Performance evaluation method and device of infrared target detection model, equipment and medium

Abstract

The invention discloses a performance evaluation method and device for an infrared target detection model, equipment and a medium, and relates to the technical field of computers, and the method comprises the steps: carrying out the simulation modeling of geometric figures in a digital environment, and determining an initial digital adversarial sample based on a corresponding geometric modeling result and an infrared detection image sample; performing iterative optimization on the initial digital countermeasure sample based on a preset sample optimization strategy to determine an optimized digital countermeasure sample; performing multi-scale domain migration transformation enhancement on the optimized digital confrontation sample based on a preset multi-scale transformation framework to determine a target physical confrontation sample; and based on the target physical confrontation sample, performing an attack test of an infrared black box query attack on an infrared target detection model, and determining a performance evaluation result corresponding to the infrared target detection model by using a corresponding test result. According to the method, the multi-scale attack test on the infrared target detection model can be efficiently and effectively realized, and the test concealment and robustness are improved.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and in particular to a method, apparatus, equipment and medium for evaluating the performance of infrared target detection models. Background Technology

[0002] Object detectors (object detection models) have become the cornerstone of many practical applications such as autonomous driving, video surveillance, and facial recognition. Therefore, evaluating the safety and other performance characteristics of object detectors is of paramount importance.

[0003] Currently, because infrared sensors are only sensitive to the temperature of the target object, when conducting physical attacks on infrared target detectors to test their performance, only heat-generating / heat-insulating materials can be used as physical countermeasures. However, these materials will appear as black or white perturbation patterns under the detection of infrared sensors, representing heat sources and cold sources respectively. As a result, there are problems such as weak perturbation patterns, insufficient concealment, and a single attack scale. Moreover, the robustness of existing attack methods is weak, making them difficult to deploy in reality. Summary of the Invention

[0004] In view of this, the purpose of this invention is to provide a method, apparatus, device, and medium for performance evaluation of infrared target detection models, which can efficiently and effectively realize multi-scale attack testing of infrared target detection models and improve the concealment and robustness of the test. The specific solution is as follows:

[0005] Firstly, this application provides a performance evaluation method for an infrared target detection model, including:

[0006] In a digital environment, geometric simulation modeling is performed, and initial digital adversarial samples are determined based on the corresponding geometric modeling results and infrared detection image samples.

[0007] The initial digital adversarial sample is iteratively optimized based on a preset sample optimization strategy to determine the optimized digital adversarial sample.

[0008] The optimized digital adversarial sample is enhanced by multi-scale domain migration transformation based on a preset multi-scale transformation framework to determine the target physical adversarial sample.

[0009] The infrared target detection model was subjected to an infrared black-box query attack based on the target physical adversarial sample, and the performance evaluation result of the infrared target detection model was determined using the corresponding test results.

[0010] Optionally, the step of performing geometric simulation modeling in a digital environment and determining initial digital adversarial samples based on the corresponding geometric modeling results and infrared detection image samples includes:

[0011] In a digital environment, simulation models of lines, triangles, and ellipses are performed separately to determine the initial modeling results;

[0012] Acquire infrared detection image samples collected by the target infrared sensor; the target infrared sensor is an infrared sensor with a fixed preset area;

[0013] The linear simulation results based on the geometric modeling results and the infrared detection image samples are synthesized to determine the first synthesized sample;

[0014] The triangle simulation results based on the geometric modeling results and the infrared detection image samples are synthesized to determine the second synthesized sample;

[0015] The ellipse simulation results based on the geometric modeling results and the infrared detection image samples are synthesized to determine the third synthesized sample;

[0016] The first, second, and third synthetic samples are processed using a particle swarm optimization algorithm to determine the initial digital adversarial sample after processing.

[0017] Optionally, the method further includes:

[0018] In determining the initial digital adversarial sample, the position of the geometric figures in the initial digital adversarial sample is constrained based on a preset mask.

[0019] Optionally, the iterative optimization of the initial digital adversarial sample based on a preset sample optimization strategy to determine the optimized digital adversarial sample includes:

[0020] Based on a preset sample optimization strategy, the initial digital adversarial sample is input into the infrared target detection model, and the target object confidence output by the model is used as the adversarial loss to iteratively optimize the initial digital adversarial sample in order to determine the optimized digital adversarial sample.

