Security assessment method and system for Internet of Things malicious detector based on evolutionary computation

Abstract

The present invention belongs to the field of security detection technology and provides a security assessment method and system for an Internet of Things malicious detector based on evolutionary computing. The present invention performs evolutionary computing on acquired malicious samples under the guidance of clustering and distance to obtain adversarial samples. During parameter evolution, if no adversarial sample corresponding to the malicious sample is found, new parameters are selected from a sequence of the number of clusters and a sequence of termination generations based on the output reason when the failure occurs and the evolutionary calculation is performed again. An adversarial attack is designed specifically for a malicious detector based on streaming text, and adversarial samples are generated using the adversarial attack method. The detector is attacked and the security of the detector is evaluated by the attack success rate, thereby solving the problem that existing assessment methods are not applicable to Internet of Things malicious detectors based on streaming text.

Description

Technical Field

[0001] The present invention belongs to the field of security detection technology, and in particular relates to a security assessment method and system for an Internet of Things malicious detector based on evolutionary computing. Background Art

[0002] In today's rapidly developing society, the emergence of the Internet of Things (IoT) has greatly facilitated human life. The success of the IoT has also led to a gradual increase in malicious attacks targeting it. This has prompted the continuous development of IoT malware detection technology, particularly machine learning-based IoT malware detectors, which have a high detection success rate. The success of machine learning-based IoT malware detectors has enabled some malicious actors to bypass and potentially compromise IoT malware detectors through adversarial attacks. To address this issue, security assessments of IoT malware detectors are crucial.

[0003] The inventors discovered that among various IoT malware detector security assessment methods, some use evolutionary computing to find solutions. Evolutionary computing mimics the biological evolutionary process to create adversarial samples and evaluate detectors. However, classic evolutionary computing can miss solutions, resulting in a low attack success rate and an inability to maximize the attack on the detector to assess its security. Specifically, existing security assessment methods in other fields are not suitable for direct application to IoT malware detectors based on streaming text. This is because these detectors have different input features, so security assessment methods tailored for detectors in other fields are not precisely applicable to these IoT malware detectors. Some methods insert perturbations and deformations into the original samples, but the insertion method affects the original malicious functionality, making them inappropriate to completely copy existing methods. Classic evolutionary computing methods on the market use a single set of parameters, which can easily miss adversarial samples and result in a low success rate. Setting a large number of evolutionary generations and evolutionary directions results in a high success rate, but the computational complexity increases. Setting a small number of evolutionary generations and evolutionary directions results in a high speed but also misses adversarial samples. Summary of the Invention

[0004] In order to solve the above problems, the present invention proposes a security assessment method and system for an Internet of Things malicious detector based on evolutionary computing. The present invention specifically designs adversarial attacks for malicious detectors based on streaming text, uses adversarial attack methods to generate adversarial samples, and attacks the detector and evaluates the security of the detector based on the attack success rate; generates adversarial samples by appending character-level perturbations rather than modifying characters, avoiding destroying the original malicious functions when generating adversarial samples; uses evolutionary computing guided by clustering and distance to search for adversarial samples; and combines several sets of evolutionary computing methods with different parameters to balance the probability of missing adversarial samples and computational overhead.

[0005] In order to achieve the above object, the present invention is implemented through the following technical solutions:

[0006] In a first aspect, the present invention provides a security assessment method for an IoT malicious detector based on evolutionary computing, comprising:

[0007] Obtain malicious samples;

[0008] Guided by clustering and distance, evolutionary calculations are performed on acquired malicious samples to generate adversarial samples. A sequence of cluster numbers and termination generations is predefined. During parameter evolution, if no adversarial sample corresponding to the malicious sample is found, new parameters are selected from the sequence of cluster numbers and termination generations based on the output reason for failure and evolutionary calculations are repeated.

[0009] Use all adversarial samples to attack IoT malicious detectors and obtain the attack success rate;

[0010] The obtained attack success rate is used to evaluate the security of IoT malicious detectors.

[0011] Furthermore, normal samples are obtained, clustered, and cluster centers are selected as landmarks; malicious samples evolve toward the landmarks.

[0012] Furthermore, each time a sample is generated, it is used to attack a real malicious detector. If the sample fails the detection of the detector, the Euclidean distance from the sample to the landmark point is calculated. If the Euclidean distance from a sample to the landmark point is greater than the Euclidean distance from its parent sample to the landmark point, the sample is retained and participates in the next round of evolutionary calculation.

[0013] Furthermore, in each generation of evolutionary computation, perturbations are added to the malicious samples.

[0014] Furthermore, perturbation is to append words to the URL, add characters representing URL parameters, and separate the appended perturbation from the original URL parameters.

