Attack detection method, attack detection system, and program
The method adjusts sampling intervals and times based on statistical variations and event information to achieve both data reduction and accurate attack detection on vehicle sensor data, addressing the limitations of fixed-interval sampling and incomplete data analysis.
Patent Information
- Application Number
- JP2021539055
- Authority / Receiving Office
- JP · JP
- Patent Type
- Patents
- Current Assignee / Owner
- Priority Date
- 2020-01-20
- Filing Date
- 2021-01-18
- Publication Date
- 2024-11-22
- Estimated Expiration
- 2041-01-18
AI Technical Summary
Existing technologies face challenges in achieving both data volume reduction and accurate attack detection on vehicle sensor data, as conventional methods may overlook attacks by sampling at fixed intervals or rely on message acquisition time, leading to incomplete data analysis.
An attack detection method that adjusts sampling intervals and times based on statistical variations and event information, calculating two types of anomaly scores from short-term and long-term sensor flows to determine attacks accurately.
This approach enables both data volume reduction and enhanced attack detection performance by dynamically adjusting sampling policies and using multiple anomaly scores, ensuring high-accuracy attack detection while minimizing monitoring costs.
Smart Images

Figure 0007588587000005 
Figure 0007588587000006 
Figure 0007588587000007
Abstract
Description
[Technical field]
[0001] The present disclosure relates to an attack detection method, an attack detection system, and a program for detecting attacks against sensor data. [Background technology]
[0002] Systems that monitor automobiles (monitoring systems) are required to detect attacks (attack detection) on vehicle sensor data that is transmitted and received for the purpose of vehicle control. A monitoring system is also characterized by the large scale of the objects that it monitors. For this reason, it is necessary to reduce monitoring costs, such as the cost of communication between the automobile and the system or the computational cost of attack detection. As a monitoring system with reduced computational costs, a technology has been disclosed that transmits a sampling log with reduced data volume to the monitoring system by extracting feature values such as peak values from the results of frequency analysis of the sensor data that is the object of monitoring (see, for example, Patent Document 1). [Prior art documents] [Patent documents]
[0003] [Patent Document 1] JP 2018-164159 A [Patent Document 2] International Publication No. 2019 / 187350 [Non-patent literature]
[0004] [Non-Patent Document 1] P.Malhotra, “Long Short Term Memory Networks for Anomaly Detection in Time Series”, 23rd European Symposium on Artificial Neural Networks, Computational Intelligence and Machine Learning, ESANN 2015. Summary of the Invention [Problem to be solved by the invention]
[0005] However, in the technology described in Patent Document 1, the sampled data is data at a unique time interval determined by the monitoring system in order to reduce the amount of data. Therefore, the sampled data may not contain attack messages, and attacks may be overlooked.
[0006] There is also known a technique for detecting attacks (attack messages) from sampled messages by using deviations from a normal cycle of sensor data (see, for example, Patent Document 2). However, even with this technique, it is difficult to accurately detect attacks depending on the acquisition interval of attack messages.
[0007] Therefore, it is desirable to achieve both a reduction in data volume and high attack detection performance, but as described above, with conventional techniques, it is difficult to achieve both a reduction in data volume and high attack detection performance.
[0008] Therefore, the present disclosure provides an attack detection method, an attack detection system, and a program that can achieve both data volume reduction and attack detection performance. [Means for solving the problem]
[0009] An attack detection method according to one aspect of the present disclosure is an attack detection method for detecting an attack against sensor data transmitted and received within a mobility for controlling the mobility, comprising: a sampling policy determination step for determining a sampling policy including a sampling interval and a sampling time for each sampling interval, in which the sensor values for detecting the attack are extracted from the sensor data based on at least one of a statistical value indicating a variation in the sensor values of the sensor data and event information of the mobility indicating a timing of a change in the sensor values, for each identifier of a sensor value included in the sensor data; and a sampling policy determination step for determining a sampling policy including a sampling interval and a sampling time for each sampling interval, and extracting two or more sensor values extracted from the sensor data based on the sampling interval and the sampling time, first order information indicating a temporal order of the two or more sensor values during the sampling time, and second order information indicating a temporal order of the two or more sensor values for each sampling interval. and an attack determination step of calculating a first anomaly score indicating the degree of anomaly of the evaluation target data based on the sensor values of the evaluation target data included in the sampling data, the first order information, and a short-term sensor flow indicating time series data of sensor values older than the evaluation target data at the sampling time when the evaluation target data was acquired, calculating a second anomaly score indicating the degree of anomaly of the evaluation target data based on the sensor values of the evaluation target data, the second order information, and a long-term sensor flow indicating changes in sensor values acquired at a sampling interval older than the sampling interval when the evaluation target data was acquired, and determining whether the evaluation target data is the result of the attack based on the calculated first anomaly score and second anomaly score, and outputting a determination result.
[0010] Moreover, an attack detection system according to one embodiment of the present disclosure is an attack detection system that detects attacks against sensor data transmitted and received within a mobility for controlling the mobility, and includes a sampling policy determination unit that determines a sampling policy including a sampling interval and a sampling time for each sampling interval, and extracts the sensor value for detecting the attack from the sensor data based on at least one of a statistical value indicating a variation in the sensor value of the sensor data and event information of the mobility indicating a timing of a change in the sensor value for each identifier of a sensor value included in the sensor data, and a sampling policy determination unit that determines a sampling policy including a sampling interval and a sampling time for each sampling interval, and extracts two or more sensor values extracted from the sensor data based on the sampling interval and the sampling time, first order information indicating a temporal order of the two or more sensor values during the sampling time, and a temporal order of the two or more sensor values for each sampling interval. and an attack determination unit that calculates a first anomaly score indicating a degree of anomaly of the evaluation target data based on sensor values of the evaluation target data included in the sampling data, the first order information, and a short-term sensor flow indicating time-series data of sensor values in the past than the sampling interval at which the evaluation target data was acquired, calculates a second anomaly score indicating a degree of anomaly of the evaluation target data based on the sensor values of the evaluation target data, the second order information, and a long-term sensor flow indicating changes in sensor values acquired at the sampling interval prior to the sampling interval at which the evaluation target data was acquired, and determines whether the evaluation target data is the result of the attack based on the calculated first anomaly score and second anomaly score, and outputs a determination result.
[0011] Moreover, a program according to one aspect of the present disclosure is a program for causing a computer to execute the above-described attack detection method. Effect of the Invention
[0012] According to one aspect of the present disclosure, it is possible to provide an attack detection method, an attack detection system, and a program that can achieve both data volume reduction and attack detection performance. [Brief description of the drawings]
[0013] [Figure 1] FIG. 1 is a block diagram showing a configuration of a monitoring system according to an embodiment. [Diagram 2] FIG. 2 is a flowchart showing an outline of the operation of the monitoring system according to the embodiment. [Diagram 3] FIG. 3 is a block diagram showing a configuration of the sampling strategy generating unit according to the embodiment. [Figure 4] FIG. 4 is a flowchart showing a sampling strategy generation process according to the embodiment. [Diagram 5] FIG. 5 is a diagram illustrating an example of vehicle sensor data according to the embodiment. [Figure 6] FIG. 6 is a diagram illustrating sampling according to the embodiment. [Figure 7] FIG. 7 is a diagram illustrating an example of sampling policy data according to the embodiment. [Figure 8] FIG. 8 is a block diagram illustrating a configuration of a sampling data generating unit according to the embodiment. [Figure 9] FIG. 9 is a flowchart showing the sampling data generation process according to the embodiment. [Figure 10] FIG. 10 is a diagram illustrating an example of sampling data according to the embodiment. [Figure 11] FIG. 11 is a block diagram illustrating a configuration of an attack detection unit according to the embodiment. [Figure 12] FIG. 12 is a flowchart showing an attack detection process according to the embodiment. [Figure 13] FIG. 13 is a diagram illustrating an example of base point data according to the embodiment. [Figure 14] FIG. 14 is a diagram illustrating an example of a detection result according to the embodiment. [Figure 15]FIG. 15 is a block diagram illustrating a configuration of a base point data updating unit according to the embodiment. [Figure 16] FIG. 16 is a flowchart showing the base point data update process according to the embodiment. [Figure 17] FIG. 17 is a schematic diagram showing a configuration of a monitoring system according to an embodiment. [Figure 18] FIG. 18 is a schematic diagram showing a configuration of an in-vehicle device according to an embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0014] (Findings on which this disclosure is based) In order to realize a monitoring system that can be implemented at low monitoring costs, it is necessary to develop both an appropriate sampling method and an attack detection method.
[0015] In Patent Document 1, in order to reduce the monitoring cost in a monitoring system, low-density data is generated by extracting representative values such as peak values at sampling intervals based on the results of frequency analysis of periodic sensor data. Therefore, it is assumed that the sensor values can be acquired periodically or have regular values.
[0016] However, an attacker may significantly disrupt the acquisition period or the regularity of the sensor values. In Patent Document 1, a representative value is simply extracted, so that the attack value injected by the attacker may be missing during sampling of periodic sensor data, making it difficult to detect the attack. In other words, the attack value injected by the attacker may not be sampled during sampling. Therefore, it is required to develop a further sampling method suitable for attack detection.
[0017] In Patent Document 2, the attack detection method involves calculating the normal cycle of sensor data in advance, and judging that an attack message exists if the actually observed cycle is not within a range that includes a margin added to the normal cycle. The margin is a range that allows for delays caused by collisions between normal messages and attack messages. This method is robust as a method that can detect added attack messages, etc.
[0018] However, in Patent Document 2, the presence or absence of an attack is determined based on the acquisition time of a message, so it is difficult to apply this to data in which the acquisition interval of a single sensor value becomes the sampling interval due to sampling of sensor data.
[0019] In addition, Non-Patent Document 1 discloses a technique for detecting attacks in which normal behavior of sensor values is learned in advance, and if the result of predicting the next sensor value based on a list of actually observed sensor values deviates from the normal behavior, it is determined to be an attack message. This technique can also be applied to time series data with a time interval that is longer than the period, such as the sampling interval.
[0020] However, the list of actually observed sensor values for predicting the next sensor value must be filled up so that it is not contaminated by attack messages.
[0021] Therefore, in order to solve the above problems, further improvements to the attack detection method are necessary in line with the construction of the sampling method. In other words, it is desirable to be able to achieve both data volume reduction and attack detection performance. Therefore, the inventors of the present application have conducted extensive research into attack detection methods and the like that can achieve both data volume reduction and attack detection performance, and have devised the attack detection method and the like described below.
[0022] An attack detection method according to one aspect of the present disclosure is an attack detection method for detecting an attack against sensor data transmitted and received within a mobility for controlling the mobility, comprising: a sampling policy determination step for determining a sampling policy including a sampling interval and a sampling time for each sampling interval, in which the sensor values for detecting the attack are extracted from the sensor data based on at least one of a statistical value indicating a variation in the sensor values of the sensor data and event information of the mobility indicating a timing of a change in the sensor values, for each identifier of a sensor value included in the sensor data; and a sampling policy determination step for determining a sampling policy including a sampling interval and a sampling time for each sampling interval, and extracting two or more sensor values extracted from the sensor data based on the sampling interval and the sampling time, first order information indicating a temporal order of the two or more sensor values during the sampling time, and second order information indicating a temporal order of the two or more sensor values for each sampling interval. and an attack determination step of calculating a first anomaly score indicating the degree of anomaly of the evaluation target data based on the sensor values of the evaluation target data included in the sampling data, the first order information, and a short-term sensor flow indicating time series data of sensor values older than the evaluation target data at the sampling time when the evaluation target data was acquired, calculating a second anomaly score indicating the degree of anomaly of the evaluation target data based on the sensor values of the evaluation target data, the second order information, and a long-term sensor flow indicating changes in sensor values acquired at a sampling interval older than the sampling interval when the evaluation target data was acquired, and determining whether the evaluation target data is the result of the attack based on the calculated first anomaly score and second anomaly score, and outputting a determination result.
