11, in the case where devices A, B, and C are connected in series, a matrix is generated with a matrix row as a transmitting device and a column as a receiving device. The eigenvector of the matrix is calculated at a predetermined cycle using the traffic amount acquired at a predetermined cycle as the matrix component. The eigenvector of the matrix does not change when all the components of the matrix are multiplied by a constant, but changes when the values of some of the components of the matrix change. Therefore, when traffic fluctuation occurs due to an abnormality or the like in a part of the network, the eigenvector changes, and the abnormality can be detected.

  In the techniques described in Non-Patent Documents 1 and 2, a moving average of the past K eigenvector time series when no abnormality has occurred is defined as a reference vector (Vec-Ref). The reference vector (Vec-Ref) is used as a criterion for determining whether or not the current eigenvector (that is, the current vector (Vec-Now)) is abnormal. Then, the abnormal score (θ) is calculated from the angle difference between the current vector (Vec-Now) and the reference vector (Vec-Ref), and if the abnormal score (θ) exceeds the threshold, an abnormality is detected in the network. It is determined that

  In addition, Non-Patent Documents 1 and 2 are new in consideration of changes in traffic volume and changes in equipment load due to increase / decrease in subscribers, service changes, etc., without fixing the reference vector (Vec-Ref). Discloses a method of calculating the latest reference vector (Vec-Ref) while learning the current vector (Vec-Now) calculated in (1). As an example of the learning, a case is described in which a moving average of reference vectors (Vec-Ref) for a certain period from the present to the past is taken. In addition, Non-Patent Documents 1 and 2 calculate the appearance frequency distribution of abnormal score (θ) values for a certain period in the past without fixing the threshold, and set the threshold when a predetermined probability (for example, 5%) is reached. It is calculated as

Ide, T. and Kashima, H. “Eigenspace-based anomaly detection in computer systems” In Proceedings of the 10th ACM SIGKDD international conference on knowledge discovery and data mining, ACM Press, New York, NY, USA, p.440-449 , 2004 Takeshi Ide and Hisayoshi Kashima, “Detection of computer system faults in eigenspace”, Proceedings of the 18th Annual Conference of the Japanese Society for Artificial Intelligence, 3F3-05, 2004

  However, in Non-Patent Documents 1 and 2, the reference vector (Vec-Ref) is calculated by the moving average using the eigenvector time series for a certain period from the present to the past, and the threshold value is the abnormal score (θ) of the past certain period. Since it is calculated from the frequency distribution of the values, one time slot (a time slot is a predetermined cycle for calculating eigenvectors) may not be suitable for determining normal / abnormal network in the future. is there. In other words, there is a possibility that the network is not sufficiently adapted to periodic fluctuations.

  For example, in a telecommunications carrier's network, the amount of video distribution traffic tends to increase at midnight. In a corporate network, the amount of traffic tends to increase with the increase in the number of accesses to the Web server and the number of e-mails during employee work hours. In this way, the amount of network traffic fluctuates greatly periodically in each time zone. Therefore, even if the network is in a normal state, the future fluctuation trend ahead of one time slot may be significantly different from the gentle fluctuation trend obtained from the entire time series for a certain period in the past. Therefore, the methods described in Non-Patent Documents 1 and 2 may not be able to correctly determine normality / abnormality with respect to periodic fluctuations in the network.

  Therefore, an object of the present invention is to provide a technique that can be adapted to periodic fluctuations in the network in network abnormality detection.

  In the present invention, the network anomaly detection device calculates an eigenvector from a matrix indicating a state between devices in the current network, the calculated eigenvector is a current vector, and the eigenvector is associated with the time when the current vector is calculated. An eigenvector calculation unit that is stored as a past eigenvector in the storage unit, and a first period that indicates a variation cycle of a state between devices in the network in a predetermined time period, among the past eigenvectors stored in the storage unit. Obtaining the past eigenvector that exists, a reference vector calculation unit that calculates a reference vector from the acquired past eigenvector, and calculating an abnormality score from an angle difference between the current vector and the reference vector, Based on the magnitude relationship between the abnormal score and the threshold value. Characterized in that it comprises an abnormality determination unit determining abnormality of the workpiece, the.

