3D point cloud defense method and device based on CLIP guidance, equipment and medium

Abstract

The invention provides a 3D point cloud defense method, device and equipment based on CLIP guidance, and a medium, relates to the technical field of point cloud defense, and aims to enhance the robustness and security of 3D point cloud data in the face of adversarial attack by using a contrast language-image pre-training model. According to the method, a new defense framework is provided by combining geometric and semantic features, triggers in backdoor attacks can be effectively identified and inhibited, meanwhile, the geometric structure of point cloud data is repaired, and the classification performance is improved. Specifically, the method includes projecting a 3D point cloud into a multi-view 2D depth map, extracting cross-modal features using a CLIP model, and performing semantically guided 3D reconstruction through a variational auto-encoder (VAE) to generate a repaired point cloud that retains key information and neutralizes perturbations. The method shows excellent performance in various complex scenes, and provides powerful support for the safety of the 3D vision technology in key application.

Description

Technical Field

[0001] The present invention relates to the technical field of point cloud defense, and particularly to a 3D point cloud defense method, device, equipment and medium guided by CLIP. Background Art

[0002] In today's digital age, three-dimensional (3D) vision technology is rapidly developing and widely used in many cutting-edge fields, such as autonomous vehicles, intelligent robots, and medical imaging. These technologies rely on accurate 3D point cloud data to achieve environmental perception, object recognition, and decision-making. However, with the continuous progress of artificial intelligence technology, the security and robustness issues of 3D deep learning models have gradually emerged, especially adversarial attacks pose a serious threat to these models.

[0003] Adversarial attack is a malicious behavior where an attacker makes a carefully designed small perturbation to the input data, causing the deep learning model to produce incorrect outputs. In the field of 3D point clouds, this attack method is particularly covert and dangerous. The attacker can inject tiny noises or triggers into the point cloud data. These perturbations are almost imperceptible to human observers but are sufficient to mislead a well-trained 3D deep neural network, resulting in serious errors in tasks such as classification and recognition. For example, in an autonomous driving scenario, the attacked model may misidentify an obstacle as a passable area, thus triggering potential traffic accidents.

[0004] To address this challenge, some defense strategies have been proposed in the market. Among them, simple random sampling (SRS) is a basic defense method that reduces the impact of noise by randomly selecting some points in the point cloud. However, this method may inadvertently discard some key information, thereby reducing the overall quality and integrity of the point cloud data. Statistical outlier removal (SOR) attempts to identify and remove outliers through statistical methods to improve data quality. Although SOR can effectively identify points that do not conform to the normal distribution in some cases, it may also erroneously delete some important structural information, especially when the difference between noise and normal points is not obvious.

[0005] Another more advanced defense method is DUP-Net (Denoiser and Upsampler Network for 3D Adversarial Point Clouds Defense), which is a deep learning-based network designed to enhance the robustness of point cloud data against adversarial attacks by learning the global and local features of the point cloud data. DUP-Net utilizes graph convolutional networks and feature fusion mechanisms to automatically learn useful features in the point cloud and identify and repair abnormal points caused by adversarial perturbations through convolutional operations. Although DUP-Net has shown certain advantages in point cloud repair, its performance largely depends on the pre-trained network model and has poor adaptability to training data and specific perturbation types. In the face of complex and changing real-world scenarios, the performance of DUP-Net often drops significantly.

[0006] Although existing defense methods can, to some extent, mitigate the impact of adversarial attacks on 3D point cloud data, they all have some limitations. These methods may either lead to the loss of effective information, or it is difficult to accurately distinguish normal points from adversarial perturbation points, or they lack sufficient generality and adaptability when dealing with complex scenarios. In addition, they usually cannot effectively cope with the hidden triggers in backdoor attacks. Therefore, there is an urgent need for a new, more efficient and robust defense technology to ensure the security and reliability of 3D vision technology in various application scenarios.

[0007] In view of this, the present application is proposed. Summary of the Invention

[0008] The present invention provides a CLIP-guided 3D point cloud defense method, device, equipment and medium, which can at least partially improve the above problems.

