Confrontation pattern generation method and device, equipment, medium and product

Abstract

The invention discloses an adversarial pattern generation method and device, equipment, a medium and a product, and relates to the technical field of adversarial attacks, and the method comprises the steps: collecting environment data through a sensor; inputting the environment data into a pre-constructed confrontation pattern generation model, performing pattern generation based on the environment data through the confrontation pattern generation model to obtain a confrontation pattern, the confrontation pattern generation model being obtained by distilling an initial confrontation pattern generation model, and the confrontation pattern generation model being obtained by distilling an initial confrontation pattern generation model; the initial confrontation pattern generation model is constructed based on a final generative confrontation network, a task pool and an attention module. Therefore, the problem that the generated confrontation pattern cannot meet the camouflage requirement in allusion to a multi-modal fusion recognition system is solved, and the generation efficiency of the confrontation pattern is improved.

Description

Technical Field

[0001] The present application relates to the field of counter-attack technology, and in particular to a method, device, equipment, medium and product for generating a counter-attack pattern. Background Art

[0002] Currently, multimodal target recognition technology based on deep learning has been widely used in the field of military security, which poses a severe challenge to traditional camouflage methods.

[0003] Existing adversarial sample generation technology has the following significant defects in defense effectiveness and actual combat adaptability: 1) Static defense methods (such as FGSM and PGD) cannot adapt to enemy algorithm upgrades in real time, and digital adversarial samples are difficult to convert into physical camouflage devices; 2) Cross-modal collaboration is insufficient, and existing technologies only target a single sensor (such as visible light / infrared) and cannot simultaneously deceive multi-modal fusion recognition systems; 3) The generation diversity is low in small sample scenarios, and there is a lack of physical parameter constraints (such as material reflectivity), resulting in poor adaptability to complex environments.

[0004] The above-mentioned technical defects make it difficult for existing adversarial sample generation methods to meet the requirements of dynamic confrontation, multimodal compatibility, and physically feasible camouflage in modern warfare environments. It is urgent to develop intelligent camouflage generation technology with environmental adaptability.

[0005] The above content is only used to assist in understanding the technical solution of this application and does not constitute an admission that the above content is prior art. Summary of the Invention

[0006] The main purpose of this application is to provide a method, device, equipment, medium and product for generating an adversarial pattern, aiming to solve the technical problem that the generated adversarial pattern cannot meet the camouflage requirements when used in a multimodal fusion recognition system.

[0007] To achieve the above objectives, the present application proposes a method for generating a countermeasure pattern, which includes:

[0008] Collect environmental data through sensors;

[0009] The environmental data is input into a pre-constructed adversarial pattern generation model, and the adversarial pattern generation model is used to generate a pattern based on the environmental data to obtain an adversarial pattern. The adversarial pattern generation model is obtained by distilling an initial adversarial pattern generation model, and the initial adversarial pattern generation model is constructed based on the final generative adversarial network, the task pool, and the attention module.

[0010] In one embodiment, before the step of inputting the environmental data into a pre-built adversarial pattern generation model and generating a pattern based on the environmental data by the adversarial pattern generation model to obtain the adversarial pattern, the method further includes:

[0011] Integrate several AI recognition models to obtain a task pool, and input training data into the task pool to obtain a task instance;

[0012] Based on the task instance, the initialization parameters of the initial generator and the initial discriminator of the initial generative adversarial network are optimized by a meta-learning algorithm to obtain a generator and a discriminator;

[0013] Update the initial generative adversarial network according to the generator and the discriminator to obtain a final generative adversarial network;

[0014] An initial adversarial pattern generation model is constructed based on the final generative adversarial network, the task pool, and the attention module;

[0015] Model distillation is performed on the initial adversarial pattern generation model to obtain an adversarial pattern generation model.

[0016] In one embodiment, the step of optimizing the initialization parameters of the initial generator and the initial discriminator of the initial generative adversarial network by a meta-learning algorithm based on the task instance to obtain the generator and the discriminator includes:

[0017] Inputting the task instance into the initial generator to obtain an adversarial sample;

[0018] Classify the adversarial sample by the initial discriminator to obtain a classification result and an evaluation result;

[0019] Calculate the adversarial sample using a diversity loss function to obtain a loss value;

[0020] Based on the classification result, the evaluation result and the loss value, gradient updating is performed on the initialization parameters of the initial generator through a meta-learning algorithm to obtain a generator;

[0021] Based on the classification result and the evaluation result, the initialization parameters of the initial discriminator are gradient updated through a meta-learning algorithm to obtain a discriminator.

[0022] In one embodiment, the step of inputting the environmental data into a pre-built adversarial pattern generation model, and generating a pattern based on the environmental data by the adversarial pattern generation model to obtain the adversarial pattern includes:

[0023] generating an initial adversarial pattern by the generator according to the environmental data;

[0024] Extracting features from the initial adversarial pattern using the attention module and the encoder to obtain pattern features;

[0025] According to the pattern features, a vector quantizer is used to reconstruct the pattern to obtain an adversarial pattern.

[0026] In one embodiment, before the step of generating the initial confrontation pattern by the generator according to the environmental data, the method further includes:

[0027] Extracting recognition model information of the adversarial end from the environmental data;

[0028] According to the identification model information, a model upgrade prediction is performed through a simulation platform to obtain a prediction result;

[0029] Based on the prediction results, reinforcement learning training is performed on the generator to obtain a trained generator.

[0030] In one embodiment, after the step of inputting the environmental data into a pre-built adversarial pattern generation model, and generating a pattern based on the environmental data by the adversarial pattern generation model to obtain the adversarial pattern, the method further comprises:

[0031] Receiving device information of the physical confrontation device, and parsing the device information to obtain device parameters;

[0032] Differentiable rendering is performed on the physical confrontation device according to the device parameters and the confrontation pattern to obtain a rendering result.

[0033] In addition, to achieve the above-mentioned purpose, the present application also proposes a countermeasure pattern generation device, which includes:

[0034] A collection module, used to collect environmental data through sensors;

[0035] A generation module is used to input the environmental data into a pre-built adversarial pattern generation model, and generate a pattern based on the environmental data through the adversarial pattern generation model to obtain an adversarial pattern. The adversarial pattern generation model is obtained by distilling an initial adversarial pattern generation model, and the initial adversarial pattern generation model is constructed based on the final generative adversarial network, the task pool and the attention module.

[0036] In addition, to achieve the above-mentioned purpose, the present application also proposes a countermeasure pattern generation device, which includes: a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the computer program is configured to implement the steps of the countermeasure pattern generation method described above.