[0015] Furthermore, a candidate adversarial sample is found in a certain generation and the evolutionary computation is terminated.

[0016] Furthermore, in a certain generation, all samples are eliminated and the evolutionary computation is terminated.

[0017] Furthermore, if the maximum generation limit is exceeded, the evolutionary computation is terminated.

[0018] Furthermore, before the attack begins, a list is defined that includes a sequence of cluster numbers and a sequence of termination generations; the evolutionary computation initially selects the first element of the list as a parameter;

[0019] If the evolution fails because the generation limit is exceeded, the next element in the termination generation sequence is selected as the parameter to restart the evolutionary calculation; if the evolution fails because the number of populations in a certain generation becomes 0, the next element in the cluster number sequence is selected as the parameter to restart the evolutionary calculation.

[0020] In a second aspect, the present invention further provides an IoT malicious detector security assessment system based on evolutionary computing, comprising:

[0021] The data collection module is configured to: obtain malicious samples;

[0022] The evolution module is configured to perform evolutionary calculations on acquired malicious samples guided by clustering and distance to generate adversarial samples. A sequence of cluster numbers and termination generations is predefined. During parameter evolution, if no adversarial sample corresponding to the malicious sample is found, new parameters are selected from the sequence of cluster numbers and termination generations based on the reason for failure and evolutionary calculation is repeated.

[0023] The attack module is configured to: use all adversarial samples to attack the IoT malicious detector and obtain the attack success rate;

[0024] The evaluation module is configured to evaluate the security of the IoT malicious detector using the obtained attack success rate.

[0025] In a third aspect, the present invention further provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the security assessment method of the Internet of Things malicious detector based on evolutionary computing described in the first aspect.

[0026] In a fourth aspect, the present invention also provides an electronic device comprising a memory, a processor, and a computer program stored in the memory and runnable on the processor. When the processor executes the program, the steps of the security assessment method of the Internet of Things malicious detector based on evolutionary computing described in the first aspect are implemented.

[0027] Compared with the prior art, the present invention has the following beneficial effects:

[0028] 1. The present invention performs evolutionary calculations on acquired malicious samples under the guidance of clustering and distance to obtain adversarial samples. During parameter evolution, if no adversarial sample corresponding to the malicious sample is found, new parameters are selected from the sequence of cluster numbers and the sequence of termination generations based on the output reason for failure and evolutionary calculation is repeated. Adversarial attacks are designed specifically for malicious detectors based on streaming text, and adversarial samples are generated using adversarial attack methods. The detector is attacked and its security is evaluated based on the attack success rate, solving the problem that existing evaluation methods are not applicable to malicious detectors based on streaming text in the Internet of Things.

[0029] 2. The present invention generates adversarial samples by adding character-level perturbations instead of modifying characters, thus avoiding destroying the original malicious functions when generating adversarial samples. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] The drawings constituting a part of the specification of this embodiment are used to provide a further understanding of this embodiment. The schematic embodiments and descriptions of this embodiment are used to explain this embodiment and do not constitute an improper limitation on this embodiment.

[0031] Figure 1 This is an overall flow chart of Example 1 of the present invention;

[0032] Figure 2 The malicious sample of Example 1 of the present invention is deformed toward the target;

[0033] Figure 3 Evolutionary calculation of malicious samples in Example 1 of the present invention generates candidate adversarial samples. DETAILED DESCRIPTION

[0034] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0035] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which the present application belongs.

[0036] Adversarial attack: An attack method targeting machine learning models. Attackers bypass the detection of machine learning models by adding perturbations to malicious samples.

[0037] Evolutionary computation: a method of finding solutions to problems by simulating the evolution of organisms.

[0038] Example 1:

[0039] This embodiment provides a security assessment method for an IoT malicious detector based on evolutionary computing, including:

[0040] Obtain malicious samples;

[0041] Guided by clustering and distance, evolutionary calculations are performed on acquired malicious samples to generate adversarial samples. A sequence of cluster numbers and termination generations is predefined. During parameter evolution, if no adversarial sample corresponding to the malicious sample is found, new parameters are selected from the sequence of cluster numbers and termination generations based on the output reason for failure and evolutionary calculations are repeated.

[0042] Use all adversarial samples to attack IoT malicious detectors and obtain the attack success rate;

[0043] The obtained attack success rate is used to evaluate the security of IoT malicious detectors.