[0021] Optionally, the step of inputting the initial digital adversarial sample into the infrared target detection model based on a preset sample optimization strategy, and using the target object confidence score output by the model as the adversarial loss, to iteratively optimize the initial digital adversarial sample includes:

[0022] The combination of physical parameters related to the geometric position of the initial digital adversarial sample is used as particles and randomly initialized to determine the initial particle swarm and the velocity vector set corresponding to the initial particle swarm.

[0023] Adversarial examples are generated for each particle in the initial particle swarm to determine the adversarial example generation result for the current iteration round.

[0024] The adversarial sample generation results are input into the infrared target detection model, and the current individual optimal solution and global optimal solution are determined by using the corresponding target object confidence scores output by the model.

[0025] The velocity and position of each particle in the initial particle swarm are updated based on the inertia factor, learning factor, velocity vector set corresponding to the initial particle swarm, the current individual optimal solution, and the global optimal solution to determine the current updated particle swarm.

[0026] Determine whether the current condition for terminating the iteration is met, and if so, determine the optimized digital adversarial sample based on the updated particle swarm.

[0027] Optionally, the step of performing multi-scale domain migration transformation enhancement on the optimized digital adversarial sample based on a preset multi-scale transformation framework to determine the target physical adversarial sample includes:

[0028] The transformation distribution is simulated using the expected transformation framework, and the optimized digital adversarial sample is enhanced with multi-scale domain migration transformation based on the simulation results to determine the target physical adversarial sample.

[0029] Optionally, the step of performing an infrared black-box query attack test on the infrared target detection model based on the target physical adversarial sample, and determining the performance evaluation result corresponding to the infrared target detection model using the corresponding test results, includes:

[0030] The target physical adversarial sample is input into the infrared target detection model to perform infrared black-box query attack tests on the infrared target detection model with line segment, triangle and / or elliptical geometric perturbations, and to determine the detection results output by the model.

[0031] Based on the detection results, the current test results are determined, and the test results are analyzed to determine the performance evaluation results corresponding to the infrared target detection model.

[0032] Secondly, this application provides a performance evaluation device for an infrared target detection model, comprising:

[0033] The initial sample generation module is used to perform geometric simulation modeling in a digital environment and determine the initial digital adversarial samples based on the corresponding geometric modeling results and infrared detection image samples.

[0034] The sample optimization module is used to iteratively optimize the initial digital adversarial sample based on a preset sample optimization strategy to determine the optimized digital adversarial sample.

[0035] The multi-scale transformation module is used to perform multi-scale domain migration transformation enhancement on the optimized digital adversarial sample based on a preset multi-scale transformation framework, so as to determine the target physical adversarial sample.

[0036] The result determination module is used to perform infrared black-box query attack tests on the infrared target detection model based on the target physical adversarial sample, and to determine the performance evaluation result corresponding to the infrared target detection model using the corresponding test results.

[0037] Thirdly, this application provides an electronic device, comprising:

[0038] Memory, used to store computer programs;

[0039] A processor is used to execute the computer program to implement the steps of the aforementioned performance evaluation method for infrared target detection models.

[0040] Fourthly, this application provides a computer-readable storage medium for storing a computer program, which, when executed by a processor, implements the steps of the aforementioned performance evaluation method for infrared target detection models.

[0041] As can be seen, in this application, geometric simulation modeling is performed in a digital environment, and initial digital adversarial samples are determined based on the corresponding geometric modeling results and infrared detection image samples. The initial digital adversarial samples are iteratively optimized based on a preset sample optimization strategy to determine optimized digital adversarial samples. Multi-scale domain migration transformation enhancements are performed on the optimized digital adversarial samples based on a preset multi-scale transformation framework to determine target physical adversarial samples. Infrared black-box query attack tests are then performed on the infrared target detection model based on the target physical adversarial samples, and the performance evaluation results corresponding to the infrared target detection model are determined using the corresponding test results. In other words, in this application, geometric simulation modeling is first performed, and initial digital adversarial samples are determined using the geometric modeling results and infrared detection image samples. Then, the samples are iteratively optimized based on a preset sample optimization strategy. Multi-scale domain migration transformation enhancements are performed on the determined optimized digital adversarial samples using a preset multi-scale transformation framework. Finally, infrared black-box query attack tests are performed on the infrared target detection model based on the determined target physical adversarial samples to complete the performance test of the infrared target detection model. This allows for efficient and effective multi-scale attack testing of the infrared target detection model and improves the concealment and robustness of the test. Attached Figure Description