  Further, the present invention is a network abnormality detection method of a network abnormality detection device for determining a network abnormality, wherein the network abnormality detection device calculates an eigenvector from a matrix indicating a state between devices in a current network, The calculated eigenvector as a current vector, the eigenvector calculation step of storing the eigenvector as a past eigenvector in association with the time when the current vector was calculated, and the past eigenvector stored in the storage unit, A reference vector calculation step of acquiring the past eigenvector existing in a predetermined time zone in a first period indicating a fluctuation cycle of a state between devices in the network, and calculating a reference vector from the acquired past eigenvector; The angle between the vector and the reference vector An abnormality score is calculated, the abnormality score is compared with a threshold value, and an abnormality determination step of determining a network abnormality based on a magnitude relationship between the abnormality score and the threshold value is executed. .

  According to such a configuration, the network abnormality detection device uses a past eigenvector existing in a predetermined time zone with a state variation cycle (first cycle) between devices in the network, and a reference for one time slot ahead A vector can be calculated. That is, it is possible to calculate a reference vector in anticipation of changes in the current vector. Therefore, the network abnormality detection device can be adapted to periodic fluctuations in the network, and can suppress a normal state from being determined as abnormal.

  Further, in the present invention, when the network abnormality detection device stores the abnormality score in the storage unit, the abnormality abnormality stored in the storage unit is acquired as a past abnormality score, and the acquired past It further includes a threshold value calculation unit that obtains the appearance frequency distribution of the abnormal score values and sets the threshold value as a predetermined probability.

  According to such a configuration, the value of the abnormality score (past abnormality score) calculated from the angle difference between the reference vector adapted to the periodic fluctuation of the network and the current vector is less varied. That is, it is possible to prevent the past abnormality score from becoming a large value due to a sudden change in the reference vector. For this reason, the network abnormality detection apparatus can stably calculate a threshold value adapted to a sudden change in the reference vector by reducing the variance of the appearance frequency distribution. Therefore, the network abnormality detection apparatus can stably detect the network abnormality by adapting to the periodic fluctuation of the network in the abnormality detection of the network.

  Further, in the present invention, when the abnormality score is stored in the storage unit in association with the time when the current vector used when calculating the abnormality score is stored, the threshold value calculation unit is stored in the storage unit. Further, the abnormality score existing in a predetermined time zone in the first period is acquired as the past abnormality score from the abnormality score.

  According to such a configuration, the threshold value is calculated based on a past abnormality score existing in a predetermined time zone in a state change cycle (first cycle) between devices in the network. Therefore, the network abnormality detection device can be adapted to the periodic fluctuation of the network in the network abnormality detection.

  Further, in the present invention, the threshold calculation unit calculates the prediction abnormality score by applying the acquired past abnormality score to the time-series prediction calculation method, obtains the appearance frequency distribution of the calculated value of the prediction abnormality score, The threshold value is set when the predetermined probability is reached.

  According to such a configuration, the threshold value is calculated based on the predicted abnormality score calculated by applying the past abnormality score to the time-series prediction calculation method. Therefore, the network abnormality detection device can be adapted to the periodic fluctuation of the network in the network abnormality detection.

  In the present invention, when the threshold value calculation unit stores the threshold value in the storage unit, the threshold value stored in the storage unit is acquired as a past threshold value, and the acquired past threshold value is time-series predicted. A new threshold value is set by applying to the calculation method.

  According to such a configuration, the new threshold value is calculated by applying the past threshold value to the time series prediction calculation method. Therefore, the network abnormality detection device can be adapted to the periodic fluctuation of the network in the network abnormality detection.

  In the present invention, when the threshold value calculation unit stores the threshold value in the storage unit in association with the time when the current vector used in the calculation of the threshold value is calculated, the threshold value calculation unit is stored in the storage unit. The threshold value existing in a predetermined time zone in the first cycle is acquired as the past threshold value from the threshold values.

  According to such a configuration, the threshold value is calculated based on a past threshold value existing in a predetermined time zone in a state change cycle (first cycle) between devices in the network. Therefore, the network abnormality detection device can be adapted to the periodic fluctuation of the network in the network abnormality detection.

  Further, the present invention is characterized in that the threshold value calculation unit generates the appearance frequency distribution as a one-dimensional normal distribution, a multidimensional normal distribution, or a multivariate normal distribution.