[0023] According to this, sampling data is generated by adjusting the sampling interval and sampling time according to the characteristics of the sensor data to be subjected to the attack detection process. Furthermore, an attack is determined by calculating two types of anomaly scores based on sampling data including two types of order information (e.g., indexes) indicating the temporal order of the sensor values at the sampling time (short term) and the sampling interval (long term). For example, the amount of data can be reduced by setting at least one of the sampling interval and the sampling time to be long. Furthermore, by calculating two types of anomaly scores, attacks can be detected more accurately than when only one type of anomaly score is calculated. This allows attacks to be detected accurately while suppressing monitoring costs. Therefore, it is possible to achieve both data amount reduction and attack detection performance.
[0024] Also, for example, the statistical value may include a variance value indicating the variation in the sensor value for each identifier, and the event information may include flag information indicating whether a change has occurred in the sensor value, and in the sampling policy determination step, the sampling interval and the sampling time may be determined based on the variance value and the flag information.
[0025] According to this, since both the variance value of the sensor value and the flag information are used, the sampling interval and the sampling time can be determined more accurately.
[0026] Also, for example, in the sampling policy determination step, the sampling time and the sampling interval may be determined according to a combination of a determination result of whether or not the variance value is equal to or greater than a first threshold value and whether or not the flag information indicates that no change has occurred in the sensor value.
[0027] According to this, the sampling interval and the sampling time can be determined by combining the judgment result of the variance value of the sensor data and the contents of the event information. Therefore, an appropriate sampling interval and sampling time can be determined based on the judgment result. For example, if the variance value is not equal to or greater than the first threshold, the sensor value is stable, so that at least one of the sampling interval and the sampling time can be determined to be long. This makes it possible to appropriately reduce the amount of sensor data to be extracted.
[0028] Also, for example, the flag information may include information indicating whether at least one of the following has occurred: a drive unit of the mobility being switched ON and OFF, a shift lever being switched, or an autonomous driving mode being switched ON and OFF.
[0029] This allows the sampling interval and sampling time to be determined according to the driving unit, the shift lever, and the switching of the autonomous driving mode of the mobility. For example, at least one of the sampling interval and the sampling time after a predetermined time has elapsed since the switching can be determined to be a long time. This makes it possible to more appropriately reduce the amount of sensor data.
[0030] Also, for example, in the sampling data generation step, a number of sensor data to be extracted within the sampling time may be determined based on a number of normally received data to be observed during the sampling time, and the determined number of sensor data may be extracted.
[0031] This makes it possible to more appropriately reduce the amount of sensor data based on the number of observed sensor data items.
[0032] Also, for example, in the sampling data generation step, it may be determined whether the number of sensor values observed during the sampling time is equal to or greater than the number of normally received data, and if it is equal to or greater than the number of normally received data, a predetermined number of the sensor values may be extracted, and if it is not equal to or greater than the number of normally received data, all of the sensor values included in the sensor data may be extracted.
[0033] According to this, when the sensor data includes attack data, sampling can be performed while suppressing loss of the attack data, thereby enabling attacks to be detected with high accuracy for the sampled data.
[0034] Furthermore, for example, the first order information is a first index indicating the order in which the two or more sensor values were acquired during the sampling time, and the second order information is a second index indicating the order of the sampling interval in which the two or more sensor values were acquired. In the attack determination step, if the first index corresponding to the evaluation target data is less than a third threshold and the second index corresponding to the evaluation target data is less than a fourth threshold, it is determined that it is unclear whether the evaluation target data is due to the attack, if a third abnormality score based on the first abnormality score and the second abnormality score is greater than a fifth threshold, it is determined that the evaluation target data is due to the attack, and if the third abnormality score is less than the fifth threshold, it is determined that the evaluation target data is normal, and the determination result corresponding to the evaluation target data indicating the determination result is output.
[0035] According to this, it is possible to perform an integrated judgment based on short-term and long-term changes in the sensor values while reducing evaluation target data in which attack detection is difficult in the sampling data. Therefore, attacks can be detected with high accuracy in the sampling data.
[0036] Also, for example, the method may further include an update step of updating the short-term sensor flow and the long-term sensor flow based on the judgment result corresponding to the evaluation target data, the first order information, and the second order information.
[0037] According to this, it is possible to update the short-term sensor flow and the long-term sensor flow used for determining attack detection. By using the updated short-term sensor flow and the long-term sensor flow, it is possible to improve the attack detection performance.
[0038] Also, for example, in the updating step, if the judgment result corresponding to the evaluation target data is an attack, a sensor value that the evaluation target data should take may be estimated using at least one of the short-term sensor flow and the long-term sensor flow, and the short-term sensor flow and the long-term sensor flow may be updated based on the estimated sensor value, the first order information, and the second order information; and if the judgment result corresponding to the evaluation target data is normal or unknown, the short-term sensor flow and the long-term sensor flow may be updated based on the sensor value of the evaluation target data, the first order information, and the second order information.
[0039] According to this, even if the sampling data includes attack data, the attack data can be removed and the short-term sensor flow and the long-term sensor flow can be updated to appropriate ones. In other words, the short-term sensor flow and the long-term sensor flow can be prevented from including attack data. Therefore, attacks can be detected more accurately from the sampling data.
[0040] Also, for example, the sensor data may be the result of decoding byte-represented network data that is stored in a device installed in the mobility and transmitted and received at any travel time interval according to at least one of the protocols CAN (Controller Area Network: registered trademark), CAN FD (CAN with Flexible Data-rate), FlexRay (registered trademark), and EtherNet (registered trademark).
[0041] This makes it possible to achieve both data volume reduction and attack detection performance in networks that communicate according to at least one of the protocols CAN, CAN FD, FlexRay, and EtherNet.
[0042] Also, for example, the mobility may be a vehicle.
[0043] This makes it possible to reduce the amount of data in a vehicle with limited computational resources while suppressing a decrease in the performance of detecting attacks against the vehicle.
[0044] Moreover, an attack detection system according to one embodiment of the present disclosure is an attack detection system that detects attacks against sensor data transmitted and received within a mobility for controlling the mobility, and includes a sampling policy determination unit that determines a sampling policy including a sampling interval and a sampling time for each sampling interval, and extracts the sensor value for detecting the attack from the sensor data based on at least one of a statistical value indicating a variation in the sensor value of the sensor data and event information of the mobility indicating a timing of a change in the sensor value for each identifier of a sensor value included in the sensor data, and a sampling policy determination unit that determines a sampling policy including a sampling interval and a sampling time for each sampling interval, and extracts two or more sensor values extracted from the sensor data based on the sampling interval and the sampling time, first order information indicating a temporal order of the two or more sensor values during the sampling time, and a temporal order of the two or more sensor values for each sampling interval. and an attack determination unit that calculates a first anomaly score indicating a degree of anomaly in the evaluation target data based on a sensor value of the evaluation target data included in the sampling data, the first order information, and a short-term sensor flow indicating time-series data of sensor values older than the evaluation target data at the sampling time when the evaluation target data is acquired, calculates a second anomaly score indicating a degree of anomaly in the evaluation target data based on the sensor value of the evaluation target data, the second order information, and a long-term sensor flow indicating a change in sensor values acquired at a sampling interval older than the sampling interval when the evaluation target data is acquired, and determines whether the evaluation target data is due to the attack based on the calculated first anomaly score and the second anomaly score, and outputs a determination result.
[0045] This provides the same effect as the attack detection method described above.
[0046] Furthermore, these general or specific aspects may be realized by a system, a method, an integrated circuit, a computer program, or a computer-readable recording medium such as a CD-ROM, or may be realized by any combination of the system, method, integrated circuit, computer program, or recording medium.
[0047] Hereinafter, a monitoring method (one example of an attack detection method) according to an embodiment will be described with reference to the drawings. The embodiments shown here are all comprehensive or specific examples. Therefore, the numerical values, components, arrangement and connection forms of the components, steps and the order of steps shown in the following embodiments are merely examples and do not limit the invention according to the present disclosure. Furthermore, among the components in the following embodiments, components that are not described in the independent claims are components that can be added arbitrarily. Furthermore, each figure is a schematic diagram and is not necessarily a precise illustration.
[0048] (Embodiment) First, the configuration of a monitoring system according to the present embodiment will be described. In this embodiment, an example will be described in which the monitoring system 100 is an information processing system for detecting attacks against sensor data (vehicle sensor data) transmitted and received inside a vehicle for vehicle control. A vehicle is an example of mobility against which the monitoring system 100 detects attacks. The monitoring system 100 is also an example of an attack detection system.
[0049] [1-1. Monitoring system configuration] FIG. 1 is a block diagram showing a configuration of a monitoring system 100 according to this embodiment.
[0050] The monitoring system 100 is a system for determining whether each evaluation target data in vehicle sensor data (see FIG. 5) is an attack. The monitoring system 100 determines whether the evaluation target data is the result of an attack. Note that determining whether the evaluation target data is the result of an attack is also referred to as determining whether the evaluation target data is attack data. The evaluation target data can also be said to be a sensor value that is the target for determining whether it is attack data.
[0051] The vehicle sensor data is stored in equipment installed in the vehicle and is the result of decoding network data expressed in bytes that is transmitted and received at any driving time interval (an example of a travel time interval) according to at least one protocol of CAN (Controller Area Network: registered trademark), CAN FD (CAN with Flexible Data-rate), FlexRay (registered trademark), and EtherNet (registered trademark).
[0052] The monitoring system 100 includes a sampling unit 101 and a sampling data monitoring unit 102. The sampling unit 101 is mounted on, for example, a vehicle, and the sampling data monitoring unit 102 is disposed, for example, in a location different from the vehicle (for example, a monitoring center in a remote location). The sampling unit 101 and the sampling data monitoring unit 102 are connected to each other so as to be able to communicate with each other (for example, wirelessly).
[0053] The sampling unit 101 accumulates vehicle sensor data transmitted and received for vehicle control of the vehicle, extracts a sensor value to be used for monitoring in the sampling data monitoring unit 102 from the accumulated multiple sensor values (vehicle sensor data) based on a sampling policy, and transmits sampling data including the extracted sensor value to the sampling data monitoring unit 102. The sampling policy is dynamically determined by the sampling unit 101. For example, the sampling unit 101 determines the sampling policy based on at least one of a statistical value indicating the variation of the sensor value of the vehicle sensor data and vehicle event information indicating the timing of a change in the characteristic of the sensor value (for example, a change in the sensor value) for each identifier of the sensor value included in the vehicle sensor data (hereinafter, also referred to as a sensor identifier). The sampling data includes, for example, two or more sensor values. The sensor identifier will be described later, and is information for uniquely identifying vehicle sensor data such as "vehicle speed" and "shift".
[0054] The vehicle sensor data includes sensor values (e.g., sensing results such as vehicle speed, acceleration, and steering angle) of various sensors mounted on the vehicle. The sampling policy will be described later (see FIG. 6), and includes a sampling interval and a sampling time for each sampling interval.
[0055] The vehicle event information includes information about non-periodic vehicle events, such as flag information indicating whether a change in a sensor value has occurred. The flag information includes, for example, information indicating whether at least one of the following has occurred: switching the vehicle engine (an example of a driving unit of a mobility) ON and OFF, switching the shift lever, and switching the autonomous driving mode ON and OFF. For example, the flag information may include "1" indicating "yes" when a vehicle event has occurred, and "0" indicating "no" when no vehicle event has occurred. The monitoring also includes a process of detecting cyber attacks on the vehicle. Note that switching the vehicle engine ON and OFF may be, for example, the engine being restarted.
[0056] The sampling unit 101 includes a sampling strategy generating unit 110 , a sampling data generating unit 111 , a vehicle sensor data storage unit 112 , a sampling strategy storage unit 113 , a vehicle information storage unit 114 , and a communication unit 115 .