[0042] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0043] Figure 1 A flowchart of a performance evaluation method for an infrared target detection model provided in this application;

[0044] Figure 2 A flowchart illustrating a specific performance evaluation method for an infrared target detection model provided in this application;

[0045] Figure 3 A schematic diagram illustrating adversarial sample generation based on geometric simulation modeling provided in this application;

[0046] Figure 4 A schematic diagram of the structure of a performance evaluation device for an infrared target detection model provided in this application;

[0047] Figure 5 This application provides a structural diagram of an electronic device. Detailed Implementation

[0048] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0049] Currently, because infrared sensors are only sensitive to the temperature of the target object, physical attacks on infrared target detectors to test their performance often only allow the use of heat-generating / heat-insulating materials as physical countermeasures. However, these materials will appear as black or white perturbation patterns under the detection of infrared sensors, representing heat and cold sources respectively. This results in weak perturbation patterns, insufficient concealment, and a single attack scale. Furthermore, existing attack methods are not robust enough for practical deployment. To address this, this application provides a performance evaluation scheme for infrared target detection models that can efficiently and effectively perform multi-scale attack testing on infrared target detection models, while improving the concealment and robustness of the test.

[0050] See Figure 1 As shown, this embodiment of the invention discloses a performance evaluation method for an infrared target detection model, including:

[0051] Step S11: Perform geometric simulation modeling in a digital environment, and determine the initial digital adversarial sample based on the corresponding geometric modeling results and infrared detection image samples.

[0052] In this embodiment, combined with Figure 2 As shown, firstly, in a digital environment, geometric simulation modeling is performed and combined with clean samples to generate initial digital adversarial samples. Clean samples are unprocessed samples. Specifically, in the digital environment, simulation modeling of lines, triangles, and ellipses is performed to determine the initial modeling results; infrared detection image samples collected by the target infrared sensor are acquired; the target infrared sensor is an infrared sensor with a fixed preset area; the simulation results of lines based on the geometric modeling results and the infrared detection image samples are synthesized to determine the first synthesized sample; the simulation results of triangles based on the geometric modeling results and the infrared detection image samples are synthesized to determine the second synthesized sample; the simulation results of ellipses based on the geometric modeling results and the infrared detection image samples are synthesized to determine the third synthesized sample; the first, second, and third synthesized samples are processed using a particle swarm optimization algorithm to determine the processed initial digital adversarial samples. It can be understood that the infrared detection image samples collected by the target infrared sensor are also clean samples.

[0053] It is important to understand that, in combination Figure 3 As shown below, the simulation modeling of geometric figures is illustrated using three geometric figures as examples: lines, triangles, and ellipses.

[0054] (1) Linear modeling;

[0055] To balance stealth and robustness, a straight-line attack uses two straight-line segments; therefore, four point coordinates are used. .in, (k=1, 2, 3, 4) represents the straight line The kth vertex The x-coordinate; (k=1, 2, 3, 4) represents the straight line The kth vertex The ordinate.

[0056] (2) Triangle modeling;

[0057] The triangle is modeled using the position coordinates of its three vertices. ,in, (k=1,2,3) represents a triangle The kth vertex The x-coordinate; and (k=1,2,3) represents a triangle The kth vertex The ordinate.

[0058] (3) Ellipse modeling;

[0059] Three parameters are used to simulate and model the ellipse: the length of the ellipse's horizontal axis, etc. Length of the ellipse's vertical axis The center of the ellipse coordinates Ellipse modeling provides a smooth curved shape that produces a smoother and more continuous perturbation effect compared to straight lines and triangles.

[0060] Furthermore, in this embodiment, during the determination of the initial digital adversarial sample, the positions of the geometric figures in the initial digital adversarial sample are constrained based on a preset mask. In a digital environment, the generated initial digital adversarial sample can be represented as:

[0061] ;

[0062] In the formula, It serves as a mask to restrict the position of geometric patterns, ensuring that these patterns do not exceed the area of ​​the detected target; The initial digital adversarial sample is generated from a simulated straight-line graph. The initial digital adversarial sample is generated based on a simulated triangle. X is the initial digital adversarial sample generated based on the simulation ellipse; X is the clean sample; S is the linear fusion function used to synthesize the generated infrared perturbation geometry with the clean sample.

