A security analysis method for intelligent network based on sensing and computing linkage

Abstract

The present invention discloses a method for security analysis of intelligent Internet of Things based on sensor-computing linkage. The method analyzes the overall vulnerability of the intelligent Internet of Things system by constructing and injecting adversarial samples for artificial intelligence algorithms from the physical domain in response to the phenomenon of linkage between sensors and artificial intelligence algorithms that is widely present in the intelligent Internet of Things. The method first summarizes the intelligent Internet of Things system into several information pathways based on the front and back links of the intelligent Internet of Things in processing information, and then performs vulnerability analysis on each information pathway, builds an actual test platform for signal injection, determines whether there is an actual threat and has professionals evaluate the degree of harm, and finally comprehensively analyzes the analysis results of each vulnerability to evaluate the overall security of the intelligent Internet of Things system. The method of the present invention can comprehensively analyze potential vulnerabilities and dig out security issues of the system more quickly.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent system security and relates to an intelligent network security analysis method based on sensing and computing linkage. Background Art

[0002] The IoT is a combination of AI technology and IoT infrastructure to achieve more efficient IoT operations, improve human-machine interaction, and enhance data management and analysis. In 2018, KPMG released a forward-looking study on the future scenario applications of AI by 2040. Analysts described in detail a scenario in which each device in the IoT is equipped with its own AI algorithm, and these AIs can autonomously connect to other AIs to perform tasks intelligently together, using group intelligence to control and execute value creation in real time. The basic requirement for realizing this scenario is the perception and intelligent algorithm of each IoT device.

[0003] Along with the benefits brought by new technologies come potential safety risks. Due to the application of artificial intelligence technology, the intelligent Internet of Things system will take on more complex tasks for humans, such as autonomous driving applications, but this also means that the user's life safety is controlled by machines. Therefore, the security of the intelligent Internet of Things system often leads to larger and more serious problems. Existing work related to artificial intelligence security mainly studies the algorithm vulnerability problem caused by the unexplainability of the artificial intelligence algorithm itself, and has never considered the security risks introduced by the vulnerability of the perception module to the artificial intelligence algorithm. There are problems such as incomplete vulnerability analysis and inefficient vulnerability mining. Summary of the invention

[0004] In view of the defects of the above-mentioned prior art, the present invention provides a security analysis method of the intelligent Internet based on sensing and computing linkage. The security analysis method based on sensing and computing linkage of the present invention can comprehensively analyze potential vulnerabilities and more quickly dig out the security problems of the system. The present invention focuses on the core scientific issues of the security of the intelligent Internet system, conducts research on the theoretical model of general security threats, and forms a systematic and effective theoretical and methodological system, which has important theoretical guiding significance for the future development of the intelligent Internet.

[0005] The present invention is implemented by the following technical solutions:

[0006] A method for security analysis of intelligent Internet of Things based on sensor-computing linkage. The method first summarizes the intelligent Internet of Things system into several information pathways according to the front and back links of information processing of the intelligent Internet of Things, then performs vulnerability analysis on each information pathway, builds an actual test platform for signal injection, determines whether there is an actual threat and has professionals evaluate the degree of harm, and finally evaluates the overall security of the intelligent Internet of Things system by combining the analysis results of each vulnerability.

[0007] In the above technical solution, further, the method specifically comprises the following steps:

[0008] 1) Obtain the hardware information such as the type, model, and function of the sensors equipped in the intelligent Internet system, obtain the configuration information such as the data type, format, and range transmitted by the sensors, obtain the software information such as the process, parameters, and structure of the artificial intelligence algorithm that receives the transmitted data, and obtain the decision information such as the behavior, timeliness, and results of the intelligent Internet system determined by the artificial intelligence algorithm. The information to be obtained in this step specifically includes the following four types:

[0009] (1) The sensor hardware information that needs to be obtained includes the working principle of the device, the basic structure of the internal circuit of the device, and the perceptible physical signal information, such as type, strength, range, etc.;

[0010] (2) The data configuration information to be acquired, including the data storage type, data precision, data expression meaning, and the data range used by subsequent algorithms;

[0011] (3) The algorithm software information that needs to be obtained, including the preprocessing of sensor data, the type of artificial intelligence algorithm, the structure and parameters of the neural network model, and the meaning of the algorithm processing results;

[0012] (4) The system decision information that needs to be obtained includes the type and parameters of the system decision model, the control method and effect of the decision on the actuator, and the impact of the decision on the system, including duration, intensity, and scope.