[0009] To achieve the above object, the present invention adopts the following technical solutions: A CLIP-guided 3D point cloud defense method, which includes: Obtain adversarial attack point cloud data, project the adversarial attack point cloud data into a preset 3D grid to obtain a 2D depth map; Compress the 2D depth map, and input the compressed 2D depth map into a preset CLIP model to generate image features and text features, and calculate the final classification Logits value based on the image features and text features; According to the classification Logits value, use a preset VAE model for reconstruction processing to generate a reconstruction result, and calculate a reconstruction error based on the reconstruction result, and perform iterative processing to generate clean reconstructed point cloud data.

[0010] The present invention also provides a CLIP-guided 3D point cloud defense device, which includes: A projection unit, configured to obtain adversarial attack point cloud data, project the adversarial attack point cloud data into a preset 3D grid to obtain a 2D depth map; A CLIP unit, configured to compress the 2D depth map, input the compressed 2D depth map into a preset CLIP model to generate image features and text features, and calculate a final classification Logits value based on the image features and text features; A reconstruction unit, configured to perform reconstruction processing using a preset VAE model according to the classification Logits value, generate a reconstruction result, calculate a reconstruction error based on the reconstruction result, perform iterative processing, and generate clean reconstructed point cloud data.

[0011] The present invention also provides a CLIP-guided 3D point cloud defense device, which includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements the CLIP-guided 3D point cloud defense method described in any one of the above.

[0012] The present invention also provides a readable storage medium, which stores a computer program that can be executed by a processor of a device where the storage medium is located to implement the CLIP-guided 3D point cloud defense method described in any one of the above.

[0013] In summary, the CLIP-guided 3D point cloud defense method combines geometric and semantic features, utilizes the Contrastive Language-Image Pre-Training (CLIP) model and the Variational Autoencoder (VAE) to effectively defend 3D point cloud data. Specifically, this method first projects 3D point cloud data into multi-view 2D depth maps, and then uses the CLIP model to extract cross-modal visual and semantic features to identify and suppress potential backdoor attack triggers. Immediately afterwards, by fusing the geometric features and semantic features of the point cloud, semantic-guided 3D reconstruction is performed using VAE to generate a repaired point cloud that not only retains key information but also neutralizes perturbations. This process not only improves the robustness and security of point cloud data, but also significantly improves the classification performance. This method has been comprehensively experimentally verified on various backdoor attack methods and datasets, and the results show that it has excellent robustness, generality, and efficiency in complex scenarios, providing strong guarantee for the security of 3D vision technology in key applications such as autonomous driving and robotics. The aim is to solve the limitations of existing defense methods in dealing with adversarial attacks. Description of the Drawings

[0014] Figure 1 It is a schematic flowchart of the CLIP-guided 3D point cloud defense method provided by the first embodiment of the present invention; Figure 2 It is an example diagram of the defense object being a backdoor attack provided by the embodiment of the present invention; Figure 3 It is a schematic framework diagram of the CLIP-guided 3D point cloud defense method provided by the embodiment of the present invention; Figure 4 It is a schematic diagram of the defense effect provided by the embodiment of the present invention; Figure 5 It is a schematic module diagram of the CLIP-guided 3D point cloud defense device provided by the second embodiment of the present invention. Detailed implementation manners

[0015] In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0016] Refer to Figures 1 to 3 As shown, the first embodiment of the present invention discloses a CLIP-guided 3D point cloud defense method, which can be executed by a CLIP-guided 3D point cloud defense device (hereinafter referred to as the defense device), and particularly, by one or more processors in the defense device to implement the following method: S1. Obtain adversarial attack point cloud data, project the adversarial attack point cloud data onto a preset 3D grid to obtain a 2D depth map; Specifically, step S1 includes: obtaining adversarial attack point cloud data P = (x, y, z), performing normalization processing on the adversarial attack point cloud data P = (x, y, z) to obtain normalized point cloud data p = , , where is the minimum value of the coordinate in the point cloud data, is the maximum value of the coordinate in the point cloud data, is the minimum value of the coordinate in the point cloud data, is the maximum value of the coordinate in the point cloud data, is the minimum value of the coordinate in the point cloud data, is the maximum value of the coordinate in the point cloud data; Construct a 3D grid initialized to zero , where H is the height of the grid, W is the width of the grid, and D is the depth of the grid; Project the normalized point cloud data onto the corresponding voxels in the 3D grid. For multiple point cloud data projected onto the same voxel, set the depth value of the voxel to the minimum depth value of all the projected points. The expression is: , is the voxel position is the depth value, and z is the minimum depth value of the projected points; Adopt a local minimum filling operation to set the depth value of the voxels in the 3D grid to the above minimum depth value, generating a 2D depth map. The formula is: , is the voxel position The local window around.