[0037] In addition, to achieve the above-mentioned purpose, the present application also proposes a storage medium, which is a computer-readable storage medium and stores a computer program. When the computer program is executed by a processor, the steps of the countermeasure pattern generation method described above are implemented.

[0038] In addition, to achieve the above-mentioned purpose, the present application also provides a computer program product, which includes a computer program. When the computer program is executed by a processor, it implements the steps of the countermeasure pattern generation method as described above.

[0039] One or more technical solutions proposed in this application have at least the following technical effects:

[0040] The embodiments of the present application propose an adversarial pattern generation method, apparatus, device, medium, and product. The method collects environmental data through sensors, inputs the environmental data into a pre-built adversarial pattern generation model, and generates a pattern based on the environmental data using the adversarial pattern generation model to obtain an adversarial pattern. The adversarial pattern generation model is obtained by distilling an initial adversarial pattern generation model, which is constructed based on a final generative adversarial network, a task pool, and an attention module. Thus, real-time environmental data is collected through sensors, and then pattern generation is performed using a pre-built adversarial pattern generation model to obtain an adversarial pattern. The adversarial pattern generation model is obtained by distilling the initial adversarial pattern generation model constructed from the final generative adversarial network, the task pool, and the attention module. This solves the problem that the generated adversarial pattern fails to meet camouflage requirements in multimodal fusion recognition systems, thereby improving the efficiency of adversarial pattern generation. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application.

[0042] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0043] Figure 1 A schematic diagram of a flow chart provided for the first embodiment of the method for generating a countermeasure pattern of the present application;

[0044] Figure 2 A schematic diagram of the flow chart provided for the second embodiment of the method for generating a countermeasure pattern of the present application;

[0045] Figure 3A schematic diagram of a brief flow chart of the method for generating a countermeasure pattern provided in Example 2 of the present application;

[0046] Figure 4 This is a schematic diagram of the module structure of the countermeasure pattern generating device according to an embodiment of the present application;

[0047] Figure 5 Schematic diagram of the device structure of the hardware operating environment involved in the countermeasure pattern generation method in the embodiment of the present application.

[0048] The purpose, features and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0049] It should be understood that the specific embodiments described herein are merely used to explain the technical solutions of the present application and are not intended to limit the present application.

[0050] In order to better understand the technical solution of the present application, a detailed description will be given below in conjunction with the accompanying drawings and specific implementation methods.

[0051] The main solution of the embodiment of the present application is: integrating several AI recognition models to obtain a task pool, inputting training data into the task pool to obtain task instances; based on the task instances, optimizing the initialization parameters of the initial generator and initial discriminator of the initial generative adversarial network through a meta-learning algorithm to obtain a generator and a discriminator; updating the initial generative adversarial network according to the generator and the discriminator to obtain a final generative adversarial network; constructing an initial adversarial pattern generation model based on the final generative adversarial network, the task pool, and the attention module; performing model distillation on the initial adversarial pattern generation model to obtain an adversarial pattern generation model. Inputting the task instance into the initial generator to obtain an adversarial sample; classifying the adversarial sample through the initial discriminator to obtain a classification result and an evaluation result; calculating the adversarial sample through a diversity loss function to obtain a loss value; based on the classification result, evaluation result, and loss value, performing a gradient update on the initialization parameters of the initial generator through a meta-learning algorithm to obtain a generator; based on the classification result and evaluation result, performing a gradient update on the initialization parameters of the initial discriminator through a meta-learning algorithm to obtain a discriminator. According to the environmental data, an initial adversarial pattern is generated by the generator; the initial adversarial pattern is subjected to feature extraction by the attention module and the encoder to obtain pattern features; according to the pattern features, the pattern is reconstructed by the vector quantizer to obtain an adversarial pattern. The recognition model information of the adversarial end in the environmental data is extracted; according to the recognition model information, a model upgrade prediction is performed through the simulation platform to obtain a prediction result; based on the prediction result, the generator is subjected to reinforcement learning training to obtain a trained generator. The device information of the physical adversarial device is received, and the device information is parsed to obtain device parameters; according to the device parameters and the adversarial pattern, the physical adversarial device is subjected to differentiable rendering to obtain a rendering result. This solves the problem that the generated adversarial pattern cannot meet the camouflage requirements for the multimodal fusion recognition system, realizes the generation of adversarial patterns, and improves the efficiency of adversarial pattern generation. Based on the solution of the present invention, a countermeasure pattern generation method is designed based on the problem that traditional methods rely on visual interference (such as oil paint, camouflage patterns) or thermal radiation shielding, but cannot cope with AI-based multimodal recognition systems, and the existing technology has difficulty in coping with multimodal systems that integrate vision, thermal imaging and gait analysis. The effectiveness of the countermeasure pattern generation method of the present invention is verified when generating countermeasure patterns. Finally, the efficiency of countermeasure pattern generation using the method of the present invention is significantly improved.

[0052] In this embodiment, for ease of description, the following description is made with the countermeasure pattern generating device as the execution subject.

[0053] Since adversarial sample generation technology in existing technologies is mainly used to deceive deep learning models, existing methods have significant defects in dynamic adaptability, cross-modal defense and physical feasibility. First, traditional methods such as the fast gradient sign method (FGSM) and projected gradient descent (PGD) rely on fixed perturbation patterns and cannot cope with real-time upgrades of enemy AI algorithms. Second, existing technologies are mostly targeted at single sensors (such as visible light or infrared) and have difficulty coping with multimodal recognition systems such as fused vision, thermal imaging, and gait analysis. Finally, traditional generative adversarial networks are prone to pattern collapse in small sample scenarios, the generated samples are highly repetitive, and the computational efficiency is low, which limits the effectiveness of practical applications.

[0054] This application provides a solution to distill the initial adversarial pattern generation model constructed by the final generative adversarial network, task pool and attention module to obtain an adversarial pattern generation model. Finally, the adversarial pattern generation model is used to generate adversarial patterns based on environmental data collected by sensors to provide users with better services.

[0055] As can be seen from the above embodiments, the present application collects environmental data through sensors; the environmental data is input into a pre-constructed adversarial pattern generation model; the adversarial pattern generation model generates a pattern based on the environmental data to obtain an adversarial pattern; the adversarial pattern generation model is obtained by distilling an initial adversarial pattern generation model, which is constructed based on a final generative adversarial network, a task pool, and an attention module. Thus, real-time environmental data is collected through sensors, and then pattern generation is performed through a pre-constructed adversarial pattern generation model to obtain an adversarial pattern, wherein the adversarial pattern generation model is obtained by distilling the initial adversarial pattern generation model constructed by the final generative adversarial network, the task pool, and the attention module. This solves the problem that the generated adversarial pattern cannot meet the camouflage requirements in a multimodal fusion recognition system, thereby improving the efficiency of adversarial pattern generation.