  According to such a configuration, when handling a distribution in multiple dimensions, a threshold value can be set finely for each vector component of the abnormality score.

  ADVANTAGE OF THE INVENTION According to this invention, the technique which can be adapted to the periodic fluctuation | variation of a network in the abnormality detection of a network can be provided.

(A) is a figure showing the structure of the network abnormality detection system containing the network abnormality detection apparatus in this embodiment, (b) is a figure showing the abnormality determination method. It is a figure which shows the outline | summary of the abnormality determination method in this embodiment, (a) represents a comparative example, (b) represents the case which predicted the threshold value, (c) represented the reference vector (Vec-Ref). Represents a case. It is a figure which shows the structural example of a network abnormality detection apparatus. It is a figure which shows an example of past eigenvector DB. It is a figure which shows an example of past abnormal score DB. It is a figure which shows an example of past threshold value DB. It is a figure which shows the example of a processing flow of a network abnormality detection apparatus. It is a figure which shows the calculation example of a reference | standard vector (Vec-Ref), (a) represents the data to be used, (b) represents the calculation method, (c) represents the combination to perform. It is a figure which shows the calculation example of a threshold value, (a) represents the data to be used, (b) represents the calculation method, (c) represents the distribution generation method, (d) represents the combination to be performed. It is a figure which shows the example in case an abnormality is not detected in a prior art, (a) represents a normal case, (b) is a case where abnormality detection is difficult even if the abnormality has occurred Represents. It is a figure which shows the example of a process for detecting abnormality in a prior art.

  A mode for carrying out the present invention (hereinafter referred to as “the present embodiment”) will be described in detail with reference to the drawings as appropriate.

(Overview)
First, a configuration of a network abnormality detection system including a network abnormality detection device of this embodiment and an abnormality determination method of the network abnormality detection device will be described with reference to FIGS.

  As shown in FIG. 1 (a), the network abnormality detection system 1 is configured by communicably connecting a device 10 (10A, 10B, 10C), which is a node in the network 11, and a network abnormality detection device 100. . The network abnormality detection apparatus 100 has a function of detecting an abnormality of the network 11 by analyzing information obtained by making an inquiry to the apparatus 10 as in the method II described above. In FIG. 1A, three devices are described, but the number is not limited to three.

  When the network abnormality detection apparatus 100 detects an abnormality in the network 11, as in the case described with reference to FIG. 11, the reference vector (Vec-Ref) is determined based on whether or not the current vector (Vec-Now) is abnormal. It is used as a criterion. The current vector (Vec-Now) is an eigenvector calculated from a matrix indicating the current state between the devices 10 in the network 11. The reference vector (Vec-Ref) is calculated based on the eigenvector calculated in the past.

  Then, as shown in FIG. 1B, the network abnormality detection device 100 has an abnormality score (θ) obtained from the angle difference between the reference vector (Vec-Ref) and the current vector (Vec-Now) larger than the threshold value. Whether or not the abnormality score (θ) is larger than the threshold is determined as abnormal.

Next, as an example of the periodic fluctuation of the network, as shown in the uppermost part of FIG. 2, a case where the traffic volume increases in the employee work hours in the company network will be taken as an example, and the abnormality determination method will be described. .
As shown in the top row of FIG. 2, the amount of traffic in a corporate network generally shows a gradual fluctuation at a small value at midnight, a sudden fluctuation when going to work, and a gradual fluctuation at a large value during the daytime. It tends to show fluctuations.

First, FIG. 2A is a comparative example in which a reference vector (Vec-Ref) is calculated by a moving average using an eigenvector time series of a past fixed period, and a threshold value is a value of an abnormal score (θ) of the past fixed period. This is a schematic representation of the case calculated from the appearance frequency distribution. In particular, the current vector (Vec-Now) changes greatly due to abrupt fluctuations due to attendance in the range surrounded by the dashed-dotted ellipse. On the other hand, since the reference vector (Vec-Ref) is calculated by a moving average of past eigenvector time series, it does not change greatly. Therefore, the abnormal score (θ) is prominently increased. Then, the abnormal score (θ) becomes larger than the threshold value (abnormal score (θ)> threshold value), and is determined to be abnormal.
However, since the network is not abnormal but is determined to be abnormal, the network maintainer is required to perform unnecessary inspection work, thereby reducing the efficiency of the network operation work.