[0057] The sampling strategy generating unit 110 is a processing unit that determines a sampling strategy for generating sampling data based on at least one of vehicle sensor data and vehicle events. In this embodiment, the sampling strategy generating unit 110 determines a sampling strategy based on both vehicle sensor data and vehicle events. The sampling strategy generating unit 110 determines a sampling strategy based on, for example, a statistical value indicating the variance of the sensor values of the vehicle sensor data and the timing of the occurrence of the vehicle event. The sampling strategy generating unit 110 is an example of a sampling strategy determining unit.
[0058] The statistical value of the sensor values may use any index that indicates the variance of the sensor values. In this embodiment, the statistical value of the sensor values uses a variance value (data variance) that indicates the variance of the sensor values for each identifier. The variance value is a value that indicates the degree of dispersion of the sensor values, and is calculated by squaring the deviation (the difference between each sensor value and the average value) and taking the average. The variance value may be, for example, a standard deviation.
[0059] The timing of the occurrence of a vehicle event indicates the timing when the vehicle event occurs, and can be said to indicate, for example, the timing of a change in the characteristic of the sensor value, or the presence or absence of the occurrence of the vehicle event.
[0060] The sampling policy generating unit 110 is, for example, a processing unit that determines a sampling policy for each sensor identifier of a sensor value included in the vehicle sensor data. For example, the sampling policy generating unit 110 determines a sampling policy for a sensor identifier based on a plurality of sensor values corresponding to the sensor identifier. The sampling policy is a sampling policy for generating sampling data from the plurality of sensor values corresponding to the sensor identifier.
[0061] The sampling data generating unit 111 is a processing unit that generates sampling data to be transmitted to the sampling data monitoring unit 102 based on the sampling policy determined by the sampling policy generating unit 110. For example, the sampling data generating unit 111 extracts two or more sensor values from a plurality of sensor values stored in the vehicle sensor data storage unit 112 based on the determined sampling policy, thereby generating sampling data including the extracted two or more sensor values.
[0062] The sampling data generating unit 111 generates sampling data including two or more sensor values extracted from a plurality of sensor values (vehicle sensor data) based on, for example, a sampling interval and a sampling time, a first index indicating the temporal order of the two or more sensor values during the sampling time, and a second index indicating the temporal order of the two or more sensor values for each sampling interval. The first index and the second index are assigned to each sensor value.
[0063] The first index indicates the order in which two or more sensor values were acquired (observed) during a sampling time, and can also be said to indicate the order of short-term movement (change) of the sensor values during the sampling time. The second index indicates the order of sampling intervals in which two or more sensor values were acquired (observed), and can also be said to indicate the order of long-term movement (change) of the sensor values during the sampling intervals. The first index is an example of first order information, and the second index is an example of second order information.
[0064] The vehicle sensor data storage unit 112 is a storage device that stores a plurality of sensor values (vehicle sensor data) transmitted and received for vehicle control of the vehicle. The vehicle sensor data storage unit 112 accumulates the plurality of sensor values.
[0065] The sampling strategy storage unit 113 is a storage device that stores the sampling strategy determined by the sampling strategy generation unit 110 .
[0066] The vehicle information storage unit 114 is a storage device that stores vehicle events. The vehicle information storage unit 114 stores vehicle event information in which the content of a vehicle event is associated with the time when the vehicle event occurred, for example.
[0067] The communication unit 115 is a processing unit having a function of communicating with the sampling data monitoring unit 102 via a public communication network such as the Internet. The communication unit 115 includes, for example, a communication circuit (communication module).
[0068] Next, the sampling data monitoring unit 102 monitors the vehicle based on the sampling data acquired from the sampling unit 101. Specifically, the sampling data monitoring unit 102 calculates a first anomaly score indicating a short-term anomaly and a second anomaly score indicating a long-term anomaly based on the sampling data, and determines whether or not a sensor value to be evaluated (hereinafter, also referred to as evaluation target data) is due to an attack based on the two calculated anomaly scores. That is, the sampling data monitoring unit 102 determines whether or not the evaluation target data is attack data. It can also be said that the sampling data monitoring unit 102 determines whether or not the vehicle is under a cyber attack. The sampling data monitoring unit 102 is characterized in that it detects an attack based on two anomaly scores.
[0069] A short-term abnormality refers to an abnormality detected based on vehicle sensor data acquired over a short period of time, for example, an abnormality detected based on vehicle sensor data acquired within one sampling time, whereas a long-term abnormality refers to an abnormality detected based on vehicle sensor data acquired over a longer period of time than the short period of time, for example, an abnormality detected based on vehicle sensor data acquired at multiple sampling intervals.
[0070] The sampling data monitoring unit 102 includes an attack detection unit 120 , a base data update unit 121 , a detector storage unit 122 , a detection result storage unit 123 , a base data storage unit 124 , and a sampling data storage unit 125 .
[0071] The attack detection unit 120 is a processing unit that determines whether or not the evaluation target data is attack data based on the sampling data acquired from the sampling unit 101. The attack detection unit 120 calculates a first anomaly score and a second anomaly score, determines whether or not the evaluation target data is due to an attack based on the calculated first anomaly score and second anomaly score, and outputs the determination result. The calculation of the first anomaly score and the second anomaly score will be described later.
[0072] The base data update unit 121 is a processing unit that updates the short-term sensor flow and the long-term sensor flow based on detection result information indicating the detection result of the attack detection unit 120 corresponding to the evaluation target data and the first index and the second index in the extracted vehicle sensor data.
[0073] The short-term sensor flow is time-series data of sensor values used to detect short-term abnormalities, and is generated, for example, based on sensor values acquired within one sampling time. The short-term sensor flow is generated, for example, based on sensor values within a sampling time when evaluation target data is acquired. The short-term sensor flow can also be said to indicate the movement of short-term sensor values that are determined to be normal or estimated to be normal.
[0074] The long-term sensor flow is time series data of sensor values used to detect long-term abnormalities longer than the short-term sensor flow, and is generated, for example, based on sensor values at multiple sampling intervals. The long-term sensor flow can also be said to indicate long-term movement of sensor values that are determined to be normal or estimated to be normal.
[0075] The data including the short-term sensor flow and the long-term sensor flow is also referred to as base data. The base data does not include sensor values that the attack detection unit 120 judges to be abnormal (attack). The base data can also be said to indicate time series data of sensor values that are judged to be normal or estimated to be normal.
[0076] The detector storage unit 122 is a storage device that stores a first detector for calculating a first anomaly score indicating the degree of deviation between the short-term sensor flow and the evaluation target data, and a second detector for calculating a second anomaly score indicating the degree of deviation between the long-term sensor flow and the evaluation target data. The first detector is realized by a trained model that is machine-learned to calculate a first anomaly score using, for example, the short-term sensor flow and the evaluation target data as inputs. The second detector is realized by a trained model that is machine-learned to calculate a second anomaly score using, for example, the long-term sensor flow and the evaluation target data as inputs. Note that storing a trained model means storing information such as network parameters in the machine learning model and a calculation algorithm (machine learning algorithm).
[0077] The detection result storage unit 123 is a storage device that stores the detection results of the attack detection unit 120.
[0078] The base data storage unit 124 is a storage device that stores the short-term sensor flows and long-term sensor flows (base data) that the attack detection unit 120 uses to determine the presence or absence of an attack.
[0079] The sampling data storage unit 125 is a storage device that stores the sampling data acquired from the sampling unit 101 .
[0080] The communication unit 126 is a processing unit having a function of communicating with the sampling unit 101 via a public communication network such as the Internet. The communication unit 126 includes, for example, a communication circuit (communication module).
[0081] The vehicle sensor data memory unit 112, the sampling policy memory unit 113, the vehicle information memory unit 114, the detector memory unit 122, the detection result memory unit 123, the base point data memory unit 124, and the sampling data memory unit 125 are composed of rewritable non-volatile memory such as a hard disk drive or a solid state drive.
[0082] The sampling policy generating unit 110, the sampling data generating unit 111, the communication unit 115, the attack detecting unit 120, the base data updating unit 121, and the communication unit 126 are configured with a processor such as a CPU, an ASIC, or an FPGA (Field Programmable Gate Array). Each functional block of the monitoring system 100 is realized by a processor such as a CPU executing a computer-readable program stored in a computer.
[0083] The communication method between the sampling unit 101 and the sampling data monitoring unit 102 is not particularly limited. When the communication method is wireless communication, the communication between the sampling unit 101 and the sampling data monitoring unit 102 may be performed by infrared communication or radio wave communication. Radio wave communication may be communication by Wi-Fi (registered trademark) (Wireless Fidelity) or Bluetooth (registered trademark). Also, for example, the communication between the sampling unit 101 and the sampling data monitoring unit 102 may be performed by wired communication.
[0084] When the sampling unit 101 and the sampling data monitoring unit 102 are configured as an integrated unit, communication between the sampling unit 101 and the sampling data monitoring unit 102 does not need to go through the communication units 115 and 126, respectively. In this case, the configuration of the communication units 115 and 126 is not necessary.
[0085] [1-2. Overview of the monitoring system operation] Next, the operation of the monitoring system 100 configured as above will be described. Fig. 2 is a flowchart showing an outline of the operation (attack detection method) of the monitoring system 100 according to this embodiment. Steps S101 and S102 shown in Fig. 2 are processes executed in the sampling unit 101, for example, and steps S103 and S104 are processes executed in the sampling data monitoring unit 102, for example.
[0086] First, the monitoring system 100 performs a sampling policy generation process (S101) for generating a sampling policy indicating a sampling interval and a sampling time, using vehicle sensor data for a predetermined driving time of the vehicle stored in the vehicle sensor data storage unit 112. The process of step S101 is performed by the sampling policy generation unit 110. The predetermined driving time may be set in advance, for example, or may be set by a monitor who monitors the vehicle using the monitoring system 100. Note that step S101 is an example of a sampling policy determination step.
[0087] Next, the monitoring system 100 performs a sampling data generation process to generate sampling data from the vehicle sensor data based on the determined sampling policy (S102). The process of step S102 is performed by the sampling data generation unit 111. Note that step S102 is an example of a sampling data generation step.
[0088] The sampling data generating unit 111 extracts a part of the vehicle sensor data (sensor values) from the vehicle sensor data based on the sampling policy, and generates the sampling data by adding a first index and a second index to each of the extracted vehicle sensor data (sensor values). That is, the sampling data includes the extracted vehicle sensor data, the first index, and the second index.
[0089] Then, the sampling data generating unit 111 transmits the generated sampling data to the communication unit 126 via the communication unit 115. That is, the sampling data generating unit 111 transmits the generated sampling data to the sampling data monitoring unit .
[0090] Next, the monitoring system 100 performs an attack detection process to calculate an anomaly score for the vehicle sensor data (evaluation target data) of the sampling data (S103). The process of step S103 is performed by the attack detection unit 120. Step S103 is an example of an attack determination step.
[0091] The attack detection unit 120 uses the base data stored in the base data storage unit 124 to calculate an anomaly score indicating the degree of abnormality in the sensor value of the evaluation target data using two different calculation methods, and performs an attack determination process using the two calculated anomaly scores to determine whether the sensor value of the evaluation target data is an attack or not.
[0092] Next, the monitoring system 100 performs a base data update process to update the base data for the attack detection process on the next evaluation target data, using the first index, the second index, and the result of the attack detection process (judgment result) of the evaluation target data (S104). The process of step S104 is performed by the base data update unit 121. Step S104 is an example of an update step.
[0093] The detailed configuration of each processing unit and the detailed operation of each step in the flowchart will be described below.
[0094] [2-1. Configuration of the sampling policy generation unit] 3 is a block diagram showing the configuration of the sampling strategy generating unit 110 according to this embodiment. As shown in FIG. 3, the sampling strategy generating unit 110 includes a vehicle sensor data acquiring unit 130, a sensor data stability calculating unit 131, and a sampling strategy determining unit 132.