[0017] In this embodiment, first, obtain adversarial attack point cloud data. These data may have been subject to adversarial attacks and contain tiny perturbations or triggers. These perturbations are almost imperceptible to human observers but sufficient to mislead deep learning models. To process these data, it is necessary to first normalize the point cloud data to obtain normalized point cloud data. The purpose of normalization is to unify the coordinate range of the point cloud data into a standard interval, such as [0, 1], for subsequent processing. Simply put, for a given perspective, normalize the 3D point cloud coordinates to the range 0.

[0018] Immediately afterwards, construct a 3D grid initialized to zero. This 3D grid is used to project the normalized point cloud data into a regular grid structure for subsequent processing. Due to the sparsity of the point cloud in the grid, in order to effectively fill these empty voxels and ensure visual continuity, during the projection process, adopt a local minimum filling operation to set the depth value in the voxel grid to the minimum depth value within this local window. Through the local filling operation, the empty voxels in the grid are effectively filled, while the occlusion relationship in the depth map is retained, ensuring that unoccluded points are projected to the correct grid positions.

[0019] Through the above steps, 2D depth maps from multiple perspectives are generated. These 2D depth maps retain the geometric information of the point cloud data. At the same time, through the local minimum filling operation, the integrity and continuity of the data are ensured. These 2D depth maps will be used as the input for subsequent processing to extract cross-modal semantic features, identify and suppress potential backdoor triggers, and ultimately improve the repair and classification performance of the point cloud data. This process not only enhances the robustness of the point cloud data but also improves the classification accuracy, significantly enhancing the defense ability against adversarial attacks.

[0020] S2. Compress the 2D depth map, and input the compressed 2D depth map into a preset CLIP model to generate image features and text features, and calculate the final classification Logits value based on the image features and text features; Specifically, step S2 includes: compressing the 2D depth map and extracting the minimum depth value of each depth channel in the 2D depth map to obtain the compressed multi-view V; Input the multi-view V into a preset CLIP model to obtain image features where C is a specific dimension and M is the number of depth maps under multi-view projection; Obtain the category label Y corresponding to the anti-attack point cloud data, and input the category label Y into a preset CLIP model to obtain text features where N is the number of iterations.

[0021] According to the image features and text features, perform feature alignment processing, calculate the correlation between the image features and text features, and the calculation formula is: where is the attention weight; Based on the image features, text features and attention weight, calculate the final classification Logits value, and the calculation formula is: where

[0022] is the text feature.

[0023] In this embodiment, the 2D depth map is compressed, and the minimum depth value in each depth channel is extracted to reduce the dimension of the data and retain key information; this compression operation not only reduces the complexity of subsequent processing, but also improves the processing efficiency, and applying this compression operation repeatedly to the RGB channels can effectively reduce the complexity of subsequent processing.

[0024] ​​​Subsequently, since features from different perspectives may be affected by a trigger (e.g., a small sphere), the CLIP-guided 3D point cloud defense method designs an adaptive weighting mechanism to perform feature alignment between multi-perspective semantic representations and text features according to modal correlations. Specifically, based on the image features and text features, feature alignment processing is carried out to calculate the correlation between the two. This process is achieved through an attention mechanism, ensuring the effective fusion of image features and text features. Finally, based on the image features, text features, and attention weights, the final classification Logits value is calculated. This process not only considers the fusion of image features and text features but also enhances the importance of key features through the attention mechanism.

[0025] S3. According to the classification Logits value, use a preset VAE model for reconstruction processing to generate a reconstruction result, and calculate the reconstruction error based on the reconstruction result, and perform iterative processing to generate clean reconstructed point cloud data.