Next, FIG. 2B schematically illustrates a case in which a change in the abnormal score (θ) is predicted and the threshold value is changed according to the change (threshold value is predicted). That is, when the abnormal score (θ) increases, the threshold value is increased to suppress the determination as abnormal.
In this case, the calculation method of the current vector (Vec-Now) and the reference vector (Vec-Ref) may be the same as in the comparative example. However, the threshold value of one time slot ahead is predicted by performing a time series prediction calculation using the abnormal score (θ) or threshold value for a certain period in the past. Then, a threshold is set in advance so that a large change in the abnormality score (θ) that occurs in a normal state is not determined to be abnormal. As a result, it is possible to prevent the normal state from being erroneously determined as an abnormal state due to an abnormal score (θ) <threshold.

FIG. 2C shows a case in which a change in the current vector (Vec-Now) is predicted, and the reference vector (Vec-Ref) is changed in accordance with the change (the reference vector (Vec-Ref) is predicted). This is a schematic representation. In other words, the abnormality score (θ) calculated from the angle difference between the current vector (Vec-Now) and the reference vector (Vec-Ref) is kept almost constant, thereby preventing the abnormality from being determined. is there.
In this case, the method for calculating the current vector (Vec-Now) and the threshold value may be the same as in the comparative example. However, a reference vector (Vec-Ref) of one time slot ahead is calculated by performing time-series prediction calculation on the reference vector (Vec-Ref) using an eigenvector for a certain past period. Therefore, the fluctuation of the abnormal score (θ) is calculated to be smaller than that in the comparative example. As a result, it is possible to prevent the normal state from being erroneously determined as an abnormal state due to an abnormal score (θ) <threshold.

(Configuration of network error detection device)
First, the configuration of the network abnormality detection device 100 will be described with reference to FIG. 3 (see FIGS. 1 and 11 as appropriate). As shown in FIG. 3, the network abnormality detection apparatus 100 includes a processing unit 110, a storage unit 120, and an input / output unit 130.
The processing unit 110 further includes a matrix generation unit 111, an eigenvector calculation unit 112, a reference vector calculation unit 113, an abnormality determination unit 114, and a threshold calculation unit 115. The processing unit 110 is configured by a CPU (Central Processing Unit) and a main memory (not shown), and an application program stored in the storage unit 120 is expanded in the main memory to embody each unit in the processing unit 110.

  The matrix generation unit 111 uses the traffic amount between the devices 10 in the network 11 acquired via the input unit 131 of the input / output unit 130 as components, and sets the rows of the matrix as the transmission side device 10 and the columns as the reception side device 10 As a matrix (see FIG. 11). The generation of the matrix is performed at a predetermined cycle or when instructed by an administrator or maintenance person of the network 11.

  The eigenvector calculation unit 112 has a function of calculating the eigenvector (see FIG. 11) of the matrix generated by the matrix generation unit 111 at a predetermined period. The calculated latest eigenvector is output to the abnormality determination unit 114 as the current vector (Vec-Now). The eigenvector computing unit 112 has a function of storing the eigenvector (current vector) as a past eigenvector in the past eigenvector DB 121 in association with the time when the eigenvector (current vector) is calculated.

  The reference vector calculation unit 113 has a function of reading a past eigenvector from the past eigenvector DB 121 and calculating a reference vector (Vec-Ref). Details of the calculation method of the reference vector (Vec-Ref) will be described later.

  The abnormality determination unit 114 acquires the current vector (Vec-Now) from the eigenvector calculation unit 112, acquires the reference vector (Vec-Ref) from the reference vector calculation unit 113, and acquires the current vector (Vec-Now) and the reference vector ( The abnormal score (θ) is calculated from the angle difference from Vec-Ref). The abnormality determination unit 114 has a function of comparing the abnormality score (θ) with a threshold value and determining a network abnormality based on the magnitude relationship between the abnormality score and the threshold value. The threshold value is calculated by a threshold value calculation unit 115 described later. For example, the abnormality determination unit 114 determines that the abnormality is normal when the abnormality score (θ) is equal to or less than the threshold, and determines that the abnormality is abnormal when the abnormality score (θ) is greater than the threshold. Further, the abnormality determination unit 114 has a function of storing the abnormality score (θ) in the past abnormality score DB 122 in association with the time when the current vector used for calculating the abnormality score (θ) is calculated. Further, the abnormality determination unit 114 has a function of outputting information such as an abnormal state to the display device via the output unit 132 of the input / output unit 130.