[0063] Furthermore, in the process of generating adversarial examples from clean samples fused based on geometric perturbations, a particle swarm optimization algorithm can be used to determine the optimal adversarial sample for each type of geometric perturbation.

[0064] Furthermore, the geometric perturbations designed above can be generated by printing simple materials (such as stickers or printed paper) and attached to the target area. For example, taking pedestrians as the detection target, in the physical world, ice stickers used to generate geometric perturbations can be fixed to the inside of pedestrians' clothing, and physical samples can be captured by infrared sensors.

[0065] Step S12: Iteratively optimize the initial digital adversarial sample based on the preset sample optimization strategy to determine the optimized digital adversarial sample.

[0066] In this embodiment, after determining the initial digital adversarial sample, specifically, the initial digital adversarial sample is input into the infrared target detection model based on a preset sample optimization strategy, and the target object confidence score output by the model is used as the adversarial loss to iteratively optimize the initial digital adversarial sample to determine the optimized digital adversarial sample. That is, firstly, the combination of physical parameters related to the geometric position in the initial digital adversarial sample is used as particles and randomly initialized to determine the initial particle swarm and the velocity vector set corresponding to the initial particle swarm; then, adversarial samples are generated for each particle in the initial particle swarm to determine the adversarial sample generation result of the current iteration; the adversarial sample generation result is input into the infrared target detection model, and the current individual optimal solution and global optimal solution are determined using the corresponding target object confidence score output by the model; the velocity and position of each particle in the initial particle swarm are updated based on the inertia factor, learning factor, the velocity vector set corresponding to the initial particle swarm, the current individual optimal solution, and the global optimal solution to determine the current updated particle swarm; it is determined whether the current iteration meets the preset iteration termination condition, and when it does, the optimized digital adversarial sample is determined based on the updated particle swarm.

[0067] It is important to understand that this embodiment proposes an ADvIG (Adversarial Infrared Geometry) scheme, targeting black-box attack scenarios, which deploys adversarial attacks by querying the physical parameters of infrared geometric shapes. The goal of this scheme is to discover the physical parameters of the most adversarial infrared blocks (represented in line attacks, triangle attacks, and elliptical attacks, respectively). , and( , , These parameters are used to simulate digital infrared perturbations, preventing infrared target detectors from detecting targets (pedestrians, etc.) in the interfered image. A more realistic and challenging black-box attack scenario is considered, where detailed model architecture and parameter information are unavailable; only the input image and the model's output detection information are accessible. ,in, The position of the target object; The confidence level of the target object; The target object category is defined by the probability output by the model. Therefore, the ADvIG scheme uses the probability from the model's output. As a countermeasure to loss, the optimization objective of the scheme is formalized as minimization. :

[0068] ;

[0069] In the formula, For the pre-defined multi-scale transformation framework; A collection of examples of domain migration for transformation; This represents a subset of the objectives in the transformation instance. Therefore, it can be seen that achieving the above optimization objective using the particle swarm optimization algorithm requires further integration with a multi-scale transformation framework. The process for AdvIG optimization using the particle swarm optimization algorithm is as follows:

[0070] 1) Random Initialization Phase: The initialization step of particle swarm optimization is initiated by randomly generating a set of candidate solutions, including the population (POP) and its corresponding velocity vector set V.

[0071] ;

[0072] ;

[0073] In the formula, Indicates population size; The first in the population POP The candidate solutions are the physical parameters of the aforementioned simulation geometry; The first in the population Particles The direction of movement.

[0074] 2) Generate adversarial examples: After random initialization, the initial digital adversarial examples are generated using the aforementioned formula, which represents each particle in the population POP, i.e., an individual. Generate corresponding adversarial examples :

[0075] ;

[0076] In the formula, b represents the current iteration number of the population; This indicates the number of generations in the b-th population. Adversarial examples corresponding to each particle; Indicates the first One particle.

[0077] 3) Updating individual and global optimal solutions: Obtaining individual and global optimal solutions guides the direction of population evolution.

[0078] ;

[0079] In the formula, This indicates the number of generations (i.e., the b-th iterations) in the population. The individual historical optimal solution for each particle; This represents the globally optimal solution in the entire population of the b-th generation. In the u-th generation population, the th generation... The current optimal solution of an individual particle (which may be the historical optimal position of the individual). In the u-th generation population, the th generation... The objective function value of the individual solution of each particle.