[0095] The vehicle sensor data acquisition unit 130 is a processing unit that acquires vehicle sampling data from the vehicle sensor data storage unit 112. The vehicle sensor data acquisition unit 130 reads out the vehicle sensor data stored in the vehicle sensor data storage unit 112.
[0096] The sensor data stability calculation unit 131 is a processing unit that calculates information used for determining a sampling policy in the sampling policy determination unit 132, based on the vehicle sensor data. For example, the sensor data stability calculation unit 131 calculates a statistical value of the sensor value for each sensor identifier. In this embodiment, the sensor data stability calculation unit 131 calculates a variance value of the sensor value for each sensor identifier.
[0097] The sensor identifiers include periodic identifiers that are periodically transmitted and received, such as vehicle speed, and event identifiers that are generated in response to operations by the driver, such as shifting gears.
[0098] The sampling policy determination unit 132 determines a sampling policy for each sensor identifier included in the vehicle sensor data by using at least one of the statistical values and the vehicle event information. In this embodiment, the sampling policy determination unit 132 determines a sampling policy by using both the statistical values and the vehicle event information.
[0099] [2-2. Operation of the Sampling Policy Generation Unit] FIG. 4 is a flowchart showing the sampling strategy generation process (S101) according to this embodiment.
[0100] First, the sampling strategy generating unit 110 acquires vehicle sensor data for a predetermined driving time of the vehicle stored in the vehicle sensor data storage unit 112 (S111). The process of step S111 is performed by the vehicle sensor data acquiring unit 130. The vehicle sensor data acquiring unit 130 acquires the vehicle sensor data from the vehicle sensor data storage unit 112, for example.
[0101] Here, the vehicle sensor data will be described. Fig. 5 is a diagram showing an example of the vehicle sensor data according to the present embodiment. The vehicle sensor data is data obtained by extracting information that can be input to the attack detection process from the vehicle network data expressed in binary. As shown in Fig. 5, the vehicle sensor data includes a sensor identifier 142, a data ID 143, an acquisition time 144, and a sensor value 145.
[0102] The sensor identifier 142 is an identifier for uniquely identifying a vehicle control target of one piece of sensor data included in the vehicle sensor data. For example, the sensor identifier 142 includes "vehicle speed", "shift", and the like.
[0103] The data ID 143 is an identifier for uniquely identifying one piece of vehicle sensor data 141 for each of the sensor identifiers included in the vehicle sensor data. For example, the data ID 143 includes a numerical value assigned in sequence starting from 1. For example, the data ID 143 is assigned in sequence starting from 1 for each sensor identifier. Also, the data ID 143 is assigned in sequence, for example, of the acquisition time.
[0104] The acquisition time 144 is time data indicating the time when the vehicle sensor data was received from the vehicle network. For example, when the sensor identifier 142 is "vehicle speed", the acquisition time 144 indicates a periodic time, and when the sensor identifier 142 is an identifier that may have an event occurrence such as "shift", the acquisition time 144 may also be a non-periodic time. For example, the acquisition time 144 can be obtained according to a timestamp of a message that transmits the vehicle sensor data.
[0105] The acquisition time 144 may be the time when the sampling unit 101 receives the vehicle sensor data from the vehicle network. Although the acquisition time 144 is an example of an elapsed time from a reference time (relative time), the acquisition time 144 is not limited to this and may be an absolute time.
[0106] The sensor value 145 is numerical data indicating vehicle control content corresponding to the sensor identifier 142. A plurality of sensor values are included in the vehicle sensor data.
[0107] 5, for example, when the sensor identifier 142 is "vehicle speed", the sensor value 145 is "22.11 km / h" when the acquisition time 144 is "25.00 ms", and the sensor value 145 is "22.50 km / h" when the next acquisition time 144 is "50.01 ms", and it can be seen that the reception cycle of the vehicle speed is about 25 ms. In the following, an example in which the reception cycle of the vehicle speed is about 25 ms will be described, but the reception cycle is not limited to this.
[0108] Returning to the description of the sampling strategy generation process, referring again to Fig. 4, the sampling strategy generation unit 110 calculates, for each sensor identifier 142 included in the vehicle sensor data, data variance (e.g., variance value) indicating the variation of the sensor value 145, and vehicle event information (e.g., flag information) indicating whether or not a vehicle event indicating an event-type sensor identifier 142 included in the vehicle sensor data has occurred (S112). The event-type sensor identifier 142 includes, for example, switching between engine ON and OFF, and shift switching. The process of step S112 is performed by the sensor data stability calculation unit 131.
[0109] Next, the sampling strategy generating unit 110 judges whether the vehicle event information (flag information) is “present” or not, or whether the data variance is equal to or greater than a first threshold value (S113). The process of step S113 is performed by the sampling strategy determining unit 132.
[0110] If the data variance is equal to or greater than the first threshold, or if the vehicle event flag information is "present", that is, if a vehicle event has occurred (Yes in S113), the sampling policy determination unit 132 sets the sampling policy to high-density sampling (high-density sampling) (S114). For example, if the data variance is equal to or greater than the first threshold, the stability of the vehicle sensor data is low (e.g., the variation is large), so high-density sampling is performed. This is because, if the data variance is equal to or greater than the first threshold, the reproducibility of the determination result of the presence or absence of an attack after sampling is low. Also, for example, it is expected that the stability of the vehicle sensor data is low for a certain period of time immediately after the occurrence of a vehicle event, so high-density sampling is performed. This is because, if the long-term sensor flow is initialized (reset) due to the occurrence of a vehicle event, the sensor values in the long-term sensor flow are insufficient.
[0111] Furthermore, in other cases (No in S113), that is, when the data variance is less than the first threshold and the vehicle event information is "absent", the sampling policy determination unit 132 sets the sampling policy to low-density sampling (low-density sampling) (S115). For example, when the data variance is less than the first threshold, the stability of the vehicle sensor data is high (e.g., the variation is small), so low-density sampling is performed. This is because, in other cases, the reproducibility of the determination result of the presence or absence of an attack after sampling is high. Also, for example, it is expected that the stability of the vehicle sensor data is high except for a certain period of time immediately after the occurrence of a vehicle event, so low-density sampling is performed. This is because, except for the certain period of time, there are sufficient sensor values in the long-term sensor flow.
[0112] Note that low density and high density here are relative terms, and high density sampling means a sampling density higher than that of low density sampling. The sampling density is realized by increasing or decreasing at least one of the sampling time and the sampling interval. High density sampling satisfies at least one of, for example, a longer sampling time and a shorter sampling interval than low density sampling. For example, high density sampling may have a longer sampling time and a shorter sampling interval than low density sampling.
[0113] Next, the sampling policy determination unit 132 stores the set sampling policy in the sampling policy storage unit 113 (S116).
[0114] As described above, the sampling strategy generating unit 110 determines the sampling interval and the sampling time based on, for example, the variance value and the flag information. For example, the sampling strategy generating unit 110 may determine the sampling time and the sampling interval according to a combination of a determination result of whether or not the variance value is equal to or greater than a first threshold value and whether or not the flag information indicates that no change has occurred in the sensor value.
[0115] Here, the details of sampling will be described. Fig. 6 is a diagram for explaining sampling according to this embodiment. A circle shown in Fig. 6 indicates that vehicle sensor data has been acquired, and the horizontal axis in Fig. 6 indicates time. Fig. 6 shows an example in which 10 pieces of vehicle sensor data (sensor values) are acquired at a sampling interval of T seconds.
[0116] As shown in FIG. 6, the sampling interval is the interval at which vehicle sensor data is sampled, and the sampling time is the time during which sampling continues at the sampling interval. The sampling time is continuous in time, and for example, does not include discrete time. That is, the vehicle sensor data sampled at the sampling time becomes continuous data in time. The sampling start time is the time when sampling of the vehicle sensor data starts, and for example, may be the disclosure time of the sampling time. Note that the sampling start time is not limited to the start time of the sampling interval (for example, the time when measurement of the time of the sampling interval starts), and it may be before (T - S) seconds have elapsed from the start time of the sampling interval. That is, if the sampling start time is N, N may be determined to satisfy the following equation.
[0117] 0 ≦ N < T - S (Equation 1)
[0118] Sampling of the vehicle sensor data is performed during the sampling time S seconds from the Nth second of the sampling start time within the sampling interval of T seconds.
[0119] Among the vehicle sensor data during the sampling interval of T seconds, from the Nth second of the sampling start time to the sampling time of S seconds, the data is used as sampling data. That is, in FIG. 6, although 10 pieces of vehicle sensor data are acquired at the sampling interval of T seconds, an example is shown in which 3 pieces of the vehicle sensor data are extracted as sampling data. In other words, among the 10 pieces of vehicle sensor data acquired at the sampling interval of T seconds, 7 pieces of the vehicle sensor data are not transmitted to the sampling data monitoring unit 102. Here, by adjusting the sampling interval and the sampling time, high-density sampling and low-density sampling are realized. Specific examples of sampling will be described later.
[0120] FIG. 7 is a diagram showing an example of sampling policy data according to the present embodiment. In the example of FIG. 7, the sampling interval in high-density sampling is 200.00 ms, the sampling time is 75.00 ms, and the sampling interval in low-density sampling is 1000.00 ms, and the sampling time is 55.00 ms. In this way, for example, the sampling interval in high-density sampling is shorter than the sampling interval in low-density sampling, for example, about 1 / 5 of the sampling interval in low-density sampling. Also, for example, the sampling time in high-density sampling is longer than the sampling time in low-density sampling, for example, about 1.35 times the sampling interval in low-density sampling. Note that the sampling interval and sampling time in high-density sampling and the sampling interval and sampling time in low-density sampling are not limited to the above.
[0121] 7, the sampling policy data includes a sensor identifier 152, a file number 153, a data distribution 154, a vehicle event flag 155, a sampling interval 156, and a sampling time 157. The sampling policy data is stored in the sampling policy storage unit 113, and is a history in which the sensor identifier 152, information used in the determination by the sampling policy determination unit 132 (the data distribution 154 and the vehicle event flag 155), and the sampling policy (the sampling interval 156 and the sampling time 157) are associated with each other.
[0122] The sensor identifier 152 corresponds to the sensor identifier 142 of the vehicle sensor data, and is an identifier for uniquely identifying the vehicle control target.
[0123] The file number 153 is a file identifier for identifying one file unit transmitted by the communication unit 115. For example, the file number is incremented according to the vehicle sensor data acquired every predetermined driving time, and a different sampling policy is calculated.
[0124] The data variance 154 is numerical data that indicates the variance of the sensor value within a predetermined travel time, calculated for each sensor identifier 152 on a file-by-file basis.
[0125] The vehicle event flag 155 is flag information indicating whether or not a vehicle event indicating an event-type sensor identifier 142 such as switching between engine ON and OFF and shift switching has occurred, separate from the sensor identifier 152.
[0126] The sampling interval 156 indicates the sampling density calculated for each sensor identifier 152 on a file-by-file basis.
[0127] The sampling time 157 indicates the sampling density calculated for each sensor identifier 152 in the file.
[0128] Moreover, the sampling strategy generating unit 110 repeatedly executes the processes of steps S111 to S115 every time the vehicle sensor data to be processed is updated.
[0129] For example, when the file number in FIG. 7 is "2" and the sensor identifier 152 is "vehicle speed", the data variance is 0.22 and the vehicle event flag is "none". If the threshold value of the data variance is set to 10, 0.22 is below the threshold value. Therefore, in the above case, the policy is to perform low-density sampling, and for example, by setting the sampling interval to "1000 ms" so as to be longer than the above-mentioned period of "approximately 25 ms" and setting the sampling time to "55 ms" so as to be approximately twice the above-mentioned period, the amount of data to be transmitted can be significantly reduced.