  The threshold value calculation unit 115 has a function of reading out the abnormality score (θ) from the past abnormality score DB 122 as a past abnormality score, executing a predetermined calculation, and calculating (setting) a threshold value for one time slot ahead. Further, the threshold value calculation unit 115 stores the calculated threshold value in the past threshold value DB 123 in association with the time when the current vector (Vec-Now) used in the calculation of the threshold value is calculated, and also stores it in the past threshold value DB 123. And reading a threshold value as a past threshold value and calculating a threshold value (new threshold value) one time slot ahead. Details of the threshold calculation method will be described later.

  The input / output unit 130 includes an input unit 131 and an output unit 132. The input unit 131 is a communication interface for connecting to the network 11 and receives information such as traffic volume from the device 10 in the network 11. The output unit 132 is an interface for connecting to the display device, and outputs information such as a processing result of the processing unit 110 to the display device.

  The storage unit 120 is a hard disk or the like, and stores a past eigenvector DB 121, a past abnormality score DB 122, and a past threshold value DB 123.

  As shown in FIG. 4, the past eigenvector DB 121 stores the time t (n) when the eigenvector computing unit 112 (see FIG. 3) calculates the eigenvector and the eigenvector U (n) in association with each other. However, the eigenvector U (n−k−1) that has already passed the predetermined time (k + 1) at the current time t (n) is deleted together with the time t (n−k−1). Note that n at time t (n) is a number for identifying the calculated cycle. The time t (n) stored in the past eigenvector DB 121 represents the present, and the eigenvector at the time t (n) is the current vector (Vec-Now). In addition, past eigenvectors U (n−1), U (n−2),..., Etc. before time t (n−1) are the reference vectors (Vec-Ref) in the reference vector calculation unit 113 (see FIG. 3). ) Is used for calculation processing.

  As shown in FIG. 5, the past abnormality score DB 122 stores time t (n) and the abnormality score θ (n) in association with each other. However, the abnormal score θ (n−k−1) that has already passed the predetermined time (k + 1) at the current time t (n) is deleted together with the time t (n−k−1).

  As shown in FIG. 6, the past threshold value DB 123 stores time t (n) and the threshold value th (n) calculated by the threshold value calculation unit 115 in association with each other. However, the threshold th (n−k−1) that has already passed the predetermined time (k + 1) at the current time t (n) is deleted together with the time t (n−k−1).

(Processing flow of network error detection device)
Next, a processing flow example of the network abnormality detection apparatus 100 will be described with reference to FIG. 7 (see FIG. 3 as appropriate).
In step S <b> 701, the matrix generation unit 111 acquires traffic information such as the traffic amount between the devices 10 in the network 11 via the input unit 131 and generates, for example, a matrix having the traffic amount as a component. The generation of the matrix is performed at a predetermined cycle or when instructed by an administrator or maintenance person of the network 11.

  In step S <b> 702, the eigenvector calculation unit 112 calculates a current vector (Vec-Now) that is an eigenvector at a predetermined cycle from the matrix generated by the matrix generation unit 111.

  In step S703, the eigenvector computing unit 112 stores the current vector (Vec-Now) as a past eigenvector in the past eigenvector DB 121 in association with the calculated time.

  In step S704, the reference vector calculation unit 113 reads a past eigenvector from the past eigenvector DB 121, executes a predetermined calculation, and calculates a reference vector (Vec-Ref). Details of the calculation method of the reference vector (Vec-Ref) will be described later.

  In step S705, the threshold value calculation unit 115 reads the abnormality score from the past abnormality score DB 122 as a past abnormality score, executes a predetermined calculation, and calculates a threshold value for one time slot ahead. Alternatively, the threshold value calculation unit 115 reads the threshold value from the past threshold value DB 123 as a past threshold value, executes a predetermined calculation, and calculates a threshold value for one time slot. Details of the threshold calculation method will be described later.