[0080] 4) Update the speed and position information of individuals: After obtaining the optimal solution for individuals and the global optimal solution, use this information to guide the population to evolve in a direction that is easier to reach the optimal solution, and use the following formula to update the speed and position of individuals in the population.

[0081] ;

[0082] In the formula, and Hyperparameters for particle swarm optimization; Indicates the inertia factor; and The learning factor representing the particle; and A random number generated from a uniform distribution in the range [0, 1]. For the b+1 generation population, the The direction of movement of each particle; For the bth generation population, the first The direction of movement of each particle; For the b+1 generation population, the One particle; For the bth generation population, the first Each particle. When applying this particle swarm optimization algorithm for optimization, in addition to inputting hyperparameters... and Clean sample X, population size Furthermore, it is necessary to utilize the infrared target detection model f and the maximum number of iterations I of the population. After multiple iterations, the optimal solution of AdvIG will be obtained, resulting in optimized digital adversarial examples. .

[0083] In this way, by using particle swarm optimization algorithms to optimize the physical parameters of these geometries in a digital environment, the physical parameters of the most adversarial geometric perturbation patterns can be efficiently determined, and then the most adversarial examples can be determined based on these physical parameters.

[0084] Step S13: Perform multi-scale domain migration transformation enhancement on the optimized digital adversarial sample based on a preset multi-scale transformation framework to determine the target physical adversarial sample.

[0085] In this embodiment, combined with Figure 2As shown, the optimized digital adversarial sample will be enhanced by multi-scale domain migration transformation to determine the target physical adversarial sample. That is, the transformation distribution is simulated through the expected transformation framework, and the optimized digital adversarial sample is enhanced by multi-scale domain migration transformation based on the simulation results to determine the target physical adversarial sample.

[0086] It is important to understand that, in order to minimize the experimental difference during the transition from digital to physical attacks when implementing infrared black-box query attack testing, this embodiment utilizes a domain transfer strategy oriented towards multi-scale transformation. This strategy is an effective method for handling the transition from the digital to the physical domain in adversarial attacks. Its core idea is to introduce an expectation transformation function during adversarial example generation to transform the original input data into a sample with expected characteristics, enabling the model to produce the same prediction results on the transformed sample as on the original input. Specifically, this is achieved by simulating a transformation distribution. This transformation effectively addresses the robustness reduction problem caused by scale transformation when migrating to the simulation domain. The approach involves performing a multi-scale transformation on a sample that can successfully execute an adversarial attack in the digital domain. If the adversarial samples generated under each scale transformation can successfully execute the adversarial attack, the adversarial sample is considered robust. Therefore, the final generated physical adversarial sample can be represented as:

[0087] .

[0088] Step S14: Perform an infrared black-box query attack test on the infrared target detection model based on the target physical adversarial sample, and use the corresponding test results to determine the performance evaluation result of the infrared target detection model.

[0089] In this embodiment, after obtaining the most adversarial sample, the target physical adversarial sample, the performance evaluation of the model will be carried out using the target physical adversarial sample. That is, the target physical adversarial sample is input into the infrared target detection model to perform infrared black-box query attack tests on the infrared target detection model with line segment, triangle and / or elliptical geometric perturbations, and the detection results output by the model are determined. Based on the detection results, the current test results are determined, and the test results are analyzed to determine the performance evaluation results corresponding to the infrared target detection model.

[0090] In summary, the AdvIG scheme described in this embodiment utilizes geometric shapes as physical perturbations to perform efficient and stable visible light physical attacks on infrared target detection models, thereby achieving attack testing of infrared black-box query attacks on the models. Key points include: 1) An attack framework supporting multiple geometric perturbation models: A unified attack framework supporting line segment, triangle, and elliptical geometric perturbations is designed and implemented. Each perturbation mode can be used independently or in combination, demonstrating efficient attack capabilities in both the digital and physical worlds; 2) Lightweight deployment advantages: The designed perturbations can be printed and attached to the target area using simple materials (such as stickers or printed paper), offering stronger deployability and concealment compared to complex physical attack tools such as light bulb boards, infrared QR (Quick Response) devices, and aerogel, while also having lower deployment costs; 3) Comprehensive robustness and transferability experiments: Through comprehensive experimental analysis, including ablation experiments, transfer attack experiments, and anti-attack defense mechanism experiments, the effectiveness, robustness, and adaptability of the scheme under model generalization and environmental changes can be verified. Regarding concealment... This scheme demonstrates its stealth by placing geometric perturbations inside the pedestrian's clothing, making them virtually undetectable to human observers without the aid of infrared sensors, while also keeping the perturbation level below the baseline. Regarding robustness, it achieved a high success rate when deploying AdvIG attacks against various advanced infrared target detection models, further validating its robustness.