[0056] It should be noted that the execution subject of this embodiment can be a computing service device with data processing, network communication, and program execution capabilities, such as a tablet computer, personal computer, or mobile phone, or an electronic device capable of performing the aforementioned functions, such as a countermeasure pattern generation device. This embodiment and the following embodiments will be described below using a countermeasure pattern generation device as an example.

[0057] Based on this, the embodiment of the present application provides a method for generating a countermeasure pattern, referring to Figure 1 , Figure 1 This is a flowchart of the first embodiment of the method for generating a countermeasure pattern of the present application.

[0058] In this embodiment, the method for generating a countermeasure pattern includes steps S01 to S02:

[0059] Step S01, collecting environmental data through sensors;

[0060] It should be clear that the current multimodal target recognition technology based on deep learning has been widely used in the field of military security, which poses a severe challenge to traditional camouflage methods.

[0061] Existing adversarial sample generation technology has the following significant defects in defense effectiveness and actual combat adaptability: 1) Static defense methods (such as FGSM and PGD) cannot adapt to enemy algorithm upgrades in real time, and digital adversarial samples are difficult to convert into physical camouflage devices; 2) Cross-modal collaboration is insufficient, and existing technologies only target a single sensor (such as visible light / infrared) and cannot simultaneously deceive multi-modal fusion recognition systems; 3) The generation diversity is low in small sample scenarios, and there is a lack of physical parameter constraints (such as material reflectivity), resulting in poor adaptability to complex environments.

[0062] Therefore, in this embodiment, sensors are used to collect specific environmental data. Among them, the sensors and environmental data in this embodiment are mainly aimed at military application scenarios, so the sensors include radar sensors (used to detect and track the position, speed and direction of enemy aircraft, ships, missiles, vehicles and other objects. Radar waves can penetrate obstacles (such as haze or at night) and reflect back information about enemy equipment for real-time monitoring), infrared sensors (based on the principle of thermal imaging, which can capture the thermal radiation of enemy equipment and identify heat sources. They are often used to detect heat signals of enemy vehicles, aircraft, drones and other equipment, especially at night or in poor visibility environments. It is very effective), sonar sensors (used in underwater environments, sonar can detect the heat radiation of enemy equipment through infrared sensors). Transmitting sound waves and receiving echoes to identify the presence and location of enemy submarines, ships and other underwater equipment), optical sensors (including optical cameras, laser rangefinders, etc., used to detect visual information of enemy equipment, such as ground vehicles, buildings, aircraft, etc., and LiDAR (LiDAR) can also be used to generate high-resolution three-dimensional maps to analyze the structure or location of enemy equipment) and radio spectrum sensors (these sensors can monitor the spectrum activities of enemy equipment such as radio communications, radar signals, data links, etc., to help analyze the enemy's communications, positioning, electronic warfare and other strategies. For example, electronic reconnaissance equipment can capture enemy electromagnetic radiation signals and identify the characteristics of enemy radar, communication equipment and other electronic equipment).

[0063] The collected environmental information includes enemy equipment information, including location and orientation (sensors can determine the geographic location and movement trajectory of enemy equipment through radar, infrared sensors, optical sensors, etc.), equipment type and characteristics (through signal analysis, the type of equipment used by the enemy (such as drones, tanks, aircraft, submarines, etc.) and its working status can be identified), electronic signals and communication content (through radio spectrum sensors, the enemy's communication signals, radar waves or other electronic signals can be captured and analyzed to obtain information such as the frequency used, communication content, equipment model, etc.) and speed and direction of movement (using radar, infrared sensors, acceleration sensors, etc., the movement status of enemy equipment can be analyzed, including its speed, acceleration and direction).

[0064] In the above-mentioned embodiment, corresponding sensors are used for military application scenarios to collect environmental information. In addition to conventional weather, humidity and images, it also includes enemy equipment information and location information, which can be used to generate countermeasure patterns in a targeted manner, solving the problem that countermeasure patterns cannot achieve camouflage effects.

[0065] Step S02: input the environmental data into a pre-constructed adversarial pattern generation model, and generate a pattern based on the environmental data through the adversarial pattern generation model to obtain an adversarial pattern. The adversarial pattern generation model is obtained by distilling the initial adversarial pattern generation model, and the initial adversarial pattern generation model is constructed based on the final generative adversarial network, task pool and attention module.

[0066] In this embodiment, after the environmental data is collected, it is added to the pre-acquired adversarial pattern generation model to generate an adversarial pattern, and an adversarial pattern used to confuse the opponent and camouflage is obtained. The adversarial pattern generation model includes a final generative adversarial network, a task pool, and an attention module. In addition, it should also include an encoder and a vector quantizer. In order to reduce the computing load, this embodiment uses knowledge distillation technology to compress the relatively bloated model into a lightweight generator and embed it into an edge AI chip (such as the NVIDIA Jetson series). In actual combat, soldiers collect environmental data (such as light intensity and enemy equipment type) in real time through the built-in sensors in the mask. The generator only needs to execute the adversarial pattern generation process to output an adversarial pattern adapted to the current scene.

[0067] In this embodiment, a pre-built adversarial pattern generation model is used to analyze the environmental information collected by the sensor, and thereby obtains an adversarial pattern for camouflage, thereby solving the problem that the generated adversarial pattern cannot meet the camouflage requirements when used in a multimodal fusion recognition system.

[0068] Before implementation of the solution in this embodiment, it is necessary to construct a countermeasure pattern generation model in advance. Therefore, before the step S02 of inputting the environmental data into the pre-constructed countermeasure pattern generation model and generating a pattern based on the environmental data by the countermeasure pattern generation model to obtain the countermeasure pattern, the method further includes:

[0069] Step S0201: Integrate several AI recognition models to obtain a task pool, and input training data into the task pool to obtain a task instance;

[0070] Step S0202: Based on the task instance, optimizing the initialization parameters of the initial generator and the initial discriminator of the initial generative adversarial network through a meta-learning algorithm to obtain a generator and a discriminator;

[0071] Step S0203, updating the initial generative adversarial network according to the generator and the discriminator to obtain a final generative adversarial network;

[0072] Step S0204, constructing an initial adversarial pattern generation model based on the final generative adversarial network, the task pool, and the attention module;

[0073] Step S0205 , performing model distillation on the initial adversarial pattern generation model to obtain an adversarial pattern generation model.