[0130] The sampling policy determination unit 132 may set different sampling densities for the case where the data variance is equal to or greater than the first threshold and the case where the vehicle event flag information is "present." In other words, multiple sample densities may be set in the high-density sampling.
[0131] [3-1. Configuration of sampling data generation unit] Fig. 8 is a block diagram showing the configuration of the sampling data generating unit 111 according to this embodiment. As shown in Fig. 8, the sampling data generating unit 111 includes a vehicle sensor data acquiring unit 130, a sampling information acquiring unit 160, a data extracting unit 161, and an index assigning unit 162.
[0132] The vehicle sensor data acquisition unit 130 is a processing unit that acquires vehicle sampling data from the vehicle sensor data storage unit 112. The vehicle sensor data acquisition unit 130 reads out the vehicle sampling data stored in the vehicle sensor data storage unit 112. Note that the vehicle sensor data acquisition unit 130 may be a component common to the vehicle sensor data acquisition unit 130 included in the sampling strategy generation unit 110, or may be a component different from the vehicle sensor data acquisition unit 130.
[0133] The sampling information acquisition unit 160 is a processing unit that acquires the sampling policy determined by the sampling policy generation unit 110 from the sampling policy storage unit 113 .
[0134] The data extraction unit 161 is a processing unit that extracts vehicle sensor data based on the sampling policy acquired by the sampling policy storage unit 113 from the vehicle sensor data acquired by the vehicle sensor data acquisition unit 130.
[0135] The index assignment unit 162 is a processing unit that assigns an index indicating a time order to the vehicle sensor data extracted by the data extraction unit 161 in association with the sensor value. The index assignment unit 162 assigns both a first index and a second index to each vehicle sensor data (sensor value).
[0136] [3-2. Operation of the sampling data generator] FIG. 9 is a flowchart showing the sampling data generation process (S102) according to the present embodiment.
[0137] First, the sampling data generation unit 111 acquires sampling policy data 151 including a sampling interval and a sampling time, and vehicle information for each sensor identifier (S121). The process of step S121 is performed by the sampling information acquisition unit 160. The sampling information acquisition unit 160 acquires, for example, the sampling policy data 151 from the sampling policy storage unit 113 and the vehicle information from the vehicle information storage unit 114.
[0138] The sampling policy data 151 includes a vehicle number (not shown) that uniquely indicates the architecture of the vehicle, and a period (sampling interval and sampling time) that indicates an acquisition period for each sensor identifier included in the vehicle sensor data. The vehicle information also includes, for example, vehicle event information.
[0139] Next, the sampling data generating unit 111 acquires vehicle sensor data for a predetermined driving time stored in the vehicle sensor data storage unit 112 (S122). The process of step S121 is performed by the vehicle sensor data acquiring unit 130. The vehicle sensor data acquiring unit 130 acquires, for example, the vehicle sensor data shown in FIG.
[0140] Next, the sampling data generating unit 111 calculates the number of pieces of received data during the sampling time for each sampling interval based on the vehicle sensor data and the sampling time included in the sampling policy data (S123). The process of step S123 is performed by the data extracting unit 161. Steps S123 and after are executed for each sampling time.
[0141] Next, the sampling data generating unit 111 judges whether the number of pieces of received data is greater than a second threshold value (S124). The process of step S124 is performed by the data extracting unit 161. If the number of pieces of received data is greater than a value obtained by adding a predetermined margin to the number of pieces of normally received data during the sampling time (an example of a second threshold value) (Yes in S124), the data extracting unit 161 extracts a predetermined number of pieces of vehicle sensor data that are continuous in time during the sampling time. If the result of step S124 is Yes, the received data (sensor value) may contain a large amount of data due to an attack. In such a case, if all the data is extracted, the number of pieces of data increases, leading to an increase in the amount of data. Therefore, if the result of step S124 is Yes, the data extracting unit 161 extracts only the predetermined number of pieces of received data (S125). This makes it possible to extract a predetermined number of pieces of received data that include data due to an attack, thereby making it possible to generate sampling data that can detect an attack while reducing the amount of data.
[0142] The predetermined number of pieces of data is, for example, a number larger than the number of pieces of normally received data. The second threshold is not limited to a value obtained by adding a predetermined margin to the number of pieces of normally received data, and may be the number of pieces of normally received data, or may be a fixed value that is set in advance for each sensor identifier. The predetermined margin may be set according to the number of pieces of normally received data, or may be a fixed value.
[0143] Furthermore, if the number of pieces of received data is equal to or less than the number of pieces of normally received data during the sampling time plus a predetermined margin (No in S124), the data extraction unit 161 extracts the number of pieces of normally received data (S126). The data extraction unit 161 extracts, for example, all of the received data (sensor values) during the sampling time.
[0144] In this way, in steps S124 to S126, the number of vehicle sensor data to be extracted within the sampling time is determined based on the number of normally received data to be observed during the sampling time, and the determined number of sensor data is extracted. For example, it may be determined whether the number of sensor values observed during the sampling time is equal to or greater than the number of normally received data, and if it is equal to or greater than the number of normally received data, a predetermined number of sensor values may be extracted, and if it is not equal to or greater than the number of normally received data, all sensor values included in the vehicle sensor data may be extracted.
[0145] Here, the number of normally received data indicates the number of normally received data within the above sampling time, and is calculated in advance based on the above cycle for each sensor identifier. For example, as shown in Figure 7, if the sampling time for the sensor identifier "vehicle speed" is "55 ms", the cycle is "approximately 25 ms", so the number of normally received data is "2".
[0146] Next, the sampling data generating unit 111 assigns to the above-mentioned extracted data a first index indicating the data order of the short-term movement of the sensor values during the sampling time, and a second index indicating the data order of the long-term movement of the sensor values at the sampling interval. That is, the sampling data generating unit 111 assigns two types of indexes to each of the sensor values included in the extracted data (S127). The process of step S127 is performed by the index assigning unit 162.
[0147] For example, when the sampling time is "55 ms", the index assignment unit 162 assigns first indexes of 1, 2, 3, etc. to the multiple data extracted during the sampling time "55 ms" in order from the top. For example, the index assignment unit 162 assigns first indexes to the multiple data extracted in order from the oldest data obtained.
[0148] The index assigning unit 162 further assigns a second index that is uniquely determined each time step S125 or S126 is executed. For example, the index assigning unit 162 assigns a second index to the extracted multiple data in the order of the oldest sampling time. Here, if the sampling interval is "1000 ms", the second index changes when processing data approximately "1000 ms" later.
[0149] Next, the sampling data generating unit 111 determines whether or not an index has been assigned to each of the extracted sensor values (S128). The process of step S128 is performed by the index assigning unit 162. The index assigning unit 162 determines, for example, whether or not the processes of steps S123 to S127 have been performed for all sampling times of each of two or more sampling intervals included in the predetermined running time.
[0150] If the index assigning unit 162 has not assigned an index to each of the extracted sensor values (No in S128), the process returns to step S123 and executes the processes from step S123 onward for the next sampling time. That is, the process of steps S123 to S127 is repeatedly executed for each sampling interval for the vehicle sensor data to be processed.
[0151] When the index assignment unit 162 assigns an index to each of the extracted sensor values (Yes in S128), the sampling data generation unit 111 transmits the extracted data as sampling data to the communication unit 126 via the communication unit 115. That is, the sampling data generation unit 111 transmits the extracted data (sampling data) to which the index has been assigned to the sampling data monitoring unit 102 (S129).
[0152] Here, the sampling data (extracted data) transmitted by the sampling data generating unit 111 will be described with reference to Fig. 10. Fig. 10 is a diagram showing an example of the sampling data according to the present embodiment.
[0153] As shown in FIG. 10, the sampling data includes a sensor identifier 172 , an evaluation ID 173 , a sensor value 174 , a first index 175 , and a second index 176 .
[0154] The sensor identifier 172 and the sensor value 174 are similar to the sensor identifier 142 and the sensor value 145 in the vehicle sensor data (see FIG. 5). The evaluation ID 173 is an identifier for uniquely identifying one piece of sensor data 171 for each sensor identifier 172 included in the sampling data.
[0155] As described above, the first index 175 indicates the order of short-term sensor value movement, and the second index 176 indicates the order of long-term sensor value movement.
[0156] For example, the sensor values indicated by evaluation IDs "1" and "2" are sensor values acquired at the same sampling time because the second indexes are both "1". In addition, the sensor value indicated by evaluation ID "1" has a first index of "1", and the sensor value indicated by evaluation ID "2" has a first index of "2", indicating that the sensor value indicated by evaluation ID "2" was acquired immediately after the sensor value indicated by evaluation ID "1".
[0157] Furthermore, when the second index is "1" or "2", the number of sensor values is two, but when the second index is "3", the number of sensor values is three or more. Since a sampling policy is determined for each sampling time, that is, for each second index, the number of sensor values can change for each sampling time.
[0158] For example, the sampling data shown in FIG.
[0159] [4-1. Configuration of the attack detection unit] Fig. 11 is a block diagram showing the configuration of the attack detection unit 120 according to this embodiment. As shown in Fig. 11, the attack detection unit 120 includes a sampling data acquisition unit 181, a base point data acquisition unit 182, a first anomaly score calculation unit 183, a second anomaly score calculation unit 184, and an attack determination unit 185.
[0160] The sampling data acquisition unit 181 is a processing unit that acquires sampling data from the sampling data storage unit 125. The sampling data acquisition unit 181 acquires, for example, the sampling data shown in FIG.
[0161] The base point data acquisition unit 182 is a processing unit that acquires base point data used for determination by the attack determination unit 185 from the base point data storage unit 124. The base point data acquisition unit 182 acquires short-term sensor flows and long-term sensor flows as base point data.
[0162] The first anomaly score calculation unit 183 is a processing unit that calculates a first anomaly score indicating the degree of deviation between the evaluation target data 191 and the short-term sensor flow, based on the evaluation target data 191 and the short-term sensor flow. That is, the first anomaly score calculation unit 183 calculates a first anomaly score for the evaluation target data 191 by using other data acquired at the sampling time when the evaluation target data 191 was acquired. The first anomaly score calculation unit 183 calculates an anomaly score for a change in the sensor value in a short period of time, such as within one sampling time. A short period of time means, for example, a time period equal to or shorter than one sampling interval (for example, one sampling time).
[0163] It can be said that the first anomaly score calculation unit 183 calculates a first anomaly score indicating the degree of anomaly of the evaluation target data, for example, based on the sensor value of the evaluation target data included in the sampling data, the first index, and a short-term sensor flow indicating time series data of sensor values prior to the evaluation target data at the sampling time when the evaluation target data was acquired.
[0164] The second anomaly score calculation unit 184 is a processing unit that calculates a second anomaly score indicating the degree of deviation between the evaluation target data 191 and the long-term sensor flow, based on the evaluation target data 191 and the long-term sensor flow. That is, the second anomaly score calculation unit 184 calculates a second anomaly score for the evaluation target data 191 using data of a plurality of sampling intervals including the sampling interval at which the evaluation target data 191 was acquired. The second anomaly score calculation unit 184 calculates an anomaly score for a change in the sensor value over a long period of time, such as a plurality of sampling intervals. A long period of time is a period of time longer than a short period of time, and means, for example, longer than one sampling interval.
[0165] It can also be said that the second anomaly score calculation unit 184 calculates a second anomaly score indicating the degree of anomaly of the evaluation target data, for example, based on the sensor value of the evaluation target data, the second index, and a long-term sensor flow indicating the change in the sensor value acquired at a sampling interval earlier than the sampling interval at which the evaluation target data was acquired.