[0091] In other words, this embodiment significantly improves deployment simplicity and practical feasibility by constructing parameterizable geometric perturbations to replace traditional complex and expensive infrared attack materials. Employing a particle swarm optimization algorithm, it achieves low-query, high-efficiency physical attacks in black-box scenarios, balancing attack success rate and stealth. It exhibits good robustness, portability, and optimization performance under different physical conditions, demonstrating broad application prospects.

[0092] Therefore, in this embodiment, geometric simulation modeling is first performed, and initial digital adversarial samples are determined using the geometric modeling results and infrared detection image samples. Then, the samples are iteratively optimized based on a preset sample optimization strategy. A preset multi-scale transformation framework is used to enhance the determined optimized digital adversarial samples through multi-scale domain migration transformation. Subsequently, based on the determined target physical adversarial samples, an infrared black-box query attack test is performed on the infrared target detection model to complete the performance test of the infrared target detection model. In this way, multi-scale attack testing of the infrared target detection model can be achieved efficiently and effectively, and the concealment and robustness of the test are improved.

[0093] See Figure 4 As shown in the figure, this application also discloses a performance evaluation device for an infrared target detection model, including:

[0094] The initial sample generation module 11 is used to perform geometric simulation modeling in a digital environment and determine the initial digital adversarial sample based on the corresponding geometric modeling results and infrared detection image samples.

[0095] The sample optimization module 12 is used to iteratively optimize the initial digital adversarial sample based on a preset sample optimization strategy to determine the optimized digital adversarial sample.

[0096] The multi-scale transformation module 13 is used to perform multi-scale domain migration transformation enhancement on the optimized digital adversarial sample based on a preset multi-scale transformation framework, so as to determine the target physical adversarial sample.

[0097] The result determination module 14 is used to perform an attack test on the infrared target detection model based on the target physical adversarial sample, and to determine the performance evaluation result corresponding to the infrared target detection model using the corresponding test results.

[0098] In some specific embodiments, the initial sample generation module 11 may specifically include:

[0099] The simulation modeling unit is used to perform simulation modeling of lines, triangles, and ellipses in a digital environment to determine the initial modeling results.

[0100] The sample acquisition unit is used to acquire infrared detection image samples collected by the target infrared sensor; the target infrared sensor is an infrared sensor with a fixed preset area.

[0101] The first synthesis unit is used to synthesize the linear simulation results based on the geometric modeling results and the infrared detection image samples to determine the first synthesized sample.

[0102] The second synthesis unit is used to synthesize the triangle simulation results based on the geometric modeling results and the infrared detection image samples to determine the second synthesized sample;

[0103] The third synthesis unit is used to synthesize the ellipse simulation result based on the geometric modeling result and the infrared detection image sample to determine the third synthesized sample.

[0104] The sample processing unit is used to process the first synthetic sample, the second synthetic sample, and the third synthetic sample based on the particle swarm optimization algorithm to determine the processed initial digital adversarial sample.

[0105] In some specific embodiments, the performance evaluation device for the infrared target detection model may further include:

[0106] A geometric constraint unit is used to constrain the position of geometric figures in the initial digital adversarial sample based on a preset mask during the process of determining the initial digital adversarial sample.

[0107] In some specific embodiments, the sample optimization module 12 may specifically include:

[0108] The iterative optimization unit is used to input the initial digital adversarial sample into the infrared target detection model based on a preset sample optimization strategy, and use the target object confidence output by the model as the adversarial loss to iteratively optimize the initial digital adversarial sample in order to determine the optimized digital adversarial sample.

[0109] In some specific embodiments, the iterative optimization unit may specifically include:

[0110] The population initialization subunit is used to randomly initialize the combination of physical parameters related to the geometric position of the initial digital adversarial sample as particles, so as to determine the initial particle swarm and the velocity vector set corresponding to the initial particle swarm.

[0111] An adversarial sample generation subunit is used to generate adversarial samples for each of the particles in the initial particle swarm, so as to determine the adversarial sample generation result of the current iteration round.