  In step S706, the abnormality determination unit 114 calculates an abnormality score (θ) from the current vector (Vec-Now) and the reference vector (Vec-Ref). Specifically, the abnormality determination unit 114 calculates an angle difference between the current vector (Vec-Now) and the reference vector (Vec-Ref), and calculates an abnormality score (θ) from the angle difference.

In step S707, the abnormality determination unit 114 determines whether or not the abnormality score (θ) calculated in step S706 is greater than the threshold calculated in step S705.
If it is determined that the abnormal score (θ) is greater than the threshold (Yes in step S707), the process proceeds to step S708. If the abnormal score (θ) is determined to be equal to or less than the threshold (No in step S707), the process proceeds to step S709.

  In step S708, the abnormality determination unit 114 executes a countermeasure at the time of abnormality. Specifically, the abnormality determination unit 114 outputs a message, alarm, or the like notifying abnormality to the display device via the output unit 132.

  In step S709, the abnormality determination unit 114 stores the abnormality score (θ) calculated in step S706 in the past abnormality score DB 122. Then, when the process of step S709 ends, the process returns to step S701.

  Although not described in FIG. 7, in the processing after step S <b> 705, the threshold value calculation unit 115 may store the threshold value calculated in step S <b> 705 in the past threshold value DB 123. Further, the order of step S704 and step S705 may be reversed.

(Calculation of reference vector)
Next, details of the calculation method of the reference vector (Vec-Ref) in step S704 will be described with reference to FIG. Here, mainly the prediction of the reference vector (Vec-Ref) ahead of one time slot will be described as described with reference to FIG.
First, a selection method for selecting data used for prediction will be described, and then a prediction value calculation method will be described.

  As shown in FIG. 8A, the selection method for selecting data to be used for prediction from the past eigenvector DB 121 is as follows. In the method A1, when selecting a time series of eigenvectors in a past predetermined period, The eigenvectors in the same time period in a predetermined period (one day, one week, etc.) within the predetermined period are selected.

Method A1 is used in the comparative example shown in FIG.
Method A2 can be used as a past time point that can coincide with the current time or a trend of one time slot ahead, for example, 1 day ago, 2 days ago, ..., the same time zone x days ago, 1 week ago, 2 weeks ago, ... The past eigenvectors in the same time zone on the same day of the week before y weeks are used. This is because, for example, in a corporate network, the traffic volume tends to fluctuate in a daily cycle or a weekly cycle (first cycle). As the same day, for example, if the current day is Wednesday, one week before, two weeks before, and so on are selected. The same time zone indicates that, for example, if the current time is 13:15, an eigenvector including 13:00 in the time data is selected. Alternatively, the same time zone may represent selecting an eigenvector within a predetermined time (for example, 30 minutes) before and after the current time.

  In addition, as shown in FIG. 8B, the calculation method of the predicted value includes a time series prediction calculation method (for example, an AR (Auto-Regressive) model, ARMA (Auto-Regressive Moving Average) model, Kalman filter, etc.) was used.

Method B1 is used in the comparative example shown in FIG.
In the method B2, when the value of each component of the current vector (Vec-Now) moves independently, for example, when the traffic output from the node A of the network flows into the neighboring node B, the vector B A one-dimensional time series prediction method such as an AR model or a Kalman filter can be applied to the component.

  In addition, when the method B2 is related to fluctuations in the value of each component of the current vector (Vec-Now), for example, there are many flows in which traffic output from the node C of the network passes through many nodes, By applying a multivariate AR model or the like, it is possible to express the future behavior of one time slot ahead.

Then, as shown in FIG. 8C, the combination for executing the calculation of the predicted value is determined by the combination of the methods A1 and A2 and the methods B1 and B2.
The first reference vector calculation method is a combination of the method A1 and the method B1, and is used in the comparative example shown in FIG.
The second reference vector calculation method is a combination of method A2 and method B1.
The third reference vector calculation method is a combination of method A2 and method B2.

(Calculation of threshold value)
Next, details of the threshold calculation method in step S705 will be described with reference to FIG. Here, mainly as described with reference to FIG. 2B, a method of predicting a threshold of one time slot ahead will be described.
First, a selection method for selecting data used for prediction will be described, and then a prediction value calculation method will be described.