[0074] In order to cope with the multimodal recognition model, it is necessary to build a model that can be upgraded. Therefore, the meta-adversarial learning framework is constructed first. The system consists of a task pool, a meta-training module and a fast adaptation module. A variety of AI recognition models (such as convolutional neural network CNN, Transformer, multimodal fusion model) are integrated in the task pool. Each model corresponds to a meta-task. The MAML algorithm is used in the meta-training stage to optimize the initialization parameters of the generator and discriminator through multiple rounds of iterations. For example, the loss function of the generator G introduces diversity loss to calculate the distance of the generated samples in the feature space to avoid mode collapse. The discriminator D adopts a dual-branch structure to handle edge detection and noise perception tasks respectively, and improves the robustness of adversarial samples through comparative learning.

[0075] To be more specific, the AI ​​recognition model is first integrated with the task pool construction. Multiple AI recognition models (such as vision, thermal imaging, gait analysis, etc.) are integrated together to form a task pool. Each recognition model is responsible for solving different types of recognition tasks (such as image classification, heat map analysis, behavior recognition, etc.). The training data is input into the task pool to generate a set of task instances. Each task instance represents a specific input data and a corresponding processing task, which can be classification, detection or other types of tasks. Subsequently, a meta-learning algorithm is used to optimize the parameters of the generative adversarial network (GAN). Based on the generated task instances, a meta-learning algorithm is used for optimization. The purpose of the meta-learning algorithm is to enable the generative adversarial network (GAN) to quickly adjust its parameters to adapt to different scenarios when facing new tasks. In this process, the initialization parameters of the initial generator and discriminator will be adjusted through the meta-learning algorithm to optimize their performance, so that the generative adversarial network can generate higher quality adversarial samples. Through multiple iterations, the optimization process continues until the parameters of the generator and discriminator reach the optimal state.

[0076] According to the optimized generator and discriminator, the initial generative adversarial network is updated. The generator is responsible for generating adversarial samples, while the discriminator determines whether the sample is a real sample or a generated sample. The updated generative adversarial network will more effectively generate adversarial samples that can deceive the AI ​​recognition system, enabling it to cope with a variety of task instances in actual use. Subsequently, the initial adversarial pattern generation model can be constructed based on the final updated generative adversarial network, task pool and attention module. The recognition capabilities of different tasks in the task pool can be weighted and integrated through the attention module to generate adversarial patterns for multiple sensors and tasks. The role of the attention module is that it can help the generative model better focus on the key features in the task, thereby improving the effectiveness of generating adversarial patterns.

[0077] Finally, the initial adversarial pattern generation model is subjected to model distillation. Model distillation is a method of training a small model to imitate the behavior of a large model. Here, the distillation process makes the generation model more streamlined and can improve inference efficiency while retaining important features. The adversarial pattern generation model after distillation will become more efficient and can generate adversarial samples that adapt to multimodal environments and complex recognition systems, thereby improving the practicality and anti-interference ability of the model.

[0078] More specifically, in step S0202, based on the task instance, the initialization parameters of the initial generator and the initial discriminator of the initial generative adversarial network are optimized by a meta-learning algorithm to obtain the generator and the discriminator, including:

[0079] Step S02021, inputting the task instance into the initial generator to obtain an adversarial sample;

[0080] Step S02022: classify the adversarial sample using the initial discriminator to obtain a classification result and an evaluation result;

[0081] Step S02023, calculating the adversarial sample using a diversity loss function to obtain a loss value;

[0082] Step S02024: Based on the classification result, the evaluation result, and the loss value, gradient update the initialization parameters of the initial generator using a meta-learning algorithm to obtain a generator;

[0083] Step S02025: Based on the classification result and the evaluation result, the initialization parameters of the initial discriminator are gradient updated by a meta-learning algorithm to obtain a discriminator.

[0084] It should be clear that task instances refer to specific task data after processing by the task pool, including possible images, text or other types of data. These task instances are input into the initial generator. The goal of the generator is to create adversarial samples based on the task instances. Adversarial samples are carefully designed and can confuse or disrupt the input data of AI recognition systems (such as classifiers, detectors, etc.).

[0085] The initial discriminator receives the adversarial sample and classifies it. The discriminator's task is to determine whether the input sample is "real" (that is, from the real data distribution) or "forged" (that is, the adversarial sample generated by the generator). The discriminator will give a classification result, indicating whether the adversarial sample is identified as a real sample or a forged sample. At the same time, it will also output an evaluation result to measure the credibility of the classification or the degree of deception of the adversarial sample.

[0086] In order to optimize the generator and discriminator, the loss value is calculated through the diversity loss function in this embodiment. The goal of the diversity loss function is to ensure that the generated adversarial samples have sufficient diversity in the feature space. This is to prevent the generator from only learning how to generate a single adversarial sample and improve the robustness of the model. The calculation of the loss value is based on the diversity of the adversarial samples, that is, how to ensure that these adversarial samples are not too similar, but can widely cover various possible perturbations, making the output of the generator more extensive and challenging.

[0087] Subsequently, the parameters of the initial generator are optimized through a meta-learning algorithm. The key to meta-learning is that it not only optimizes a single task, but also optimizes the model by learning how to quickly adapt to changes in different tasks. Based on the classification results, evaluation results and diversity loss values ​​given by the discriminator, the meta-learning algorithm adjusts the initialization parameters of the initial generator through a gradient update mechanism. This optimization process allows the generator to generate more confusing adversarial samples, thereby better "deceiving" the discriminator.

[0088] Similarly, based on the classification and evaluation results of the discriminator, the meta-learning algorithm optimizes the discriminator's parameters. The discriminator's goal is to distinguish adversarial samples from real samples as accurately as possible. Therefore, the meta-learning algorithm adjusts the discriminator's initialization parameters through a gradient update mechanism to improve its discrimination ability when facing adversarial samples. This embodiment, through the above-mentioned scheme, specifically collects environmental data through sensors; inputs the environmental data into a pre-built adversarial pattern generation model; and generates an adversarial pattern based on the environmental data using the pre-built adversarial pattern generation model. The adversarial pattern generation model is obtained by distilling the initial adversarial pattern generation model constructed from the final generative adversarial network, the task pool, and the attention module. Thus, by collecting real-time environmental data through sensors, the pre-built adversarial pattern generation model is used to generate an adversarial pattern. The adversarial pattern generation model is obtained by distilling the initial adversarial pattern generation model constructed from the final generative adversarial network, the task pool, and the attention module. This solves the problem of generated adversarial patterns failing to meet camouflage requirements in multimodal fusion recognition systems and improves the efficiency of adversarial pattern generation.