[0112] The optimal solution determination subunit is used to input the adversarial sample generation result into the infrared target detection model, and use the corresponding target object confidence score output by the model to determine the current individual optimal solution and the global optimal solution;

[0113] The particle update subunit is used to update the velocity and position of each particle in the initial particle swarm based on the inertia factor, the learning factor, the velocity vector set corresponding to the initial particle swarm, the current individual optimal solution, and the global optimal solution, so as to determine the current updated particle swarm.

[0114] The condition judgment subunit is used to determine whether the preset iteration termination condition is met, and when it is met, to determine the optimized digital adversarial sample based on the updated particle swarm.

[0115] In some specific embodiments, the multi-scale transformation module 13 may specifically include:

[0116] The transformation enhancement unit is used to simulate the transformation distribution through the desired transformation framework and perform multi-scale domain migration transformation enhancement on the optimized digital adversarial sample based on the simulation results to determine the target physical adversarial sample.

[0117] In some specific embodiments, the result determination module 14 may specifically include:

[0118] The detection result determination unit is used to input the target physical adversarial sample into the infrared target detection model to perform infrared black-box query attack tests on the infrared target detection model with line segment, triangle and / or elliptical geometric perturbations, and determine the detection result output by the model.

[0119] The result analysis unit is used to determine the current test result based on the detection result, and to analyze the test result to determine the performance evaluation result corresponding to the infrared target detection model.

[0120] Furthermore, embodiments of this application also disclose an electronic device, Figure 5 This is a structural diagram of an electronic device 20 according to an exemplary embodiment. The content of the diagram should not be construed as limiting the scope of this application.

[0121] Figure 5 This is a schematic diagram of the structure of an electronic device 20 provided in an embodiment of this application. Specifically, the electronic device 20 may include: at least one processor 21, at least one memory 22, a power supply 23, a communication interface 24, an input / output interface 25, and a communication bus 26. The memory 22 stores a computer program, which is loaded and executed by the processor 21 to implement the relevant steps in the performance evaluation method of the infrared target detection model disclosed in any of the foregoing embodiments. Furthermore, the electronic device 20 in this embodiment may specifically be an electronic computer.

[0122] In this embodiment, the power supply 23 is used to provide operating voltage for each hardware device on the electronic device 20; the communication interface 24 can create a data transmission channel between the electronic device 20 and external devices, and the communication protocol it follows can be any communication protocol applicable to the technical solution of this application, and is not specifically limited here; the input / output interface 25 is used to acquire external input data or output data to the outside world, and its specific interface type can be selected according to specific application needs, and is not specifically limited here.

[0123] In addition, the memory 22, as a carrier for resource storage, can be a read-only memory, random access memory, disk or optical disk, etc. The resources stored thereon can include operating system 221, computer program 222, etc., and the storage method can be temporary storage or permanent storage.

[0124] The operating system 221 is used to manage and control the various hardware devices on the electronic device 20 and the computer program 222, which may be Windows Server, Netware, Unix, Linux, etc. In addition to including a computer program capable of performing the performance evaluation method of the infrared target detection model executed by the electronic device 20 as disclosed in any of the foregoing embodiments, the computer program 222 may further include computer programs capable of performing other specific tasks.

[0125] Furthermore, this application also discloses a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, it implements the performance evaluation method of the aforementioned disclosed infrared target detection model. Specific steps of this method can be found in the corresponding content disclosed in the foregoing embodiments, and will not be repeated here.

[0126] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section.

[0127] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0128] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented directly by hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.

[0129] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0130] The technical solutions provided in this application have been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for performance evaluation of an infrared target detection model, characterized in that, The method comprises the steps of: carrying out simulation modeling of geometric figures in a digital environment, and determining initial digital adversarial samples based on corresponding geometric modeling results and infrared detection image samples; iteratively optimizing the initial digital adversarial samples based on a preset sample optimization strategy to determine optimized digital adversarial samples; performing multi-scale domain migration transformation enhancement on the optimized digital adversarial samples based on a preset multi-scale transformation framework to determine target physical adversarial samples; performing an infrared black-box query attack test on an infrared target detection model based on the target physical adversarial samples, and determining performance evaluation results of the infrared target detection model using corresponding test results.