[0013] 2) Based on the hardware, configuration, software, and decision information obtained in step 1), several "state→perception→algorithm→decision→execution" information pathways of the intelligent network system are formed, and each information pathway is independently analyzed, including perception vulnerability analysis, algorithm vulnerability analysis, and decision vulnerability analysis. This step specifically includes:

[0014] The information pathway is defined as a series of processes that a physical state undergoes in each link of the intelligent Internet of Things system after being converted into information, specifically including: ① "state → perception", specifically: a physical attribute that reflects the environment or its own state is captured by the relevant sensors of the intelligent Internet of Things system and converted into sensor readings representing the perception information, ② "perception → algorithm", specifically: the sensor readings representing the perception information are converted into the data format specified by the artificial intelligence algorithm after preprocessing, ③ "algorithm → decision", specifically: the artificial intelligence algorithm recognizes the perception data and passes the recognition results into the decision model, ④ "decision → execution", specifically: finally, the decision model outputs decision instructions to control the actuator to complete the corresponding behavior based on the recognition results of the artificial intelligence algorithm;

[0015] The perceptual vulnerability analysis specifically refers to the use of common physical signals, such as sound, light, electromagnetics, and heat, to test the sensor's response behavior to signals, explore signal injection methods other than those required for the design function, and test the sensor's response behavior to common playback and interference attacks. For example, for a microphone sensor, the signal required for its design function is a voice signal, but the use of the laser resonance effect may also cause the microphone sensor to capture voice information that does not exist;

[0016] The algorithm vulnerability analysis specifically refers to constructing adversarial samples using white-box or black-box methods according to the type of artificial intelligence algorithm, and testing the robustness of the algorithm when it is subjected to adversarial samples with different interference levels;

[0017] The decision vulnerability analysis specifically refers to testing whether the decision model will produce decision-level errors after inputting erroneous artificial intelligence algorithm recognition results, and testing the self-correction ability of the decision model.

[0018] 3) Based on the vulnerability analysis results of step 2), design an implementation plan for injecting signals from the sensor according to the physical mechanism and working principle of the sensor, and construct adversarial samples for subsequent algorithms under the physical constraints of the injected signal, and build an actual test platform. This step specifically includes the following four steps:

[0019] (1) Based on the analysis of sensor vulnerability in step 2), design a corresponding implementation plan for physical signal injection for the sensor;

[0020] (2) for the physical signal injection scheme described in step (1), construct a mapping relationship from the physical signal to the sensor reading and determine the actual constraints of the physical signal;

[0021] (3) Based on the actual constraints of the physical signal constructed in step (2), construct a constrained optimization problem to solve the adversarial example;

[0022] (4) For the adversarial sample solved in step (3), the specific parameters of the physical signal actually injected are obtained according to the mapping relationship from physical signal to sensor reading constructed in step (2), the physical signal is injected using the actual test platform, and the feasibility of the actual test is verified according to the response of the sensor reading.

[0023] 4) For the actual test device built in step 3), determine whether the vulnerability of the information path will pose a security threat to the intelligent network system, and evaluate the threat level of the vulnerability based on the working environment information of the intelligent network system. This step specifically includes the following two steps:

[0024] (1) The vulnerability analyst shall make a judgment on whether the vulnerability will have a substantial impact on the IoT system based on the actual test results;

[0025] (2) The product developers of the IoT system rate the severity of the vulnerability.