[0089] Based on the first embodiment of the present application, in the second embodiment of the present application, the same or similar contents as those in the first embodiment can be referred to the above introduction and will not be described in detail later. Figure 2 In step S02, the method for generating a countermeasure pattern further includes steps S021 to S023:

[0090] Step S021, generating an initial confrontation pattern through the generator according to the environmental data;

[0091] Step S022: extracting features from the initial adversarial pattern using the attention module and the encoder to obtain pattern features;

[0092] Step S023 : Reconstructing the pattern through a vector quantizer according to the pattern features to obtain a countermeasure pattern.

[0093] In the above embodiment, it is clear that the adversarial pattern in this embodiment is generated by the generator. In actual use, environmental data refers to external information related to the task, which may include visual data, sensor data or other types of input. When generating adversarial patterns, environmental data provides the generator with the necessary background and constraints.

[0094] The task of the generator is to generate an initial adversarial pattern based on these environmental data. This pattern may be an image, graph, or other type of visual representation. Its purpose is to be challenging and able to disturb the existing system. After the initial adversarial pattern is generated, it needs to be converted into information that is easier to process through a feature extraction step. The goal of this step is to extract important features from the pattern so that subsequent processing can utilize these features more efficiently.

[0095] The attention module is used to highlight the most relevant parts of the pattern. Through the attention mechanism, the model can dynamically focus on the areas or features that have the greatest impact on the task, thereby improving the representation ability of the pattern. The encoder maps and compresses the pattern and converts it into a more compact representation. The encoder usually converts the original pattern into a high-dimensional feature vector through a convolutional neural network (CNN) or other neural network structures. These feature vectors contain the key information of the pattern for subsequent processing.

[0096] Finally, after obtaining the pattern's features, the next step is to reconstruct the pattern through a vector quantizer. The function of a vector quantizer is to map the pattern's feature vectors to a predefined discrete set, which typically includes common pattern templates or representative features. The goal of the vector quantizer is to optimize the pattern based on its characteristics, ensuring that the reconstructed pattern remains visually consistent with the original while still possessing a certain degree of perturbation or adversarial properties, achieving the desired interference effect.

[0097] Therefore, in this embodiment, the generator receives environmental data (including mask images and random noise), extracts features through the encoder, and then discretizes and reconstructs them into multi-modal compatible adversarial samples through a vector quantizer. For example, in the infrared mode, the generator outputs a pattern containing an insulation layer, and balances the consistency of the surface temperature with the background environment through linear transformation. At the same time, the CBAM module fuses channels and spatial attention to ensure that the interference signal is effective in both visible light and infrared bands.

[0098] Specifically, before the step S021, in which the generator generates an initial confrontation pattern according to the environmental data, the method further includes:

[0099] Step S0211, extracting recognition model information of the adversarial end from the environmental data;

[0100] Step S0212: performing model upgrade prediction on a simulation platform based on the identified model information to obtain a prediction result;

[0101] Step S0213: Based on the prediction result, reinforcement learning training is performed on the generator to obtain a trained generator.

[0102] A simulation platform is deployed to simulate the enemy AI upgrade path (such as introducing liveness detection or gait analysis). The generator, as an intelligent agent, predicts the enemy strategy through reinforcement learning and optimizes it in advance. In this step, it is first necessary to collect environmental data related to the adversarial system. This data may include sensor input, real-time feedback, network status, or other contextual information related to the adversarial task. Environmental data can include various forms of input, such as images, sensor data, network traffic, attack information, etc.

[0103] The adversarial end refers to the part that performs the attack or interference. It can be a black box model or a model of the attacker's strategy in the system. The goal of this step is to extract recognition model information related to the adversarial end from the environmental data. This information may include attack patterns, adversarial strategies, vulnerability identification, system responses, and so on. Extracting recognition model information can use various algorithms, such as feature selection, pattern recognition, or deep learning techniques, to extract key information from the environmental data that helps predict the effectiveness of the attack.

[0104] After obtaining the recognition model information of the adversarial end, the next step is to predict model upgrades. "Upgrade" here generally refers to optimizing or improving the current system to cope with potential adversarial attacks. The model's behavior is further analyzed by simulating the attack environment through a simulation platform. The simulation platform is a virtual environment used to create and test adversarial models. Within this platform, various adversarial attack scenarios can be simulated, and the impact of adversarial attacks on the system can be predicted through simulation. These simulation results will help predict the model's possible responses under different adversarial scenarios, providing a basis for subsequent generator training. The simulation platform will perform a series of inferences and calculations based on the information from the recognition model to determine the model's behavior and response to different attack strategies, ultimately obtaining predictions. These predictions will provide guidance for model improvements, which may include how to adjust defense strategies, identify attack patterns, and optimize system responses.

[0105] Based on the prediction results, the generator is trained with reinforcement learning to obtain the background of the reinforcement learning training of the trained generator. Reinforcement learning is a training method that optimizes model strategies by interacting with the environment and continuously trial and error. The generator in this embodiment (usually a model that generates adversarial samples) will be trained through reinforcement learning based on previous prediction results. The purpose is to enable the generator to generate more effective adversarial samples and ultimately improve the effectiveness of the adversarial attack. During the reinforcement learning process, the generator will take actions (generate adversarial patterns or execute specific strategies) according to the prediction results given by the simulation platform, and adjust its own strategies based on the feedback of these actions. The core goal of the training is to enable the generator to perform more intelligently and efficiently in the face of adversarial attacks through continuous learning and adjustment.

[0106] During specific usage, when it is detected that an enemy drone cluster uses a heterogeneous recognition model, the system automatically switches to the confrontation mode, generates a comprehensive interference signal, and extracts the recognition model information of the confrontation end in the environmental data.

[0107] After training through the above scheme, the generator will obtain an optimized model that can generate more aggressive and disruptive adversarial samples based on the predicted adversarial scenarios. These trained generators will be able to effectively generate adversarial attacks in practical applications, achieving the purpose of interfering with or misleading the target system.

[0108] More specifically, after the step S02 of inputting the environmental data into a pre-built adversarial pattern generation model, and generating a pattern based on the environmental data by the adversarial pattern generation model to obtain the adversarial pattern, the method further includes:

[0109] Step S03: receiving device information of the physical confrontation device, and parsing the device information to obtain device parameters;

[0110] Step S04 , performing differentiable rendering on the physical confrontation device according to the device parameters and the confrontation pattern to obtain a rendering result.