[0026] Specifically, step S3 includes: guiding a preset VAE model for reconstruction based on the classification Logits value, and the objective function formula of the VAE model is: , , where is the objective function of the VAE model, is the expectation operator, is the given latent variable and under the condition, generate the logarithmic probability of, is the KL divergence, is the given data , the latent variable conditions respectively, is the given category , the latent variable conditions respectively, is the gradient of the decoder for ; Converge to obtain a reconstruction result consistent with the target true sample: , is the logarithmic conditional probability of the target true sample.

[0027] Input the reconstruction result into the preset VAE model, calculate the KL divergence value as the reconstruction error, and the calculation formula is: , , where is the calculation result of the KL divergence, is the regularization constraint, is the mean of the latent variable distribution of the original point cloud data, is the standard deviation of the latent variable distribution of the original point cloud data, is the mean of the latent variable distribution of the reconstructed point cloud data, is the standard deviation of the latent variable distribution of the reconstructed point cloud data; Calculate the similarity between the real point cloud data and the predicted point cloud data, and the formula is: , where Q is the predicted point cloud data set, P is the real point cloud data set, q is a data in the predicted point cloud, and p is a data in the real point cloud; Perform iterative training according to the reconstruction error and similarity to minimize the target loss function , and generate clean reconstructed point cloud data based on the trained target loss function.

[0028] In this embodiment, first, VAE is used to integrate cross-modal semantics into the latent space to facilitate reconstruction, that is, based on the classification Logits value, a preset VAE model is guided to perform reconstruction. The goal of the VAE model is to maximize the conditional likelihood of generating real data while minimizing the KL divergence (i.e., reconstruction error) between the model-predicted point cloud and the original point cloud. By using the guidance of semantic information, the point cloud containing the trigger can be repaired or the influence of the trigger can be eliminated during the reconstruction process, and finally converge to a result consistent with the target real sample.

[0029] Among them, VAE is a latent feature reconstruction framework. Due to the unpredictability of the trigger type, it relies heavily on the prior distribution of real data during use. In order to further strengthen the reconstruction process of the CLIP-guided 3D point cloud defense method, the reconstructed point cloud is re-input into the VAE model to calculate the KL divergence, and by constraining the latent space distribution, it is made to align more closely with the original distribution.

[0030] Please refer to Figure 4, In summary, the CLIP-guided 3D point cloud defense method effectively neutralizes hidden triggers and improves classification performance by innovatively introducing cross-modal semantic features and combining geometric and semantic features. It aims to solve problems in existing 3D point cloud defense methods, such as potential loss of effective information during denoising, difficulty in accurately distinguishing normal points from adversarial perturbation points, lack of generality and adaptability to complex scenes, and ineffectiveness in dealing with backdoor attacks. Specifically, the CLIP-guided 3D point cloud defense method can effectively identify and suppress triggers in backdoor attacks through cross-modal semantic features extracted by the CLIP model, ensuring the structural integrity and semantic consistency of the point cloud. Secondly, the VAE model is used for reconstruction processing, which not only repairs the attacked point cloud data but also significantly improves the data integrity and classification performance. In addition, this method has been comprehensively experimentally verified on various backdoor attack methods and datasets, demonstrating its robustness, generality, and efficiency in complex scenes, providing strong guarantees for the security of 3D vision technology in critical applications.

[0031] Briefly, the CLIP-guided 3D point cloud defense method effectively improves the robustness and classification performance of 3D point cloud data under adversarial attacks through innovative feature extraction and reconstruction methods, providing strong support for the security of 3D vision technology in critical applications such as autonomous driving and robotics, and promoting the continuous development and popularization of 3D vision technology in practical applications.

[0032] Please refer to Figure 5 , The second embodiment of the present invention provides a CLIP-guided 3D point cloud defense device, which includes: A projection unit 201, configured to obtain adversarial attack point cloud data, project the adversarial attack point cloud data into a preset 3D grid to obtain a 2D depth map; A CLIP unit 202, configured to compress the 2D depth map, input the compressed 2D depth map into a preset CLIP model to generate image features and text features, and calculate the final classification Logits value based on the image features and text features; A reconstruction unit 203, configured to perform reconstruction processing using a preset VAE model according to the classification Logits value, generate a reconstruction result, calculate a reconstruction error based on the reconstruction result, perform iterative processing, and generate clean reconstructed point cloud data.