  As shown in FIG. 9A, there are four selection methods for selecting data used for prediction. When the method a11 selects a time series of threshold values in the past predetermined period from the past threshold value DB 123, the method a12 sets the threshold value in the same period of the predetermined period (one day, one week, etc.) in the past predetermined period as the past threshold value. This is a case of selecting from the DB 123. In the method a21, when a time series of abnormal scores (θ) in the past predetermined period is selected from the past abnormal score DB 122, the method a22 is performed in a predetermined cycle (one day, one week, etc.) in the past predetermined period. This is a case where an abnormal score (θ) in the same time zone is selected from the past abnormal score DB 122.

In the methods a11 and a12, the predicted value of the threshold is directly calculated using the past threshold. On the other hand, in the methods a21 and a22, after calculating a prediction abnormality score indicating a prediction value of the past abnormality score (θ), an appearance frequency distribution of the value of the prediction abnormality score is obtained, and a predetermined probability (for example, 5 %) Is calculated as a threshold value.
In the methods a12 and a22, for example, in a corporate network, the traffic amount tends to fluctuate in a daily cycle or a weekly cycle (first cycle). The past time points that can be matched include, for example, the same time zone of 1 day ago, 2 days ago, ... x days ago, 1 week ago, 2 weeks ago, ... It is assumed that a past value (existing in a predetermined time zone in one cycle) is used.

  Further, as shown in FIG. 9B, the prediction value is calculated by using a weighted moving average method in the method b1, and a time series prediction calculation method (eg, AR model, ARMA model, Kalman filter, etc.) in the method b2. ).

  FIG. 9C shows two distribution generation methods. The method c1 is a one-dimensional normal distribution, and the method c2 is a multidimensional normal distribution or a multivariate normal distribution. This distribution generation method is used when generating an appearance frequency distribution of abnormal score values and predicted abnormal score values in the methods a21 and a22.

  As shown in FIG. 9D, the combination for calculating the predicted value of the threshold is determined by the combination of the methods a11, a12, a21, a22, the methods b1, b2, and the methods c1, c2.

The first threshold value calculation method is a combination of the method a21, the method b1, and the method c1, and is used in the comparative example shown in FIG.
The second threshold value calculation method is a combination of the method a11 and the method b2.
The third threshold value calculation method is a combination of the method a12 and the method b2.
Note that the second and third threshold value calculation methods directly calculate the threshold value, and thus it is not necessary to use the method c1 or the method c2.

The fourth threshold value calculation method is a combination of the method a21, the method b2, and the method c1.
The fifth threshold value calculation method is a combination of the method a21, the method b2, and the method c2.
The sixth threshold value calculation method is a combination of the method a22, the method b2, and the method c1.
The seventh threshold value calculation method is a combination of the method a22, the method b2, and the method c2.

  The combination of the reference vector (Vec-Ref) used in step S706 shown in FIG. 7 and the threshold used in step S707 is the reference vector (Vec-Ref) calculated by the first to third reference vector calculation methods. And a threshold value calculated by the first to seventh threshold value calculation methods. However, since the combination of the first reference vector calculation method and the first threshold value calculation method is a comparative example shown in FIG. 2A, it is not included in this embodiment.

  As described above, the network abnormality detection apparatus 100 described in the present embodiment uses a matrix row as a transmission side apparatus, a column as a reception side apparatus, a traffic amount as a matrix component, and a matrix eigenvector (current vector (Vec -Now)), and the abnormal score (θ) is calculated from the angle difference between the current vector (Vec-Now) and the reference vector (Vec-Ref). Then, the network abnormality detection device 100 determines that the network 11 is abnormal when the abnormality score (θ) is larger than the threshold value. At that time, the network anomaly detection apparatus 100 performs a time series prediction operation on the past eigenvectors so that it can be applied to periodic fluctuations in the traffic amount in the network 11, and performs a reference vector (Vec) one time slot ahead. -Ref). Alternatively, in order to change the threshold value so that it can be applied to the periodic fluctuation of the traffic amount in the network 11, the network abnormality detection device 100 performs time series prediction calculation on the past threshold value, and performs one time slot ahead. Calculate the threshold. Alternatively, the network abnormality detection apparatus 100 performs time series prediction calculation on the past abnormality score (θ) in order to change the threshold value so that it can be applied to periodic fluctuations in the traffic amount in the network 11, and performs prediction. An abnormality score is calculated, an appearance frequency distribution of the predicted abnormality score value is obtained, and a time when a predetermined probability is reached is set as a threshold of one time slot ahead. That is, since the network abnormality detection device 100 has such a configuration, it is possible to suppress periodic fluctuations in a normal network from being determined as abnormal.