[0111] First, the system needs to receive and parse the device information of the physical adversarial device. This device information includes the device's geometric shape, material properties, the working status of the temperature control module, the properties of the photochromic material, external lighting conditions, temperature distribution, etc. By analyzing the device information, the system can extract the necessary device parameters and build a mathematical model for differentiable rendering. Among them, the key device parameters that need to be extracted include optical parameters (light absorption rate, refractive index and color transformation model, etc.), thermal parameters (thermal conductivity and thermal radiation model, etc.) and electrical parameters (temperature control module power and module response speed, etc.).

[0112] Then, differentiable rendering can be performed based on the device parameters and the adversarial pattern. The core of differentiable rendering technology is to simulate and optimize the performance of the device through physical models, while taking into account the dynamic response of the device. The system uses the optical, thermal and other parameters of the device to render the photochromic material coated on the surface of the camouflage mask, and adjust the pattern and color according to environmental conditions.

[0113] The specific rendering process can be:

[0114] (1) Lighting rendering: Based on the ambient light and the absorption characteristics of the photochromic material on the camouflage mask surface, the system simulates the pattern changes under different lighting conditions. The system calculates the light absorption rate and reflectivity of each pixel and dynamically adjusts the pattern according to the lighting angle.

[0115] (2) Thermal radiation rendering: The thermal radiation signal of the mask is calculated by combining the thermal conductivity of the material and the response of the temperature control module. The micro temperature control module suppresses infrared radiation by adjusting the temperature of the mask, reducing the probability of being detected by the enemy's infrared sensor;

[0116] (3) Adversarial pattern generation: The system generates specific camouflage patterns by combining digital adversarial sample patterns with physical device parameters. The generation of these patterns needs to consider the adversarial effects with enemy sensors (such as infrared, visual, etc.).

[0117] In order to optimize the camouflage effect, the system needs to incorporate optical and thermal parameters as constraints into the loss function when generating adversarial samples. The goal of the loss function is to minimize the probability of being detected by enemy sensors (such as visible light and infrared detectors) while maintaining the realism and dynamic responsiveness of the camouflage pattern. The elements of the loss function include: visual visibility loss (i.e., measuring the visibility of the camouflage pattern under different lighting conditions, with the goal of reducing the probability of detection in enemy visible light sensors), infrared radiation loss (i.e., measuring the matching degree of the infrared radiation of the camouflage mask with the surrounding environment, with the goal of reducing the detection probability of infrared sensors), and thermal control efficiency loss (i.e., measuring the response speed and energy efficiency of the temperature control module in adjusting the temperature, with the goal of optimizing the power consumption and thermal radiation suppression effect of the temperature control module).

[0118] By optimizing the loss function, the system updates the adversarial samples and maps the digital adversarial samples to physical devices such as camouflage masks through a differentiable rendering process. Through backpropagation, the adversarial samples are gradually optimized until the optimal camouflage effect is achieved.

[0119] Ultimately, after multiple iterations of optimization, the adversarial samples generated by the system will be able to dynamically adjust the camouflage effect in the actual environment, allowing the camouflage mask to effectively conceal the target under different lighting and temperature conditions, minimizing the risk of being discovered by enemy sensors.

[0120] This embodiment, through the above-mentioned solution, specifically generates an initial adversarial pattern based on the environmental data through the generator; extracts features from the initial adversarial pattern through the attention module and encoder to obtain pattern features; and reconstructs the pattern based on the pattern features through a vector quantizer to obtain the adversarial pattern. Thus, real-time environmental data is collected by sensors, and then pattern generation is performed using a pre-built adversarial pattern generation model to obtain the adversarial pattern. The adversarial pattern generation model is obtained by distilling the initial adversarial pattern generation model constructed by the final generative adversarial network, the task pool, and the attention module. This solves the problem of the generated adversarial pattern failing to meet camouflage requirements in multimodal fusion recognition systems, thereby improving the efficiency of adversarial pattern generation.

[0121] For example, in order to help understand the implementation process of the method for generating a countermeasure pattern obtained by combining this embodiment with the above-mentioned embodiment 1, please refer to Figure 3 , Figure 3 A brief flowchart of a method for generating an adversarial pattern is provided, specifically:

[0122] Including the model construction plan,

[0123] First, several AI recognition models (convolutional neural network CNN, Transformer, multimodal fusion model) are integrated to obtain a task pool. For model selection and integration, for the convolutional neural network CNN part, a suitable architecture is selected according to specific task requirements, such as VGG, ResNet, etc., and its hyperparameters such as the number of layers, convolution kernel size, and step size are determined. Image features are extracted from public data sets or proprietary data sets. For the Transformer part, its parameters such as the number of heads, number of layers, and embedding dimension are determined. Its self-attention mechanism is used to process sequence data or perform feature extraction and conversion on image and other data. For the multimodal fusion model, a fusion strategy should be designed, such as early fusion (fusion of different modal data at the input layer), mid-term fusion (fusion at the feature extraction layer), and late fusion (fusion at the decision layer). Parameters such as the corresponding fusion weights are determined. These different models are integrated together to form a task pool so that it can process diverse input data and provide rich feature representations.

[0124] Subsequently, the data is normalized, standardized, and other preprocessing operations are performed according to the model requirements in the task pool, and the preprocessed training data is input into the task pool in batches. Each batch of data is processed by various models in the task pool to obtain the corresponding task instance, and the relevant features and intermediate results of each task instance are recorded.

[0125] Then, select a suitable meta-learning algorithm, such as MAML (model-independent meta-learning) or Reptile, determine its hyperparameters such as learning rate and number of iterations, design the meta-learning training process, including how to sample small batches of tasks from task instances for meta-training, how to evaluate model performance in the meta-testing phase, etc., and optimize the initial generator and initial discriminator in the initial generative adversarial network. For the initial generator, adjust its network parameters (such as the weights of the fully connected layer, the parameters of the convolutional layer, etc.) according to the gradient information feedback from the meta-learning algorithm to generate more realistic samples. You can improve the performance of the generator by increasing the depth of the generator, introducing new activation functions (such as LeakyReLU, etc.), using batch normalization, etc., and regularly evaluate the quality of the samples generated by the generator, such as by calculating the similarity index between the generated samples and the real samples (such as Fréchet Inception Distance, etc.) to monitor the optimization effect.

[0126] The optimization of the discriminator is also based on the gradient information of the meta-learning algorithm to optimize the parameters of the discriminator to improve its ability to distinguish true and false samples. Different discriminator architectures can be tried, such as multi-layer perceptrons, convolutional neural networks, etc., and their corresponding hyperparameters (such as the number of layers, the number of neurons, etc.) can be adjusted. In this embodiment, the discrimination accuracy of the discriminator is measured by indicators such as cross entropy loss, and optimization is performed accordingly.