[0044] Detectors in different fields have different input features, which means that security assessment methods customized for other detectors are not exactly suitable for IoT malicious detectors based on streaming text. Some methods insert perturbations and deformations into the original samples, but the insertion method affects the original malicious functions, so the methods available on the market cannot be completely copied. To address the problem that security assessment methods in other fields on the market are not suitable for direct application to IoT malicious detectors based on streaming text, in this embodiment, the acquired malicious samples are subjected to evolutionary calculations under the guidance of kmeans clustering and Euclidean distance to generate candidate adversarial samples. The sequence of the number of clusters L = [l1, l2, ..., ln] and the sequence of the termination generation Y = [y1, y2, y3, .., yn] are pre-defined. In the evolutionary calculation of setting l and y, if no adversarial sample corresponding to the malicious sample is found, new l and y are selected from L and Y as new parameters based on the reason for failure and the evolutionary calculation is restarted. Each sample generated by evolutionary computation attacks the target detector until it succeeds or a termination condition is met. The security of the target detector can be assessed by the attack success rate. For example, if the attack success rate is less than a preset value, the target detector is deemed secure; otherwise, it is deemed unsafe. In this embodiment, the clustering method can be kmeans clustering; in other embodiments, the clustering method can be any clustering method, such as Brich clustering. In this embodiment, the distance can be Euclidean distance; in other embodiments, the distance can be any distance measurement method, such as Mahalanobis distance and cosine distance.

[0045] To address this issue, the classic evolutionary computation method uses a single set of parameters, which can easily lead to the problem of missing adversarial examples. Furthermore, if a large number of generations and directions are set, the success rate is high, but the computational complexity increases. If a small number of generations and directions are set, the speed is high, but adversarial examples may also be missed.

[0046] like Figure 2 As shown, in this embodiment, first, normal samples are clustered and landmark points are selected. Specifically, in this embodiment, a labeled feature data set is mastered, normal samples in the data set are taken out, and these samples are clustered into N clusters through kmeans, and the i-th cluster is represented by Ci. For each cluster, the cluster center point is selected as the landmark point Fi. The landmark point can be any sample point, such as the median point or the n-th point in the cluster. For the convenience of understanding, the cluster center point is used as the landmark point in this article. The malicious sample will evolve towards the landmark point Fi.

[0047] Secondly, if Figure 3As shown, evolutionary computation searches for candidate samples. In each generation, perturbations are added to malicious samples. Perturbations essentially consist of appending words to the Uniform Resource Locator (URL). These words are selected from a bag-of-words, which is composed of words segmented from the HTTP header in the dataset using ngram segmentation. These words cannot be appended directly to the URL because they must not interfere with the original malicious functionality of the HTTP header. Therefore, characters representing URL parameters are added to separate the appended perturbations from the original URL parameters. Each time a sample is generated, it is used to attack a real malware detector. If the sample fails the detector, the Euclidean distance from the sample to the landmark Fi is calculated. If the Euclidean distance from a sample to Fi is greater than the Euclidean distance from its parent sample to Fi, the sample is retained and participates in the next round of evolutionary computation. By appending character-level perturbations rather than modifying characters, adversarial samples are generated. This avoids destroying the original malicious functionality when generating adversarial samples.

[0048] Again, set the termination conditions for evolutionary computation: if a candidate adversarial sample is found in a certain generation, the evolutionary computation is terminated; or, if all samples are eliminated in a certain generation, the evolutionary computation is terminated; or, if the maximum generation limit is exceeded, the evolutionary computation is terminated.

[0049] Finally, the attack results are recorded and the target detector score is evaluated. After the user uses the malicious samples they know to attack the target detector, the percentage of successfully transformed malicious samples out of the total number of malicious samples (sucrate) is recorded. The target detector score is calculated as Scrore = (1-sucrate) * 100.

[0050] Several evolutionary computation methods with different parameters are combined to balance the probability of missing adversarial samples and computational overhead. Specifically, most malicious samples can be found at lower generations and fewer evolutionary directions, so there is no need to set large parameters for them to find adversarial samples. A few malicious samples cannot be found at lower generations, so when the search fails at low generations and in evolutionary directions n, the parameters are adjusted based on the reason for the evolutionary failure.

[0051] Before the attack begins, the user will customize a list, such as L = [5, 10, 15], Y = [1, 3, 5]; evolutionary computing will initially select the first element of the list as a parameter.

[0052] If the evolution fails because the generation limit is exceeded, the next element in the list Y is selected as the parameter to restart the evolution calculation.

[0053] If the evolution fails because the population size becomes 0 in a certain generation, the next element in the list L is selected as the parameter to restart the evolution calculation.

[0054] Example 2:

[0055] This embodiment provides an IoT malicious detector security assessment system based on evolutionary computing, including:

[0056] The data collection module is configured to: obtain malicious samples;

[0057] The evolution module is configured to perform evolutionary calculations on acquired malicious samples guided by clustering and distance to generate adversarial samples. A sequence of cluster numbers and termination generations is predefined. During parameter evolution, if no adversarial sample corresponding to the malicious sample is found, new parameters are selected from the sequence of cluster numbers and termination generations based on the reason for failure and evolutionary calculation is repeated.