[0033] The third embodiment of the present invention provides a CLIP-guided 3D point cloud defense device, which includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements the CLIP-guided 3D point cloud defense method as described in any one of the above.

[0034] The fourth embodiment of the present invention provides a readable storage medium storing a computer program, which can be executed by a processor of the device where the storage medium is located to implement the CLIP-guided 3D point cloud defense method described in any one of the above.

[0035] Exemplarily, each of the above devices and each process step can be implemented by a computer program, which can be divided into one or more units. The one or more units are stored in the memory and executed by the processor to complete the present invention.

[0036] The so-called processor may be a central processing unit (CPU), or may also be other general-purpose processors, digital signal processors (DSPs), application specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0037] The memory can be used to store the computer program and / or module. The processor realizes various functions of the present invention by running or executing the computer program and / or module stored in the memory and calling the data stored in the memory. The memory mainly includes a program storage area and a data storage area. Among them, the program storage area can store an operating system, application programs required for at least one function (such as a sound playback function, an image playback function, etc.); the data storage area can store data created according to the use of the mobile phone (such as audio data, phone book, etc.). In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, at least one magnetic disk storage device, a flash device, or other volatile solid-state storage devices.

[0038] Among them, if the units integrated in the electronic device or the printer are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on such an understanding, to implement all or part of the processes in the above-described embodiment methods of the present invention, it can also be completed by a computer program instructing relevant hardware. The computer program can be stored in a computer-readable storage medium. When the computer program is executed by a processor, the steps of the above-described various method embodiments can be implemented. Among them, the computer program includes computer program code, and the computer program code can be in the form of source code, object code, executable file, or some intermediate form, etc. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signal, telecommunication signal, and software distribution medium, etc. It should be noted that the content included in the computer-readable medium can be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, the computer-readable medium does not include electrical carrier signals and telecommunication signals.

[0039] It should be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated. The components shown as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. In addition, in the attached drawings of the device embodiments provided by the present invention, the connection relationship between the modules indicates that they have a communication connection, which can be specifically implemented as one or more communication buses or signal lines. Those of ordinary skill in the art can understand and implement it without creative efforts.

[0040] The above is the preferred implementation manner of the present invention. It should be pointed out that for those of ordinary skill in the art in the technical field, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements are also regarded as the protection scope of the present invention.

Claims

1. A 3D point cloud defense method based on CLIP guidance, characterized in that: include: Acquire adversarial attack point cloud data, project the adversarial attack point cloud data into a preset 3D grid to obtain a 2D depth map, and preprocess the 2D depth map; Compressing the preprocessed 2D depth map, and inputting the compressed 2D depth map into a preset CLIP model, generating image features and text features, and calculating the final classification Logits value based on the image features and text features; According to the classified Logits value, a preset VAE model is used to perform reconstruction processing to generate a reconstruction result, and a reconstruction error is calculated based on the reconstruction result, and an iterative process is performed to generate clean reconstructed point cloud data.

2. The CLIP-guided 3D point cloud defense method according to claim 1, characterized in that: Acquire the adversarial attack point cloud data, project the adversarial attack point cloud data into a preset 3D grid, and obtain a 2D depth map, specifically: Obtain the adversarial attack point cloud data P=(x, y, z), and perform normalization processing on the adversarial attack point cloud data P=(x, y, z) to obtain normalized point cloud data p= , ,in, For point cloud data The minimum value of the coordinates, For point cloud data The maximum value of the coordinates, For point cloud data The minimum value of the coordinates, For point cloud data The maximum value of the coordinates, For point cloud data The minimum value of the coordinates, For point cloud data The maximum value of the coordinates; Construct a zero-initialized 3D grid , where H is the height of the grid, W is the width of the grid, and D is the depth of the grid; The normalized point cloud data is projected onto the corresponding voxel in the 3D grid, wherein for multiple point cloud data projected onto the same voxel, the depth value of the voxel is set to the minimum depth value of all projected points, and the expression is: , is the voxel position The depth value of z is the minimum depth value of the point projection point; The local minimum filling operation is adopted to set the depth value of the voxel in the 3D grid to the above minimum depth value to generate a 2D depth map. The formula is: , is the voxel position Local windows around.