[0127] The optimized generator and discriminator are combined into the final generative adversarial network to ensure the normal operation of the interaction and training mechanism between the two. The performance of the final generative adversarial network is evaluated on the test dataset, including indicators such as the quality of generated samples, the discrimination accuracy of the discriminator, and the convergence speed of the network.

[0128] The initial adversarial pattern generation model is constructed by combining the final generative adversarial network, task pool and attention module. In addition, it is necessary to design appropriate attention modules, such as spatial attention module, channel attention module, etc., determine their parameters and structure, fuse the attention module with the final generative adversarial network and task pool, clarify the location and method of fusion (for example, inserting attention modules in certain layers of the generator or discriminator), and optimize the entire initial adversarial pattern generation model through joint training and other methods, so that it can use the attention mechanism to better focus on important features and improve the quality and accuracy of pattern generation.

[0129] Finally, in order to facilitate practical use, the initial adversarial pattern generation model is distilled to obtain the adversarial pattern generation model for actual use.

[0130] It also includes the process of using the model:

[0131] First, the sensor is responsible for collecting two types of information, namely enemy device information and actual environment information, and sending this information to the adversarial pattern generation model. Based on the enemy device information, the generator in the adversarial pattern generation model is optimized so that it can better adapt to relevant conditions.

[0132] The optimized generator is then used to generate the initial adversarial pattern in combination with the actual environment information, and then the encoder and attention module are used to extract the features of the initial adversarial pattern.

[0133] Finally, based on the extracted graph features, a vector quantizer is used to reconstruct the graph and finally obtain a complete adversarial pattern.

[0134] It should be noted that the above examples are only used to understand the present application and do not constitute a limitation on the anti-pattern generation method of the present application. More simple transformations based on this technical concept are all within the scope of protection of the present application.

[0135] This application also provides a countermeasure pattern generation device, please refer to Figure 4 , the countermeasure pattern generating device includes:

[0136] The acquisition module 10 is used to collect environmental data through sensors;

[0137] The generation module 20 is used to input the environmental data into a pre-constructed adversarial pattern generation model, and generate a pattern based on the environmental data through the adversarial pattern generation model to obtain an adversarial pattern. The adversarial pattern generation model is obtained by distilling the initial adversarial pattern generation model, and the initial adversarial pattern generation model is constructed based on the final generative adversarial network, the task pool and the attention module.

[0138] The adversarial pattern generation device provided in this application, employing the adversarial pattern generation method described in the aforementioned embodiments, can address the technical issue of adversarial patterns failing to meet camouflage requirements in multimodal fusion recognition systems. Compared to the prior art, the adversarial pattern generation device provided in this application offers the same beneficial effects as the adversarial pattern generation method described in the aforementioned embodiments. Other technical features of the adversarial pattern generation device are the same as those disclosed in the aforementioned embodiments and are not further elaborated here.

[0139] The present application provides a countermeasure pattern generation device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the countermeasure pattern generation method in the above-mentioned embodiment 1.

[0140] Reference below Figure 5, which shows a schematic diagram of the structure of a countermeasure pattern generation device suitable for implementing embodiments of the present application. The countermeasure pattern generation device in the embodiments of the present application may include, but is not limited to, mobile terminals such as mobile phones, laptop computers, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Descriptions), PMPs (Portable Media Players), in-vehicle terminals (e.g., in-vehicle navigation terminals), and fixed terminals such as digital TVs and desktop computers. Figure 5 The countermeasure pattern generating device shown is merely an example and should not limit the functions and scope of use of the embodiments of the present application.

[0141] like Figure 5 As shown, the countermeasure pattern generation device may include a processing device 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes based on programs stored in a read-only memory 1002 or programs loaded from a storage device 1003 into a random access memory 1004. The random access memory 1004 also stores various programs and data required for the operation of the countermeasure pattern generation device. The processing device 1001, the read-only memory 1002, and the random access memory 1004 are interconnected via a bus 1005. An input / output interface 1006 is also connected to the bus. Typically, the following systems may be connected to the input / output interface 1006: an input device 1007 including, for example, a touch screen, touchpad, keyboard, mouse, image sensor, microphone, accelerometer, gyroscope, etc.; an output device 1008 including, for example, a liquid crystal display (LCD), speaker, vibrator, etc.; a storage device 1003 including, for example, a magnetic tape, hard disk, etc.; and a communication device 1009. Communication device 1009 can allow the countermeasure pattern generation device to communicate with other devices wirelessly or wired to exchange data. Although the figure shows a countermeasure pattern generation device with various systems, it should be understood that it is not required to implement or have all of the systems shown. More or fewer systems may be implemented or have alternatively.

[0142] In particular, according to the embodiments disclosed in the present application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, the embodiments disclosed in the present application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program comprising program code for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from a network via a communication device, or installed from a storage device 1003, or installed from a read-only memory 1002. When the computer program is executed by the processing device 1001, the above-mentioned functions defined in the method of the embodiment disclosed in the present application are performed.

[0143] The adversarial pattern generation device provided in this application, employing the adversarial pattern generation method described in the aforementioned embodiment, can resolve the technical issue of adversarial patterns failing to meet camouflage requirements in multimodal fusion recognition systems. Compared to the prior art, the adversarial pattern generation device provided in this application achieves the same beneficial effects as the adversarial pattern generation method described in the aforementioned embodiment. Other technical features of this adversarial pattern generation device are the same as those disclosed in the aforementioned embodiment and are not further elaborated here.

[0144] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any one or more embodiments or examples in a suitable manner.

[0145] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

[0146] The present application provides a computer-readable storage medium having computer-readable program instructions (ie, a computer program) stored thereon, wherein the computer-readable program instructions are used to execute the method for generating a countermeasure pattern in the above-mentioned embodiment.

[0147] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems or devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system or device. The program code contained on the computer-readable storage medium may be transmitted using any appropriate medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0148] The computer-readable storage medium may be included in the countermeasure pattern generation device, or may exist independently without being incorporated into the countermeasure pattern generation device.

[0149] The above-mentioned computer-readable storage medium carries one or more programs. When the above-mentioned one or more programs are executed by the adversarial pattern generation device, the adversarial pattern generation device: collects environmental data through sensors; inputs the environmental data into a pre-constructed adversarial pattern generation model, and generates a pattern based on the environmental data through the adversarial pattern generation model to obtain an adversarial pattern. The adversarial pattern generation model is obtained by distilling an initial adversarial pattern generation model, and the initial adversarial pattern generation model is constructed based on the final generative adversarial network, the task pool and the attention module.