2. The method of claim 1, wherein, The method of carrying out simulation modeling of geometric figures in a digital environment, and determining initial digital adversarial samples based on corresponding geometric modeling results and infrared detection image samples comprises the steps of: carrying out simulation modeling of straight lines, triangles and ellipses respectively in a digital environment to determine initial modeling results; obtaining infrared detection image samples collected by a target infrared sensor; the target infrared sensor is an infrared sensor fixed in a preset area; synthesizing the straight line simulation result of the geometric modeling result and the infrared detection image samples to determine a first synthesized sample; synthesizing the triangle simulation result of the geometric modeling result and the infrared detection image samples to determine a second synthesized sample; synthesizing the ellipse simulation result of the geometric modeling result and the infrared detection image samples to determine a third synthesized sample; processing the first synthesized sample, the second synthesized sample and the third synthesized sample based on a particle swarm optimization algorithm to determine processed initial digital adversarial samples.

3. The method of claim 2, wherein, The method further comprises the steps of: constraining the position of the geometric figure in the initial digital adversarial samples based on a preset mask during the determination of the initial digital adversarial samples.

4. The method of claim 1 to 3, wherein, The method of iteratively optimizing the initial digital adversarial samples based on a preset sample optimization strategy to determine optimized digital adversarial samples comprises the steps of: inputting the initial digital adversarial samples into an infrared target detection model based on a preset sample optimization strategy, and iteratively optimizing the initial digital adversarial samples using the target object confidence output by the model as an adversarial loss to determine optimized digital adversarial samples.

5. The method of claim 4, wherein, The method of inputting the initial digital adversarial samples into an infrared target detection model based on a preset sample optimization strategy, and iteratively optimizing the initial digital adversarial samples using the target object confidence output by the model as an adversarial loss comprises the steps of: randomly initializing a combination of physical parameters related to the position of the geometric figure in the initial digital adversarial samples as particles to determine an initial particle swarm and a set of velocity vectors corresponding to the initial particle swarm; generating adversarial samples for each particle in the initial particle swarm to determine adversarial sample generation results of the current iteration round; inputting the adversarial sample generation results into an infrared target detection model, and determining a current individual optimal solution and a global optimal solution using the corresponding target object confidence output by the model; updating the speed and position of each particle in the initial particle swarm based on an inertia factor, a learning factor, the set of speed vectors corresponding to the initial particle swarm, the current individual optimal solution and the global optimal solution to determine a current updated particle swarm; determining whether a preset iteration termination condition is met, and when the preset iteration termination condition is met, determining the optimized digital adversarial sample based on the updated particle swarm.

6. The method of claim 1, wherein, the multi-scale domain migration transformation enhancement of the optimized digital adversarial sample based on a preset multi-scale transformation framework to determine a target physical adversarial sample, including: simulating a transformation distribution through an expected transformation framework, and performing multi-scale domain migration transformation enhancement on the optimized digital adversarial sample based on the simulation result to determine a target physical adversarial sample.

7. The method of claim 1, wherein, the attack test of the infrared target detection model based on the target physical adversarial sample, and the performance evaluation result corresponding to the infrared target detection model is determined by using the corresponding test result, including: inputting the target physical adversarial sample into the infrared target detection model to perform an infrared black box query attack test of the infrared target detection model on line segment type, triangle type and / or ellipse type geometric perturbation, and determining the detection result output by the model; determining the current test result based on the detection result, and analyzing the test result to determine the performance evaluation result corresponding to the infrared target detection model.

8. An apparatus for performance evaluation of an infrared target detection model, the apparatus comprising: a target detection model; a target detection model performance evaluation module configured to evaluate performance of the target detection model. including: an initial sample generation module configured to simulate geometric modeling in a digital environment, and determine an initial digital adversarial sample based on the corresponding geometric modeling result and an infrared detection image sample; a sample optimization module configured to iteratively optimize the initial digital adversarial sample based on a preset sample optimization strategy to determine an optimized digital adversarial sample; a multi-scale transformation module configured to perform multi-scale domain migration transformation enhancement on the optimized digital adversarial sample based on a preset multi-scale transformation framework to determine a target physical adversarial sample; a result determination module configured to perform an attack test of an infrared target detection model based on the target physical adversarial sample, and determine a performance evaluation result corresponding to the infrared target detection model by using a corresponding test result.

9. An electronic device, comprising: including: a memory configured to save a computer program; a processor configured to execute the computer program to implement the performance evaluation method of the infrared target detection model according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, a computer program is saved, and the computer program is executed by a processor to implement the performance evaluation method of the infrared target detection model according to any one of claims 1 to 7.

Images