3. The CLIP-guided 3D point cloud defense method according to claim 2, characterized in that: The 2D depth map is preprocessed, specifically: The edge enhancement algorithm is used to extract and enhance the boundary information in the 2D depth map, and the bilateral filtering algorithm is used to smooth the distribution of depth values ​​to eliminate noise interference; The pixel values ​​of the 2D depth map are scaled to a preset range of 0 to 1 based on the minimum-maximum normalization technique, and the normalized 2D depth map is resized to adjust its resolution to the resolution required by the CLIP model.

4. The CLIP-guided 3D point cloud defense method according to claim 3, characterized in that: The 2D depth map is compressed and input into a preset CLIP model to generate image features and text features, specifically: Compress the 2D depth map and extract each depth channel in the 2D depth map The minimum depth value of , get the compressed multi-view V; The multi-view V is input into the preset CLIP model to obtain image features , C is the specific dimension, M is the number of depth maps under multi-view projection; Obtain the category label Y corresponding to the anti-attack point cloud data, and input the category label Y into the preset CLIP model to obtain the text feature , N is the number of iterations.

5. The CLIP-guided 3D point cloud defense method according to claim 4, characterized in that: The final classification Logits value is calculated based on the image features and text features, specifically: According to the image features and text features, feature alignment processing is performed to calculate the correlation between the image features and text features. The calculation formula is: ,in, is the attention weight; Based on the image features, text features and attention weights, the final classification Logits value is calculated using the following formula: , , is a text feature.

6. The CLIP-guided 3D point cloud defense method according to claim 5, characterized in that: According to the classified Logits value, the preset VAE model is used to perform reconstruction processing to generate a reconstruction result, which is specifically: Based on the classification Logits value, the preset VAE model is guided to perform reconstruction, and the objective function formula of the VAE model is: , ,in, is the objective function of the VAE model, is the expectation operator, For a given latent variable and Under these conditions, the generation The logarithmic probability of is the KL divergence, For the given data , latent variables The conditions are respectively, For a given category , latent variables The difference, For the decoder The gradient of Converge to get the reconstruction result consistent with the target real sample: , is the logarithmic conditional probability of the target true sample.

7. The CLIP-guided 3D point cloud defense method according to claim 6, characterized in that: The reconstruction error is calculated based on the reconstruction result, and iterative processing is performed to generate clean reconstructed point cloud data, specifically: The reconstruction result is input into the preset VAE model, and the KL divergence value is calculated as the reconstruction error. The calculation formula is: , ,in, is the calculation result of KL divergence, is the regularization constraint, is the mean of the latent variable distribution of the original point cloud data, is the standard deviation of the latent variable distribution of the original point cloud data, is the mean of the latent variable distribution of the reconstructed point cloud data, is the standard deviation of the latent variable distribution of the reconstructed point cloud data; Calculate the similarity between the real point cloud data and the predicted point cloud data. The formula is: , Q is the predicted point cloud data set, P is the real point cloud data set, q is a data in the predicted point cloud, and p is a data in the real point cloud; Iterative training is performed based on the reconstruction error and similarity to reduce the target loss function , based on the trained target loss function, clean reconstructed point cloud data is generated.

8. A 3D point cloud defense device based on CLIP guidance, characterized in that: include: A projection unit, used to obtain the adversarial attack point cloud data, project the adversarial attack point cloud data into a preset 3D grid to obtain a 2D depth map, and pre-process the 2D depth map; A CLIP unit, used to compress the preprocessed 2D depth map, input the compressed 2D depth map into a preset CLIP model, generate image features and text features, and calculate the final classification Logits value based on the image features and text features; The reconstruction unit is used to perform reconstruction processing using a preset VAE model according to the classified Logits value to generate a reconstruction result, and calculate the reconstruction error based on the reconstruction result, perform iterative processing, and generate clean reconstructed point cloud data.

9. A 3D point cloud defense device based on CLIP guidance, characterized in that: The invention comprises a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein when the processor executes the computer program, the CLIP-guided 3D point cloud defense method according to any one of claims 1 to 7 is implemented.

10. A readable storage medium, characterized in that: A computer program is stored, and the computer program can be executed by a processor of the device where the storage medium is located to implement the CLIP-guided 3D point cloud defense method as described in any one of claims 1 to 7.

Images