[0026] 5) Based on the analysis results of steps 2), 3), and 4) for each information channel, combined with the frequency and probability of each security threat under actual working conditions, the overall security level of the intelligent network system is evaluated. This step specifically includes the following two steps:

[0027] (1) Vulnerability analysts assign weight scores to each vulnerability based on the access frequency of each information path under working conditions, the conditions for signal injection, and the difficulty of subsequent protection;

[0028] (2) The vulnerability weight w of step (1) i And the vulnerability severity rating x in step 4) i , calculate the weighted sum of the two and give the overall safety assessment score s:

[0029]

[0030] The beneficial effects of the present invention are:

[0031] (1) The present invention innovatively proposes a security analysis method for an intelligent Internet of Things system based on sensor-computing linkage. By decomposing a complex system into a number of information pathways of "state → perception → algorithm → decision → execution", the difficulty of security analysis is simplified and the general working mode of the intelligent Internet of Things is grasped. The method is general and generally applicable to currently popular intelligent Internet of Things systems, such as self-driving cars, smart voice assistants, etc.

[0032] (2) The present invention constructs an injection signal under physical constraints based on the vulnerability analysis results of the algorithm part. On the one hand, the added physical constraints are in line with the actual scenario and avoid the problem of exaggerating security threats. On the other hand, the injection signal construction guided by the algorithm vulnerability is more efficient than brute force attempts, avoiding the neglect and misjudgment of potential security threats and preventing them before they occur. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 is a flow chart of the method of the present invention. DETAILED DESCRIPTION

[0034] In order to facilitate ordinary technicians in this field to understand and implement this invention, the present invention is further elaborated in detail below with reference to the drawings and embodiments. It should be understood that the implementation examples described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0035] The method flow of the embodiment of the present invention is as follows: Figure 1 shown.

[0036] The intelligent network security analysis method based on sensing and computing linkage of the present invention takes an autonomous driving car as an example. The autonomous driving car has many information pathways, such as "object→camera image→obstacle detection→driving decision model→turn / no turn", "pedestrian→lidar captured point cloud→point cloud classification→driving decision model→brake / accelerate", etc. The specific analysis method includes the following steps:

[0037] 1) Obtaining hardware information such as the type, model, and function of the sensors equipped in the autonomous vehicle, obtaining configuration information such as the type, format, and range of the data transmitted by the sensors, obtaining software information such as the process, parameters, and structure of the artificial intelligence algorithm that receives the transmitted data, and obtaining decision information such as the behavior, timeliness, and results of the intelligent Internet system determined by the artificial intelligence algorithm:

[0038] (1) Sensor data configuration information, the data stored is an RGB image, the data range is an integer value from 0 to 255, and the value represents the size of the component of each channel;

[0039] (2) The algorithm for processing the RGB image data is a convolutional neural network, which uses the Faster RCNN structure. The function of the algorithm is to detect obstacles based on image information;

[0040] (3) The decision model determines whether to stop or change lanes based on the output of the obstacle detection algorithm within the last 10 frames.

[0041] 2) Perform independent analysis on each information path, including perception vulnerability analysis, algorithm vulnerability analysis, and decision vulnerability analysis. Here, we take "object → camera image → obstacle detection → driving decision model → turn / no turn" as an example to perform independent analysis on the information path:

[0042] (1) Perception vulnerability analysis. For example, according to tests, the camera of the system is not equipped with an infrared filter and will receive injection of infrared light signals that are invisible to the human eye.

[0043] (2) Algorithm vulnerability analysis. For example, after testing, the convolutional neural network used in the obstacle detection module of the system is vulnerable to adversarial samples;

[0044] (3) Decision vulnerability analysis. For example, after testing, the decision model of the system does not use sensor fusion and only relies on the output results of the obstacle detection module.