[0166] The attack determination unit 185 is a processing unit that determines whether or not the evaluation target data 191 is the result of an attack, that is, whether or not the vehicle is under a cyber-attack, based on the calculated first abnormality score and second abnormality score, and outputs a determination result. In this embodiment, the attack determination unit 185 calculates a third abnormality score based on the first abnormality score and the second abnormality score, compares the calculated third abnormality score with a predetermined threshold (fifth threshold), and determines whether or not the evaluation target data 191 is the result of an attack based on the comparison result. Note that the attack determination unit 185 is not limited to calculating the third abnormality score, and may compare each of the first abnormality score and the second abnormality score with the corresponding threshold, and determine whether or not the evaluation target data 191 is the result of an attack based on the comparison result.
[0167] [4-2. Operation of the attack detection unit] Fig. 12 is a flowchart showing the attack detection process (S103) according to this embodiment. The attack detection process shown in Fig. 12 may be performed, for example, on any sensor value included in the sampling data, or may be performed on all of the sensor values included in the sampling data. When the attack detection process is performed on all of the sensor values, the process shown in Fig. 12 may be performed in order from the sensor value obtained the oldest, for example.
[0168] First, the attack detection unit 120 acquires sampling data, and extracts evaluation target data (see evaluation target data 191 shown in FIG. 13), which is one data item in the sampling data (S131). The processing of step S131 is performed by the sampling data acquisition unit 181.
[0169] Next, the attack detection unit 120 acquires base point data corresponding to the evaluation target data 191 (S132). The process of step S132 is performed by the base point data acquisition unit 182.
[0170] FIG. 13 is a diagram showing an example of base point data according to the present embodiment.
[0171] As shown in FIG. 13, the baseline data includes a sensor identifier 192 , an assessment ID 193 , a sensor value 194 , a first index 195 , a second index 196 , a short-term sensor flow 197 , and a long-term sensor flow 198 .
[0172] The sensor identifier 192, evaluation ID 193, sensor value 194, first index 195, and second index 196 are similar to the sensor identifier 172, evaluation ID 173, sensor value 174, first index 175, and second index 176 in the sampling data (see FIG. 10).
[0173] The short-term sensor flow 197 has the same second index as the evaluation target data 191, and is a list of past sensor values in which a unique sensor value is determined using an index earlier than the first index of the evaluation target data 191 as a reference key.
[0174] 13, for evaluation target data with evaluation ID "82", the second index has "3", and the first index has "1" which is before "2" as a reference key, and the representative value of the sensor values of evaluation target data with evaluation ID 81 is stored as the short-term sensor flow. The calculation method of the representative value will be described later.
[0175] On the other hand, the long-term sensor flow 198 is a list of representative values of past sensor values having unique values, using an index before the second index of the evaluation target data 191 as a reference key.
[0176] For example, in Fig. 13, in the case of evaluation target data with evaluation ID "82", the representative value of the sensor value of the evaluation target data is stored as a long-term sensor flow using the second indexes "1" and "2" that precede "3" as reference keys. The method of calculating the representative value will be described later.
[0177] FIG. 14 is a diagram illustrating an example of the detection result information according to the present embodiment.
[0178] 14, the detection result information includes a sensor identifier 202, an evaluation ID 203, a sensor value 204, a first anomaly score 205, a second anomaly score 206, a third anomaly score 207, and a detection result 208. The detection result 208 is an example of a determination result corresponding to the evaluation target data.
[0179] The sensor identifier 202, the evaluation ID 203, and the sensor value 204 are similar to the sensor identifier 172, the evaluation ID 173, and the sensor value 174 (see FIG. 10) in the sampling data.
[0180] The first anomaly score 205, the second anomaly score 206, the third anomaly score 207, and the detection result 208 will be described later.
[0181] 12 again, next, the attack detection unit 120 determines whether at least one of the first index and the second index is equal to or greater than a predetermined value (S133). The predetermined value is set in advance and stored in the storage unit. The predetermined value of the first index and the predetermined value of the second index may be the same value or may be different values.
[0182] If at least one of the first index and the second index for the evaluation target data is equal to or greater than a predetermined value (Yes in S133), the attack detection unit 120 calculates an anomaly score (at least one of the first anomaly score and the second anomaly score) described below, and if each of the first index and the second index is less than the predetermined value (No in S133), the detection result is set to "Unknown" (S134). If the result in step S133 is Yes, the attack detection unit 120 determines whether the first index is equal to or greater than a predetermined value (S135). The predetermined value is set in advance and stored in the storage unit. The predetermined value used in step S135 is an example of a third threshold value. The third threshold value is set, for example, to a value that allows the first anomaly score to be accurately calculated.
[0183] For example, in Fig. 13, when data with evaluation ID "1" is the evaluation target data, the first index is "1" and the second index is "1". Therefore, for example, when the predetermined value is 2, the detection result of the corresponding data is "unknown" as shown in Fig. 14. Note that when it is determined to be "unknown", the first anomaly score, the second anomaly score, and the third anomaly score may be "0".
[0184] Next, when at least one of the first index and the second index for the evaluation target data is equal to or greater than a predetermined value (Yes in S133) and the first index is equal to or greater than a predetermined value (Yes in S135), the attack detection unit 120 reads out a first detector that calculates a first anomaly score indicating the degree of deviation between the evaluation target data and a normal flow of short-term sensor values (short-term sensor flow), and calculates the first anomaly score based on the evaluation target data 191 and the short-term sensor flow (S136). The process of step S136 is performed by the first anomaly score calculation unit 183.
[0185] If the results of steps S133 and S135 are Yes, the first anomaly score calculation unit 183 reads out the first detector from the detector storage unit 122. The first detector may be realized by a trained model trained to receive the evaluation target data 191 and the short-term sensor flow as inputs and output a first anomaly score indicating the degree of deviation between the evaluation target data 191 and the short-term sensor flow. The first detector is trained to output a first anomaly score with a larger value as the degree of deviation increases, but is not limited thereto. The first anomaly score is, for example, a value from 0 to 100 (for example, a normalized value), but may also be a level (such as "high", "medium", or "low").
[0186] 13, in the case of evaluation target data with an evaluation ID of "82", the first index is "2", the short-term sensor flow includes "35.11", and the sensor value is "100.50". In this case, as shown in FIG. 14, the first anomaly score output by the first detector is, for example, "95". The first anomaly score calculation unit 183 calculates that the first anomaly score is 95.
[0187] The first detector may be an outlier detector using a machine learning model such as OC-SVM (One Class-Support Vector Machine) based on the difference between the sensor value of the evaluation target data and the short-term sensor flow. The first detector may be a detector that uses the acquisition time instead of the sensor value and determines whether the difference in acquisition time between the short-term sensor flow and the evaluation target data is based on a period. The above-mentioned detectors may be used in combination.
[0188] Note that the first anomaly score calculation unit 183 is not limited to calculating the first anomaly score using a machine learning model. For example, the first anomaly score calculation unit 183 may calculate the first anomaly score using a table (lookup table) in which the evaluation target data 191 and the short-term sensor flow are associated with the first anomaly score, or may calculate the first anomaly score by performing a predetermined calculation using the evaluation target data 191 and the short-term sensor flow.
[0189] Furthermore, if at least one of the first index and the second index for the evaluation target data is equal to or greater than a predetermined value (Yes in S133) and the first index is less than the predetermined value (No in S135), the attack detection unit 120 proceeds to step S137. If step S135 is No, the number of vehicle sensor data acquired prior to the evaluation target data within the sampling time in which the evaluation target data was acquired is small, making it difficult to accurately determine whether the evaluation target data is abnormal or not. Therefore, if step S135 is No, the attack detection unit 120 may not calculate the first anomaly score. Note that if step S135 is No, the attack detection unit 120 may set the first anomaly score to "0".
[0190] Next, the attack detection unit 120 determines whether the second index is equal to or greater than a predetermined value (S137). The predetermined value is set in advance and stored in the storage unit. The predetermined value used in step S137 may be the same as the predetermined value used in step S133, or may be different. The predetermined value used in step S137 is an example of a fourth threshold. The fourth threshold is set, for example, to a value that allows the second anomaly score to be accurately calculated.
[0191] Next, when the second index is equal to or greater than a predetermined value (Yes in S137), the attack detection unit 120 reads out a second detector that calculates a second anomaly score that indicates the degree of deviation between the evaluation target data and a normal flow of long-term sensor values (long-term sensor flow), and calculates the second anomaly score based on the evaluation target data 191 and the long-term sensor flow (S138). The process of step S138 is performed by the second anomaly score calculation unit 184.
[0192] If the result of step S137 is Yes, the second anomaly score calculation unit 184 reads out the second detector from the detector storage unit 122. The second detector may be realized by a trained model trained to receive the evaluation target data 191 and the long-term sensor flow as inputs and output a second anomaly score indicating the degree of deviation between the evaluation target data 191 and the long-term sensor flow. The second detector is trained to output a second anomaly score with a larger value as the degree of deviation increases, but is not limited to this. In addition, the second anomaly score is, for example, a value from 0 to 100 (for example, a normalized value), but may also be a level (such as "high", "medium", or "low").
[0193] 13, when data with evaluation ID "82" is the evaluation target data, the second index is "3", the long-term sensor flow includes "22.50" and "33.50", and the sensor value is "100.50". In this case, as shown in FIG. 14, the second anomaly score output by the second detector is "80". The second anomaly score calculation unit 184 calculates that the second anomaly score is 80.
[0194] The second detector may be a change point detector that uses a recurrent neural network model (RNN) or an auto regressive integrated moving average model (ARIMA) to predict normal behavior of the sensor value and determine an abnormality based on the error with the actual measured value.
[0195] Furthermore, if the second index for the evaluation target data is less than a predetermined value (No in S137), the attack detection unit 120 proceeds to step S139. If the result in step S137 is No, there is little vehicle sensor data acquired prior to the evaluation target data during the sampling interval in which the evaluation target data was acquired, making it difficult to accurately determine whether the evaluation target data is abnormal or not. Therefore, if the result in step S137 is No, the attack detection unit 120 may not need to calculate the second anomaly score. Note that, if the result in step S137 is No, the attack detection unit 120 may set the second anomaly score to "0".
[0196] The first detector and the second detector may be different detectors from each other, or may be a single detector.
[0197] Next, the attack detection unit 120 calculates a third anomaly score from the first anomaly score and the second anomaly score (S139). The process of step S139 is performed by the attack determination unit 185.
[0198] The third abnormality score is an abnormality score based on the first abnormality score and the second abnormality score. The third abnormality score may be, for example, the first abnormality score, the second abnormality score, or a statistical value based on the first abnormality score and the second abnormality score. The statistical value is, for example, an average value (for example, a weighted average), but may also be any of a median, a mode, a maximum value, and a minimum value.
[0199] It should be noted that it is not essential to calculate the third anomaly score.
[0200] Next, the attack detection unit 120 uses the third anomaly score to determine whether or not an attack has occurred (S140). The process of step S139 is performed by the attack determination unit 185. It can also be said that the attack determination unit 185 determines whether or not the evaluation target data is data resulting from an attack. For example, when the third anomaly score is equal to or greater than a fifth threshold, the attack determination unit 185 determines that there is an "attack" (there is an attack), and when the third anomaly score is less than the fifth threshold, the attack determination unit 185 determines that there is a "normal" (there is no attack).
[0201] 14, when the evaluation ID is "82", both the first anomaly score of "95" and the second anomaly score of "80" of the corresponding data are calculated (when S135 is Yes and S137 is Yes), and the average value of "87" is calculated as the third anomaly score. In this case, when the fifth threshold is "60", the attack determination unit 185 determines that the sensor value of the evaluation ID "82" is an "attack". In other words, the attack determination unit 185 determines that the sensor value "100.50" is due to an attack.
[0202] 14, when the evaluation ID is "81", only the second anomaly score of "6" is calculated for the corresponding data (when S135 is No and S137 is Yes), and the third anomaly score is calculated as "6". In this case, when the fifth threshold is "60", the attack determination unit 185 determines that the sensor value "35.11" is not due to an attack.