DESCRIPTION OF SYMBOLS 1 Network abnormality detection system 10 (10A, 10B, 10C) Apparatus 11 Network 100 Network abnormality detection apparatus 110 Processing part 111 Matrix generation part 112 Eigenvector calculation part 113 Reference vector calculation part 114 Abnormality determination part 115 Threshold calculation part 120 Storage part 121 Past Eigenvector DB
122 Past abnormal score DB
123 Past threshold DB
130 Input / output unit 131 Input unit 132 Output unit

Claims (8)

Eigenvector calculation that calculates an eigenvector from a matrix indicating the state between devices in the current network, uses the calculated eigenvector as the current vector, and associates the eigenvector with the time when the current vector is calculated, and stores the eigenvector as a past eigenvector in the storage unit And
From the past eigenvectors stored in the storage unit, the past eigenvectors existing in a predetermined time zone with a first period indicating a fluctuation cycle of a state between devices in the network are acquired, and the acquired past A reference vector calculation unit for calculating a reference vector from the eigenvector;
An abnormality determination that calculates an abnormality score from an angle difference between the current vector and the reference vector, compares the abnormality score with a threshold value, and determines a network abnormality based on a magnitude relationship between the abnormality score and the threshold value And
A network abnormality detection device comprising: When the abnormality score is stored in the storage unit, the abnormality score stored in the storage unit is acquired as a past abnormality score, an appearance frequency distribution of the acquired values of the past abnormality score is obtained, and a predetermined The network abnormality detection device according to claim 1, further comprising a threshold value calculation unit that sets the probability as the threshold value. When the abnormality score is stored in the storage unit in association with the time when the current vector used when calculating the abnormality score is calculated,
The threshold calculation unit acquires the abnormality score existing in a predetermined time zone in the first period as the past abnormality score from the abnormality score stored in the storage unit. Item 3. The network abnormality detection device according to Item 2. The threshold calculation unit calculates the prediction abnormality score by applying the acquired past abnormality score to the time series prediction calculation method, obtains the appearance frequency distribution of the calculated value of the prediction abnormality score, and has a predetermined probability The network abnormality detection device according to claim 2, wherein the threshold value is set as the threshold value. When the threshold value is stored in the storage unit, the threshold value calculation unit acquires the threshold value stored in the storage unit as a past threshold value, and applies the acquired past threshold value to the time-series prediction calculation method. The network abnormality detection device according to claim 2, wherein a new threshold is set. When the threshold value calculation unit stores the threshold value in the storage unit in association with the time when the current vector used in the calculation of the threshold value is calculated, the threshold value calculation unit is selected from the threshold values stored in the storage unit. The network abnormality detection device according to claim 5, wherein the threshold existing in a predetermined time zone in the first period is acquired as the past threshold. 5. The network abnormality detection device according to claim 3, wherein the threshold value calculation unit generates the appearance frequency distribution as a one-dimensional normal distribution, a multidimensional normal distribution, or a multivariate normal distribution. A network abnormality detection method of a network abnormality detection device for judging a network abnormality,
The network abnormality detection device is:
Eigenvector calculation that calculates an eigenvector from a matrix indicating the state between devices in the current network, uses the calculated eigenvector as the current vector, and associates the eigenvector with the time when the current vector is calculated, and stores the eigenvector as a past eigenvector in the storage unit Steps,
From the past eigenvectors stored in the storage unit, the past eigenvectors existing in a predetermined time zone with a first period indicating a fluctuation cycle of a state between devices in the network are acquired, and the acquired past A reference vector calculation step for calculating a reference vector from the eigenvector;
An abnormality determination that calculates an abnormality score from an angle difference between the current vector and the reference vector, compares the abnormality score with a threshold value, and determines a network abnormality based on a magnitude relationship between the abnormality score and the threshold value Steps,
The network abnormality detection method characterized by performing.

Images