[0058] The attack module is configured to: use all adversarial samples to attack the IoT malicious detector and obtain the attack success rate;

[0059] The evaluation module is configured to evaluate the security of the IoT malicious detector using the obtained attack success rate.

[0060] The working method of the system is the same as the security assessment method of the Internet of Things malicious detector based on evolutionary computing in Example 1, and will not be repeated here.

[0061] Example 3:

[0062] This embodiment provides a computer-readable storage medium having a computer program stored thereon. When the program is executed by a processor, the steps of the method for security assessment of an IoT malicious detector based on evolutionary computing described in Example 1 are implemented.

[0063] Example 4:

[0064] This embodiment provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the steps of the method for security assessment of an IoT malicious detector based on evolutionary computing described in Example 1 are implemented.

[0065] The above description is merely a preferred embodiment of this embodiment and is not intended to limit this embodiment. Those skilled in the art will readily appreciate that this embodiment may be modified and varied in various ways. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of this embodiment shall be within the scope of protection of this embodiment.

Claims

1. A security assessment method for Internet of Things malicious detectors based on evolutionary computing, characterized in that: include: Obtain malicious samples; Under the guidance of clustering and distance, evolutionary calculation is performed on the acquired malicious samples to obtain adversarial samples; The sequence of the number of clusters and the sequence of the termination generations are pre-defined. During parameter evolution, if no adversarial sample corresponding to the malicious sample is found, new parameters are selected from the sequence of the number of clusters and the sequence of the termination generations based on the output reason when the failure occurs, and the evolutionary calculation is repeated; In each generation of evolutionary computation, perturbations are added to the malicious samples; Perturbation is to append words to the URL, add characters representing URL parameters, and separate the appended perturbation from the original URL parameters; Use all adversarial samples to attack IoT malicious detectors and obtain the attack success rate; The obtained attack success rate is used to evaluate the security of IoT malicious detectors.

2. The security assessment method of an Internet of Things malicious detector based on evolutionary computing according to claim 1 is characterized in that: Obtain normal samples, cluster them, and select the cluster center as the landmark point; The malicious sample evolves towards the landmark point.

3. The security assessment method of an Internet of Things malicious detector based on evolutionary computing according to claim 2 is characterized in that: Each time a sample is generated, it is used to attack a real malicious detector. If the sample fails the detector's detection, the Euclidean distance from the sample to the landmark point is calculated. If the Euclidean distance from a sample to the landmark point is less than the Euclidean distance from its parent sample to the landmark point, the sample is retained and participates in the next round of evolutionary calculation.

4. The security assessment method of an Internet of Things malicious detector based on evolutionary computing according to claim 1, characterized in that: Find a candidate adversarial sample in a certain generation and terminate the evolutionary computation; Or, in a certain generation, all samples are eliminated, terminating the evolutionary computation; Or, if the maximum number of generations is exceeded, the evolutionary computation is terminated.

5. The security assessment method of an Internet of Things malicious detector based on evolutionary computing according to claim 1, characterized in that: Before the attack begins, a list is defined that includes the number of clusters and the termination generation sequence. Evolutionary computation initially selects the first element of the list as a parameter. If the evolution fails because the generation limit is exceeded, the next element in the termination generation sequence is selected as the parameter to restart the evolution calculation; If the evolution fails because the number of populations in a certain generation becomes 0, the next element in the cluster number sequence is selected as a parameter to restart the evolution calculation.

6. An IoT malicious detector security assessment system based on evolutionary computing, characterized in that: include: The data collection module is configured to: obtain malicious samples; The evolution module is configured to perform evolutionary calculations on acquired malicious samples guided by clustering and distance to generate adversarial samples. A sequence of cluster numbers and termination generations is predefined. During parameter evolution, if no adversarial sample corresponding to the malicious sample is found, new parameters are selected from the sequence of cluster numbers and termination generations based on the reason for failure and evolutionary calculation is repeated. In each generation of evolutionary computation, perturbations are added to the malicious samples; Perturbation is to append words to the URL, add characters representing URL parameters, and separate the appended perturbation from the original URL parameters; The attack module is configured to: use all adversarial samples to attack the IoT malicious detector and obtain the attack success rate; The evaluation module is configured to evaluate the security of the IoT malicious detector using the obtained attack success rate.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the security assessment method of the Internet of Things malicious detector based on evolutionary computing as described in any one of claims 1 to 5 are implemented.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the steps of the security assessment method of the Internet of Things malicious detector based on evolutionary computing as described in any one of claims 1 to 5 are implemented.

Images