[0203] When step S135 is No and step S137 is Yes, it is assumed that the situation is, for example, immediately after the start time of the sampling interval when the second index is equal to or greater than a predetermined value. In this case, all of the vehicle sensor data within the sampling time of the sampling interval has not been acquired (for example, the state at the beginning of the sampling time), but there is enough vehicle sensor data for each sampling interval to determine whether or not there is an attack. In other words, when step S135 is No and step S137 is Yes, it is a case where it is impossible to detect an attack using the first detector, and it is possible to detect an attack using the second detector.
[0204] At this time, the attack detection unit 120 can determine whether or not the evaluation target data is the result of an attack using the second anomaly score from the second detector.
[0205] Furthermore, when step S135 is Yes and step S137 is No, it is assumed that, for example, the number of sampling intervals that have elapsed is small, but all vehicle sensor data within the sampling time of the elapsed sampling interval has been acquired. In this case, the amount of vehicle sensor data for each sampling interval is small, but there is enough vehicle sensor data within the sampling time to determine whether or not an attack has occurred. In other words, when step S135 is Yes and step S137 is No, it is a case where an attack can be detected using the first detector, but an attack cannot be detected using the second detector.
[0206] At this time, the attack detection unit 120 can determine whether or not the evaluation target data is the result of an attack using the first anomaly score from the first detector.
[0207] In this way, the first detector and the second detector can complement each other's attack detection during the time when each detector is unable to detect an attack. In other words, the attack detection unit 120 can reduce the time during which it is unable to detect an attack. This improves the attack detection performance of the attack detection unit 120.
[0208] Next, the attack detection unit 120 stores the determination result of step S134 or S140 in the detection result storage unit 123 (S141).
[0209] In this way, when the first index corresponding to the evaluation target data is less than the third threshold and the second index corresponding to the evaluation target data is less than the fourth threshold, the attack detection unit 120 determines that it is unclear whether the evaluation target data is the result of an attack (S134). Furthermore, when the third abnormality score based on the first abnormality score and the second abnormality score is equal to or greater than the fifth threshold, the attack detection unit 120 determines that the evaluation target data is the result of an attack, and when the third abnormality score is less than the fifth threshold, the attack detection unit 120 determines that the evaluation target data is normal (S140), and outputs a detection result indicating the result of the determination.
[0210] In FIG. 12, the attack detection unit 120 may further use vehicle event information to calculate the anomaly score. For example, when the vehicle event information indicates that at least one of the following has occurred: switching the vehicle engine ON and OFF, switching the shift lever, and switching the automatic driving mode ON and OFF, the attack detection unit 120 may initialize (reset) the sensor value included in the long-term sensor flow. In other words, when at least one of the events has occurred, the attack detection unit 120 may calculate a second anomaly score based on the sensor value acquired after that time. In this case, since it is necessary to accumulate the sensor value included in the long-term sensor flow, the second anomaly score cannot be calculated during the accumulation period (for example, a period corresponding to two or more sampling intervals), but the first anomaly score can be calculated when the first index is equal to or greater than a predetermined value.
[0211] In this way, when at least one of the above occurs, the attack detection unit 120 detects the attack based on the first anomaly score, thereby reducing the time during which the attack cannot be detected. Therefore, the attack detection performance of the attack detection unit 120 can be further improved.
[0212] [5-1. Configuration of base point data update section] 15 is a block diagram showing the configuration of the base point data update unit 121 according to the present embodiment. As shown in FIG. 15, the base point data update unit 121 includes a base point data acquisition unit 182, a detection result acquisition unit 201, and a base point data addition unit 302.
[0213] The base point data acquisition unit 182 is a processing unit that acquires the base point data from the base point data storage unit 124 .
[0214] The detection result acquisition unit 201 is a processing unit that acquires the detection result of the attack detection unit 120 from the detection result storage unit 123 .
[0215] The base point data adding unit 302 is a processing unit that adds data corresponding to the detection result acquired by the detection result acquiring unit 201 to the base point data acquired by the base point data acquiring unit 182. It can also be said that the base point data adding unit 302 updates the base point data.
[0216] [5-2. Operation of base point data update section] 16 is a flowchart showing the base point data update process (S104) according to this embodiment. The base point data update process is performed by the base point data update unit 121.
[0217] First, the base data update unit 121 initializes the base data according to the evaluation target data acquired by the base data acquisition unit 182 (S151). The base data update unit 121 acquires base data corresponding to the evaluation target data 191, and initializes the short-term sensor flow in the base data when the first index of the evaluation target data 191 is 1, and initializes the long-term sensor flow in the base data when the second index of the evaluation target data 191 is 1. Note that initialization here means making the short-term sensor flow or the long-term sensor flow into a state in which no sensor value is included. For example, the short-term sensor flow 197 and the long-term sensor flow when the evaluation ID in FIG. 13 is "1", and the short-term sensor flow 197 when the evaluation ID is "81" ([ ] shown in FIG. 13) indicate that they have been initialized.
[0218] Next, the base data update unit 121 acquires a detection result corresponding to the evaluation target data (S152). The process of step S152 is performed by the detection result acquisition unit 201. The detection result acquisition unit 201 reads out the detection result corresponding to the evaluation target data 191 from the detection result storage unit 123. The detection result acquisition unit 201 reads out, for example, the detection result shown in FIG.
[0219] Next, the base point data update unit 121 judges whether the acquired detection result is “attack” (S153). The processes from step S153 to S157 are performed by the base point data addition unit 302. The base point data addition unit 302 performs the judgment of step S153 based on, for example, the detection result 208 corresponding to the evaluation ID 203 of the evaluation target data.
[0220] Next, if the detection result is "attack" (Yes in S153), the base data addition unit 302 sets the estimated sensor value as the additional data as a value to complement the sensor value (S154), and if the detection result is "normal" or "unknown" (No in S153), it sets the sensor value (actual measured value) of the data to be evaluated as the additional data (S155).
[0221] The estimated sensor value may be a value added to the end of the short-term sensor flow corresponding to the evaluation target data, or an average value of the short-term sensor flows.The estimated sensor value may also be a predicted value of the sensor value of the evaluation target data calculated by the second detector based on the long-term sensor flow.
[0222] For example, in FIG. 13, when the evaluation ID is “82”, the detection result is “attack”, so the base data addition unit 302 sets the estimated sensor value to “35.11” as the value added to the end of the short-term sensor flow corresponding to the evaluation target data.
[0223] Next, the base point data adding unit 302 updates the short-term sensor flow (Equation 1) used to determine the next evaluation target data according to the following (Equation 2) (S156).
[0224]
number
number
[0225] Here, si indicates one sensor value included in the short-term sensor flow, i indicates a first index of the evaluation target data, and j indicates a second index of the evaluation target data. If the above detection result labeli-j is "Attack", the base point data adding unit 302 updates si with the above estimated sensor value pi-j. Also, if the labeli-j is "Normal" or "Unknown", the base point data adding unit 302 updates si with the above sensor value xi-j of the evaluation target data. Note that si is an example of a representative value.
[0226] Next, the base point data adding unit 302 updates the long-term sensor flow (Equation 3) according to the following (Equation 4) (S157).
[0227]
number
number
[0228] Here, hj indicates one sensor value included in the long-term sensor flow. If the detection result labeli-j is "Attack", the base point data adding unit 302 updates hj with the estimated sensor value pi-j. If labeli-j is "Normal" or "Unknown", the base point data adding unit 302 updates hj with the sensor value xi-j of the evaluation target data. Note that hj is an example of a representative value.
[0229] In this way, the base point data adding unit 302 updates two types of base point data each time it judges whether or not there is an attack, and uses the updated data for the next judgment.
[0230] In the above description, an example has been shown in which the sampling process and the attack detection process are performed by different devices in the monitoring system 100, but they may be performed by the same device (for example, one device).
[0231] As described above, sampling data is generated by adjusting the sampling interval and sampling time according to the characteristics of the vehicle sensor data that is the target of the monitoring process (for example, attack detection process). Furthermore, attacks are judged by calculating two types of anomaly scores based on the sampling data that includes two types of indexes that indicate the short-term and long-term movements of the sensor values. This makes it possible to accurately detect attacks while keeping monitoring costs down.
[0232] Furthermore, when the detection result corresponding to the evaluation target data is an attack, the base data update unit 121 estimates a sensor value that the evaluation target data should take using at least one of the short-term sensor flow and the long-term sensor flow, and updates the short-term sensor flow and the long-term sensor flow based on the estimated sensor value, the first index, and the second index. Furthermore, when the detection result corresponding to the evaluation target data is normal or unknown, the base data update unit 121 updates the short-term sensor flow and the long-term sensor flow based on the sensor value (actual measurement value) of the evaluation target data, the first index, and the second index.
[0233] This makes it possible to prevent outliers such as attacks from being mixed into the base point list. Therefore, by using such a base point list, attacks can be detected with high accuracy while keeping monitoring costs down.
[0234] The following describes how the above-mentioned monitoring system 100 is used.
[0235] Fig. 17 is a schematic diagram showing the configuration of a monitoring system 100 according to this embodiment. Fig. 17 is a block diagram showing the configuration of the monitoring system 100 including a sampling unit 401 and a sampling data monitoring unit 501. The monitoring system 100 includes a server 500 existing on the cloud, and an in-vehicle device 400. The monitoring system 100 can acquire extracted vehicle sensor data (sampling data) via the in-vehicle device 400, and detect an attack based on the acquired vehicle sensor data.
[0236] The in-vehicle device 400 and the server 500 are communicably connected via a public communication network such as the Internet. The in-vehicle device 400 is a security ECU (Electronic Control Unit) or a car navigation system.
[0237] The in-vehicle device 400 includes a sampling unit 401 and an attack notification device 402 .
[0238] The sampling unit 401 is a device that transmits the results of sampling the vehicle network data (sampling data) to the sampling data monitoring unit 501. For example, the sampling unit 401 converts binary-expressed CAN data into sensor values and further converts the values into sampling data every predetermined driving time. Then, the sampling unit 401 transmits the sampling data to the sampling data monitoring unit 501.
[0239] The attack notification device 402 is a device for notifying the driver of the occurrence of an attack on the vehicle network received from the server 500. The attack notification device 402 outputs, for example, a voice or a screen notifying the driver of the occurrence of the attack and instructing the driver to drive in a safety mode disconnected from the network. The attack notification device 402 is realized by a display device, a speaker device, or the like.
[0240] The sampling data monitoring unit 501 determines whether or not there is an attack on the vehicle network data (eg, sampling data) received from the in-vehicle device 400, and transmits the result of the attack detection to the attack notification device 402 if there is an attack.
[0241] The configuration of the sampling unit 401 may be the same as that of the sampling unit 101. The configuration of the sampling data monitoring unit 501 may be the same as that of the sampling data monitoring unit .
[0242] 18 is a schematic diagram showing the configuration of an in-vehicle device 600 according to this embodiment. Fig. 18 is a block diagram showing the configuration of an in-vehicle device 600 including a monitoring device 601. The in-vehicle device 600 includes the monitoring device 601 and an attack notification device 602.
[0243] The monitoring device 601 determines whether or not there is an attack based on the results of sampling vehicle network data only inside the vehicle, and if there is an attack, transmits the attack detection result to the attack notification device 602. The monitoring device 601 is configured to have, for example, the functions of the sampling unit 101 and the sampling data monitoring unit 102. In this way, the functions of the sampling unit 101 and the sampling data monitoring unit 102 may be realized by one device (for example, a device mounted on the vehicle). The monitoring device 601 is an example of an attack detection system.
[0244] The attack notification device 602 is a device for notifying the driver of the occurrence of an attack on the vehicle network received from the monitoring device 601. The attack notification device 602 is realized by a display device, a speaker device, or the like.
[0245] In this way, the monitoring system 100 in Fig. 17 and the monitoring device 601 in Fig. 18 can detect attacks at low monitoring costs by appropriately determining whether an attack has occurred based on the results of sampling the vehicle network. Furthermore, the monitoring system 100 and the monitoring device 601 can also notify the driver that an attack has been detected by notifying the driver of the detection result, thereby making it possible to reduce the occurrence of accidents caused by attacks.