[0150] Computer program code for performing the operations of the present application may be written in one or more programming languages, or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, C++, and conventional procedural programming languages ​​such as "C" or similar programming languages. The program code may be executed entirely on the user's computer, partially on the user's computer, as a stand-alone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0151] The flow charts and block diagrams in the accompanying drawings illustrate the possible architecture, functions and operations of the systems, methods and computer program products according to various embodiments of the present application. In this regard, each box in the flow chart or block diagram can represent a module, program segment or a part of code, and the module, program segment or a part of code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order than that marked in the accompanying drawings. For example, two boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or flow chart, and the combination of the boxes in the block diagram and / or flow chart can be implemented by a dedicated hardware-based system that performs the specified function or operation, or can be implemented by a combination of dedicated hardware and computer instructions.

[0152] The modules described in the embodiments of the present application may be implemented in software or hardware, wherein the name of a module does not necessarily limit the unit itself.

[0153] The computer-readable storage medium provided in this application stores computer-readable program instructions (i.e., a computer program) for executing the aforementioned adversarial pattern generation method. This computer-readable storage medium can address the technical issue of adversarial patterns failing to meet camouflage requirements in multimodal fusion recognition systems. Compared to the prior art, the beneficial effects of the computer-readable storage medium provided in this application are similar to those of the adversarial pattern generation method provided in the aforementioned embodiments and are not further elaborated here.

[0154] The present application also provides a computer program product, including a computer program, which implements the steps of the above-mentioned countermeasure pattern generation method when executed by a processor.

[0155] The computer program product provided in this application can address the technical issue of generating adversarial patterns that fail to meet camouflage requirements for multimodal fusion recognition systems. Compared to the prior art, the beneficial effects of the computer program product provided in this application are the same as those of the adversarial pattern generation method provided in the aforementioned embodiments, and are not further elaborated here.

[0156] The above description is only part of the embodiments of the present application and does not limit the patent scope of the present application. All equivalent structural transformations made by using the contents of the present application specification and drawings under the technical concept of the present application, or direct / indirect application in other related technical fields are included in the patent protection scope of the present application.

Claims

1. A method for generating a countermeasure pattern, characterized in that: The method for generating a countermeasure pattern comprises: Collect environmental data through sensors; The environmental data is input into a pre-constructed adversarial pattern generation model, and the adversarial pattern generation model is used to generate a pattern based on the environmental data to obtain an adversarial pattern. The adversarial pattern generation model is obtained by distilling an initial adversarial pattern generation model, and the initial adversarial pattern generation model is constructed based on the final generative adversarial network, the task pool, and the attention module.

2. The method for generating a countermeasure pattern according to claim 1, wherein: Before the step of inputting the environmental data into a pre-built adversarial pattern generation model and generating a pattern based on the environmental data by the adversarial pattern generation model to obtain a adversarial pattern, the method further includes: Integrate several AI recognition models to obtain a task pool, and input training data into the task pool to obtain a task instance; Based on the task instance, the initialization parameters of the initial generator and the initial discriminator of the initial generative adversarial network are optimized by a meta-learning algorithm to obtain a generator and a discriminator; Update the initial generative adversarial network according to the generator and the discriminator to obtain a final generative adversarial network; An initial adversarial pattern generation model is constructed based on the final generative adversarial network, the task pool, and the attention module; Model distillation is performed on the initial adversarial pattern generation model to obtain an adversarial pattern generation model.

3. The method for generating a countermeasure pattern according to claim 2, wherein: The step of optimizing the initialization parameters of the initial generator and the initial discriminator of the initial generative adversarial network by a meta-learning algorithm based on the task instance to obtain the generator and the discriminator includes: Inputting the task instance into the initial generator to obtain an adversarial sample; Classify the adversarial sample by the initial discriminator to obtain a classification result and an evaluation result; Calculate the adversarial sample using a diversity loss function to obtain a loss value; Based on the classification result, the evaluation result and the loss value, gradient updating is performed on the initialization parameters of the initial generator through a meta-learning algorithm to obtain a generator; Based on the classification result and the evaluation result, the initialization parameters of the initial discriminator are gradient updated through a meta-learning algorithm to obtain a discriminator.

4. The method for generating a countermeasure pattern according to claim 2, wherein: The step of inputting the environmental data into a pre-built adversarial pattern generation model, and generating a pattern based on the environmental data by the adversarial pattern generation model to obtain the adversarial pattern comprises: generating an initial adversarial pattern by the generator according to the environmental data; Extracting features from the initial adversarial pattern using the attention module and the encoder to obtain pattern features; According to the pattern features, a vector quantizer is used to reconstruct the pattern to obtain an adversarial pattern.

5. The method for generating a countermeasure pattern according to claim 4, wherein: Before the step of generating an initial confrontation pattern by the generator according to the environmental data, the method further includes: Extracting recognition model information of the adversarial end from the environmental data; According to the identification model information, a model upgrade prediction is performed through a simulation platform to obtain a prediction result; Based on the prediction results, reinforcement learning training is performed on the generator to obtain a trained generator.

6. The method for generating a countermeasure pattern according to claim 1, wherein: After the step of inputting the environmental data into a pre-built adversarial pattern generation model, and generating a pattern based on the environmental data by the adversarial pattern generation model to obtain a adversarial pattern, the method further comprises: Receiving device information of the physical confrontation device, and parsing the device information to obtain device parameters; Differentiable rendering is performed on the physical confrontation device according to the device parameters and the confrontation pattern to obtain a rendering result.

7. A countermeasure pattern generating device, characterized in that: The countermeasure pattern generating device comprises: A collection module, used to collect environmental data through sensors; A generation module is used to input the environmental data into a pre-built adversarial pattern generation model, and generate a pattern based on the environmental data through the adversarial pattern generation model to obtain an adversarial pattern. The adversarial pattern generation model is obtained by distilling an initial adversarial pattern generation model, and the initial adversarial pattern generation model is constructed based on the final generative adversarial network, the task pool and the attention module.

8. A countermeasure pattern generating device, characterized in that: The device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program is configured to implement the steps of the method for generating a countermeasure pattern according to any one of claims 1 to 6.

9. A storage medium, characterized in that: The storage medium is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by a processor, the steps of the method for generating a countermeasure pattern according to any one of claims 1 to 6 are implemented.

10. A computer program product, characterized in that The computer program product comprises a computer program, and when the computer program is executed by a processor, the steps of the method for generating a countermeasure pattern according to any one of claims 1 to 6 are implemented.

Images