[0045] 3) Based on the vulnerability analysis results of step 2), design an implementation plan for injecting signals from the sensor according to the physical mechanism and working principle of the sensor, and construct adversarial samples for subsequent algorithms under the physical constraints of the injected signal, so as to build an actual test platform:

[0046] (1) Based on the analysis of the perception vulnerability in step 2), design an implementation scheme using an infrared laser transmitter as a signal injection;

[0047] (2) For the physical signal injection scheme described in step (1), a mapping relationship from the physical signal to the sensor reading is constructed, and the actual constraints of the physical signal are determined. In this embodiment, due to the injected infrared light, only the reading of the red channel in the image can be changed;

[0048] (3) Based on the actual constraints of the physical signal constructed in step (2), construct a constrained optimization problem to solve the adversarial example;

[0049] (4) For the adversarial sample solved in step (3), the specific parameters of the physical signal actually injected are obtained according to the mapping relationship from the physical signal to the sensor reading constructed in step (2). In this embodiment, the specific parameters include the incident angle and emission light intensity of the infrared laser transmitter. The physical signal is injected using an infrared laser transmitter, and the feasibility of the actual test is verified based on the response of the sensor reading.

[0050] 4) For the actual test platform built in step 3), analyze whether the vulnerability of the information path will pose a security threat to the actual system, and evaluate the threat level of the vulnerability based on the working environment information of the intelligent network system:

[0051] (1) According to actual test results, it is found that the injection of the infrared light signal does cause the autonomous driving system to misjudge the obstacle ahead;

[0052] (2) The product developers of the IoT system rate the severity of the vulnerability.

[0053] 5) Based on the analysis results of steps 2), 3), and 4) for each information path of the autonomous vehicle, combined with the frequency and probability of each security threat under actual working conditions, the overall security level of the intelligent network system is evaluated. This step specifically includes the following two steps:

[0054] (1) Vulnerability analysts assign weight scores to each vulnerability based on the access frequency of each information path under working conditions, the conditions for signal injection, and the difficulty of subsequent protection;

[0055] (2) The vulnerability weight w of step (1) iAnd the vulnerability severity rating x in step 4) i , calculate the weighted sum of the two and give the overall safety assessment score s.

[0056]

[0057] It should be understood that parts not elaborated in detail in this specification belong to the prior art.

[0058] It should be understood that the above description of the embodiments is relatively detailed and cannot be regarded as limiting the scope of patent protection of the present invention. Under the enlightenment of the present invention, ordinary persons in the field may make substitutions or modifications without departing from the scope of protection of the claims of the present invention, which shall fall within the scope of protection of the present invention. The scope of protection requested by the present invention shall be based on the attached claims.

Claims

1. A security analysis method for intelligent network based on sensor-computing linkage, characterized in that: The method first summarizes the intelligent Internet of Things system into several information pathways according to the front and back links of the intelligent Internet of Things information processing, and then conducts vulnerability analysis on each information pathway, builds an actual test platform for signal injection, determines whether there is an actual threat and sends it to professionals to evaluate its degree of harm, and finally comprehensively analyzes the analysis results of each vulnerability to evaluate the overall security of the intelligent Internet of Things system; The method specifically comprises the following steps: 1) Obtaining sensor hardware information equipped with the intelligent Internet of Things system, obtaining data configuration information transmitted by the sensor, obtaining software information of the artificial intelligence algorithm that receives the transmitted data, and obtaining decision information of the intelligent Internet of Things system determined by the artificial intelligence algorithm; 2) Based on the hardware, configuration, software, and decision-making information obtained in step 1), several information pathways of "state→perception→algorithm→decision→execution" of the intelligent network system are formed, and each information pathway is independently analyzed, including perception vulnerability analysis, algorithm vulnerability analysis, and decision vulnerability analysis; 3) Based on the vulnerability analysis results of step 2), design an implementation plan for injecting signals from the sensor according to the physical mechanism and working principle of the sensor, and construct adversarial samples for subsequent algorithms under the physical constraints of the injected signal, so as to build a practical test platform; 4) For the actual test platform built in step 3), determine whether the vulnerability of the information path will pose a security threat to the intelligent Internet of Things system, and evaluate the threat level of the vulnerability based on the working environment information of the intelligent Internet of Things system; 5) Comprehensively analyzing the analysis results of each information channel in steps 2), 3) and 4), combined with the frequency and probability information of each security threat under actual working conditions, conduct an overall security assessment of the intelligent network system.