[0246] (Other embodiments) Although the monitoring device according to the embodiment has been described above, the present disclosure is not limited to this embodiment.
[0247] For example, in the above embodiment, the present invention has been described as a security measure in an in-vehicle network mounted on a vehicle such as an automobile, but the scope of application is not limited thereto. For example, the present invention may be applied to mobility such as construction machinery, agricultural machinery, ships, trains, and airplanes as moving bodies, in addition to vehicles such as automobiles. In other words, the present invention may be applied as a cybersecurity measure in a mobility network and a mobility network system.
[0248] Furthermore, each processing unit included in the monitoring device according to the above embodiment is typically realized as an LSI, which is an integrated circuit. These may be implemented individually as single chips, or may be implemented as a single chip that includes some or all of the processing units.
[0249] The integrated circuit is not limited to an LSI, and may be realized by a dedicated circuit or a general-purpose processor. An FPGA that can be programmed after the LSI is manufactured, or a reconfigurable processor that can reconfigure the connections and settings of the circuit cells inside the LSI may also be used.
[0250] The present disclosure may also be realized as a monitoring method (e.g., an attack detection method) executed by a monitoring system. For example, one aspect of the present disclosure may be a computer program that causes a computer to execute each of the characteristic steps included in the monitoring method. Another aspect of the present disclosure may be a computer-readable non-transitory recording medium, such as a DVD, on which such a program is recorded. For example, such a program may be recorded on a recording medium and distributed or circulated. For example, such a program and a digital signal consisting of the program may be transmitted via a telecommunication line, a wireless or wired communication line, a network such as the Internet, data broadcasting, or the like. For example, the distributed program may be installed in a device having another processor and the program may be executed by the processor, thereby making it possible to cause the device to perform each of the above-mentioned processes.
[0251] In the above embodiment, the first order information is a first index, but the present invention is not limited to this, and the first order information may be, for example, an absolute time. In the above embodiment, the second order information is a second index, but the present invention is not limited to this, and the second order information may be, for example, an absolute time when a sampling interval starts.
[0252] In the above embodiment, each component may be realized by a dedicated hardware or by executing a software program suitable for each component. Each component may be realized by a program execution unit such as a CPU or a processor reading and executing a software program recorded on a recording medium such as a hard disk or a semiconductor memory.
[0253] In addition, the division of functional blocks in the block diagram is an example, and multiple functional blocks may be realized as one functional block, one functional block may be divided into multiple blocks, or some functions may be transferred to another functional block. Furthermore, the functions of multiple functional blocks having similar functions may be processed in parallel or in a time-sharing manner by a single piece of hardware or software.
[0254] Furthermore, when the monitoring system is realized by a plurality of devices, the functions of the sampling unit and the functions of the sampling data monitoring unit may be distributed in any manner among the plurality of devices.
[0255] In addition, the order in which each step is performed in the flowchart is merely an example for specifically explaining the present disclosure, and an order other than the above may be used. In addition, some of the steps may be performed simultaneously (in parallel) with other steps.
[0256] Although the monitoring system according to one or more aspects has been described based on the embodiment, the present disclosure is not limited to this embodiment. As long as it does not deviate from the spirit of the present disclosure, various modifications conceived by a person skilled in the art to this embodiment and forms constructed by combining components in different embodiments may also be included within the scope of one or more aspects. [Industrial Applicability]
[0257] The present disclosure can be applied to a monitoring device for vehicle sensor data, and can appropriately determine whether vehicle control data included in the vehicle sensor data is an attack. For example, the present disclosure can be used in a monitoring device or a monitoring system, such as an in-vehicle device such as a security ECU, or a security information and event management (SIEM) in a remote monitoring server. [Explanation of symbols]
[0258] 100 Surveillance System 101, 401 Sampling section 102, 501 Sampling data monitoring unit 110 Sampling Policy Generation Unit 111 Sampling data generation unit 112 Vehicle sensor data storage unit 113 Sampling Policy Memory Unit 114 Vehicle information storage unit 115, 126 Communications Department 120 Attack Detection Unit 121 Base point data update section 122 Detector memory unit 123 Detection result storage unit 124 Base point data storage unit 125 Sampling data storage unit 130 Vehicle sensor data acquisition unit 131 Sensor data stability calculation unit 132 Sampling policy decision unit 141 Vehicle Sensor Data 142, 152, 172, 192, 202 Sensor Identifiers 143 Data ID 144 Acquisition time 145, 174, 194, 204 Sensor Values 151 Sampling Policy Data 153 File number 154 Data Distribution 155 Vehicle Event Flag 156 Sampling Interval 157 Sampling Time 160 Sampling information acquisition unit 161 Data Extraction Unit 162 Indexing Department 173, 193, 203 Rating ID 175, 195 First Index 176, 196 Second Index 181 Sampling data acquisition unit 182 Base point data acquisition section 183 First Anomaly Score Calculation Unit 184 Second Anomaly Score Calculation Unit 185 Attack Judgment Section 191 Data to be evaluated 197 Short-Term Sensor Flow 198 Long Term Sensor Flow 201 Detection result acquisition unit 205 First Anomaly Score 206 Second Anomaly Score 207 Third Anomaly Score 208 Detection Results 302 Base point data addition section 400, 600 In-vehicle equipment 402, 602 Attack notification device 500 servers 601 Monitoring equipment
Claims
1. An attack detection method for detecting an attack on sensor data transmitted and received within a mobility for controlling the mobility, comprising: a sampling policy determination step of determining a sampling policy including a sampling interval and a sampling time for each sampling interval, for extracting the sensor value for detecting the attack from the sensor data based on at least one of a statistical value indicating a variation in the sensor value of the sensor data and the mobility event information indicating a timing of a change in the sensor value for each identifier of the sensor value included in the sensor data; a sampling data generating step of generating sampling data including two or more sensor values extracted from the sensor data based on the sampling interval and the sampling time, first order information indicating a temporal order of the two or more sensor values during the sampling time, and second order information indicating a temporal order of the two or more sensor values for each sampling interval; an attack determination step of calculating a first anomaly score indicating a degree of anomaly in the evaluation target data based on a sensor value of the evaluation target data included in the sampling data, the first order information, and a short-term sensor flow indicating time-series data of sensor values older than the evaluation target data at the sampling time when the evaluation target data was acquired; calculating a second anomaly score indicating a degree of anomaly in the evaluation target data based on the sensor value of the evaluation target data, the second order information, and a long-term sensor flow indicating a change in sensor value acquired at a sampling interval older than the sampling interval when the evaluation target data was acquired; and determining whether or not the evaluation target data is due to the attack based on the calculated first anomaly score and second anomaly score, and outputting a determination result. Attack detection methods.
2. the statistical value includes a variance value indicating a variance of the sensor value for each of the identifiers; the event information includes flag information indicating whether a change in the sensor value has occurred; In the sampling policy determination step, the sampling interval and the sampling time are determined based on the variance value and the flag information. The method of detecting an attack according to claim 1 .
3. In the sampling policy determination step, the sampling time and the sampling interval are determined according to a combination of a determination result of whether or not the variance value is equal to or greater than a first threshold value and whether or not the flag information indicates that no change has occurred in the sensor value. The attack detection method according to claim 2 .
4. The flag information includes information indicating whether or not at least one of the following has occurred: switching of a drive unit of the mobility between ON and OFF, switching of a shift lever, and switching of an autonomous driving mode between ON and OFF. The attack detection method according to claim 2 or 3.
5. In the sampling data generating step, a number of pieces of sensor data to be extracted within the sampling time is determined based on a number of pieces of normal reception data to be observed during the sampling time, and the determined number of pieces of sensor data is extracted. The attack detection method according to any one of claims 1 to 4.
6. In the sampling data generating step, it is determined whether or not the number of the sensor values observed during the sampling time is equal to or greater than the number of normally received data, and if the number is equal to or greater than the number of normally received data, a predetermined number of the sensor values are extracted, and if the number is not equal to or greater than the number of normally received data, all of the sensor values included in the sensor data are extracted. The attack detection method according to claim 5 .
7. the first order information is a first index indicating an order in which the two or more sensor values were acquired during the sampling time; the second order information is a second index indicating an order of the sampling intervals in which the two or more sensor values were acquired; In the attack determination step, if the first index corresponding to the evaluation target data is less than a third threshold and the second index corresponding to the evaluation target data is less than a fourth threshold, it is determined that it is unclear whether the evaluation target data is the result of the attack; if a third abnormality score based on the first abnormality score and the second abnormality score is equal to or greater than a fifth threshold, it is determined that the evaluation target data is the result of the attack; if the third abnormality score is less than the fifth threshold, it is determined that the evaluation target data is normal; and the determination result corresponding to the evaluation target data indicating the determination result is output. The attack detection method according to any one of claims 1 to 6.
8. Further, the method includes an update step of updating the short-term sensor flow and the long-term sensor flow based on the determination result corresponding to the evaluation target data, the first order information, and the second order information. The attack detection method according to any one of claims 1 to 7.
9. In the updating step, when the judgment result corresponding to the evaluation target data is an attack, a sensor value that the evaluation target data should take is estimated using at least one of the short-term sensor flow and the long-term sensor flow, and the short-term sensor flow and the long-term sensor flow are updated based on the estimated sensor value, the first order information, and the second order information; when the judgment result corresponding to the evaluation target data is normal or unknown, the short-term sensor flow and the long-term sensor flow are updated based on the sensor value of the evaluation target data, the first order information, and the second order information. The attack detection method according to claim 8.
10. The sensor data is stored in a device mounted on the mobility, and is a result of decoding network data in byte representation transmitted and received at any running time interval according to at least one protocol of CAN (Controller Area Network: registered trademark), CAN FD (CAN with Flexible Data-rate), FlexRay (registered trademark), and EtherNet (registered trademark). The attack detection method according to any one of claims 1 to 9.
11. The mobility is a vehicle. The attack detection method according to any one of claims 1 to 10.
12. An attack detection system for detecting an attack on sensor data transmitted and received within a mobility for controlling the mobility, comprising: a sampling policy determination unit that determines a sampling policy including a sampling interval and a sampling time for each sampling interval, for extracting the sensor value for detecting the attack from the sensor data based on at least one of a statistical value indicating a variation in the sensor value of the sensor data and the mobility event information indicating a timing of a change in the sensor value for each identifier of the sensor value included in the sensor data; and a sampling data generating unit that generates sampling data including two or more sensor values extracted from the sensor data, first order information indicating a time order of the two or more sensor values during the sampling time, and second order information indicating a time order of the two or more sensor values for each sampling interval, based on the sampling interval and the sampling time; an attack determination unit that calculates a first anomaly score indicating a degree of anomaly in the evaluation target data based on a sensor value of the evaluation target data included in the sampling data, the first order information, and a short-term sensor flow indicating time-series data of sensor values older than the evaluation target data at the sampling time when the evaluation target data is acquired; calculates a second anomaly score indicating a degree of anomaly in the evaluation target data based on the sensor value of the evaluation target data, the second order information, and a long-term sensor flow indicating a change in sensor value acquired at a sampling interval older than the sampling interval when the evaluation target data is acquired; and determines whether the evaluation target data is the result of the attack based on the calculated first anomaly score and second anomaly score, and outputs a determination result. Attack detection system.
13. A program for causing a computer to execute the attack detection method according to any one of claims 1 to 11.
Citation Information
Patent Citations
Monitoring system, processing device, monitoring device, monitoring method and program
JP2018164159A
Communication device, communication method, and program
JP2019146145A
Methods and apparatus to detect anomalies of a monitored system
US20190138423A1
Information processing device and information processing method
WO2018105319A1
Illicit act detection method, illicit act detection device, and program
WO2019187350A1