2. The intelligent network security analysis method based on sensing and computing linkage according to claim 1 is characterized in that: The step 1) comprises: 1.1) The sensor hardware information required includes the device working principle, the basic structure of the device's internal circuit, and the perceptible physical signal information; 1.2) The data configuration information required includes the data storage type, data precision, data expression meaning, and the data range used by subsequent algorithms; 1.3) The algorithm software information required to be obtained, including the pre-processing algorithm for sensor data, the type of artificial intelligence algorithm, the structure and parameters of the neural network model, and the meaning of the algorithm processing results; 1.4) The system decision information that needs to be obtained includes the type and parameters of the system decision model, the control method and effect of the decision on the actuator, and the impact of the decision on the system.

3. The intelligent network security analysis method based on sensing and computing linkage according to claim 1 is characterized in that: In the step 2): The information pathway is defined as a series of processes that a physical state undergoes in each link of the intelligent Internet of Things system after being converted into information, specifically including: ① "state→perception", specifically: a physical attribute that reflects the environment or its own state is captured by the relevant sensors of the intelligent Internet of Things system and converted into sensor readings representing the perception information, ② "perception→algorithm", specifically: the sensor readings representing the perception information are converted into the data format specified by the artificial intelligence algorithm after preprocessing, ③ "algorithm→decision", specifically: the artificial intelligence algorithm recognizes the perception data and passes the recognition results into the decision model, ④ "decision→execution", specifically: finally, the decision model outputs decision instructions to control the actuator to complete the corresponding behavior based on the recognition results of the artificial intelligence algorithm; The sensor vulnerability analysis specifically includes using common physical signals to test the sensor's response behavior to signals, exploring signal injection methods other than the signal types required for the design function, and testing the sensor's response behavior to common replay and interference attacks; The algorithm vulnerability analysis is specifically to construct adversarial samples using white box or black box methods according to the type of artificial intelligence algorithm, and test the robustness of the algorithm when it is subjected to adversarial samples with different interference levels; The decision vulnerability analysis is specifically to test whether the decision model will generate erroneous decision instructions after receiving erroneous artificial intelligence algorithm recognition results, and to test the self-correction ability of the decision model.

4. The intelligent network security analysis method based on sensing and computing linkage according to claim 1 is characterized in that: The step 3) specifically comprises the following steps: 3.1) Based on the analysis of the sensor vulnerability in step 2), design a corresponding implementation plan for physical signal injection of the sensor; 3.2) For the physical signal injection scheme described in step 3.1), construct a mapping relationship from the physical signal to the sensor reading and determine the actual constraints of the physical signal; 3.3) Based on the actual constraints of the physical signal constructed in step 3.2), construct a constrained optimization problem to solve the adversarial example; 3.4) For the adversarial sample solved in step 3.3), obtain the specific parameters of the physical signal actually injected according to the mapping relationship from physical signal to sensor reading constructed in step 3.2), use the actual test platform to inject the physical signal, and verify the feasibility of the actual test based on the response of the sensor reading.

5. The intelligent network security analysis method based on sensing and computing linkage according to claim 1 is characterized in that: The step 4) comprises the following steps: 4.1) The vulnerability analyst shall make a judgment on whether the vulnerability will have a substantial impact on the IoT system based on the actual test results; 4.2) The product developers of the IoT system rate the severity of the vulnerability.

6. The intelligent network security analysis method based on sensing and computing linkage according to claim 5 is characterized in that: The step 5) comprises the following steps: 5.1) Vulnerability analysts weight each vulnerability based on the access frequency of each information channel under working conditions, the conditions of signal injection, and the difficulty of subsequent protection, thereby obtaining the vulnerability weight w i ; 5.2) Comprehensive vulnerability weight w of step 5.1) i And the vulnerability hazard rating x obtained in step 4) i , calculate the weighted sum of the two and give the overall safety assessment score s:

Images