Point cloud encryption method and device based on Gaussian mixture flow model, medium and product

Abstract

The invention discloses a point cloud encryption method and device based on a Gaussian mixture flow model, a medium and a product, and relates to the technical field of data processing, and the method comprises the steps: mapping an original point cloud to a hidden space obeying Gaussian mixture distribution based on a preset point cloud protection model, and obtaining an encrypted point cloud, the point cloud protection model being constructed based on the Gaussian mixture flow model; and rotating each point of the encrypted point cloud through a preset target orthogonal matrix to obtain a target point cloud. According to the method, the Gaussian mixture flow model is introduced into 3D point cloud data processing, the difficulty that a traditional flow model cannot be directly applied to the point cloud due to the irregularity, sparsity and disorder of the point cloud data is overcome, the target point cloud can be restored to the original point cloud under the condition that a user holds a decoder and a target orthogonal matrix, and the user experience is improved. And the encrypted point cloud can be normally used in a downstream application scene, so that the practicability of the encrypted point cloud is improved.

Description

Technical Field

[0001] The present application relates to the technical field of data processing, and particularly to a point cloud encryption method, device, medium and product based on a Gaussian mixture flow model. Background Art

[0002] The development of three-dimensional point cloud technology benefits from the progress of sensor devices and augmented reality and mixed reality technologies, and is widely used in tasks such as classification, segmentation, localization, vehicle positioning, modeling, and three-dimensional vision language learning. However, due to the irregularity, sparsity, and disorder of point cloud data, it is more challenging to achieve privacy protection in the field of three-dimensional vision. Existing three-dimensional privacy protection research mainly focuses on camera positioning tasks. For devices such as HoloLens, MagicLeap1, and iRobotRoomba, the technology of elevating point clouds to line clouds is used to hide the appearance of objects. However, the original intention of such methods is limited to specific positioning tasks, and the enhanced processing will damage the geometric information of the point cloud, resulting in irreversible point cloud information, making it impossible for downstream tasks to use the point cloud information; another type of method simplifies the protection problem to an encryption problem and uses a chaotic system to obfuscate the position of the point cloud. However, the obfuscated point cloud does not retain its semantic information, and the encrypted data needs to be restored before being used in downstream tasks. The obfuscated point cloud leads to the complexity of point cloud restoration, thus increasing the complexity of the application. In summary, the encrypted point cloud in current three-dimensional point cloud privacy protection has the problem of low practicality.

[0003] The above content is only used to assist in understanding the technical solution of the present application, and does not represent an admission that the above content is prior art. Summary of the Invention

[0004] The main purpose of the present application is to provide a point cloud encryption method, device, medium and product based on a Gaussian mixture flow model, aiming to solve the technical problem of low practicality of the encrypted point cloud in current three-dimensional point cloud privacy protection.

[0005] To achieve the above purpose, the present application proposes a point cloud encryption method, and the method includes:

[0006] Mapping the original point cloud to a latent space subject to a mixture of Gaussian distributions based on a preset point cloud protection model to obtain an encrypted point cloud, where the point cloud protection model includes a shape representation component constructed based on a Gaussian mixture flow model and a point transformation component constructed based on a Gaussian mixture flow model. In the process of mapping the original point cloud through the point cloud protection model, the original point cloud is mapped to a latent vector through the shape representation component, and the original point cloud is mapped to an encrypted point cloud based on the latent vector through the point transformation component;

[0007] Rotating each point of the encrypted point cloud through a preset target orthogonal matrix to obtain a target point cloud.

[0008] In one embodiment, the step of mapping the original point cloud to a latent space subject to a mixture of Gaussian distributions based on a preset point cloud protection model to obtain an encrypted point cloud includes:

[0009] Input the original point cloud into the shape representation component, and map the original point cloud to a latent space based on the shape representation component to obtain a latent vector;

[0010] Input each point in the original point cloud and the latent vector into the point transformation component, and map each point in the original point cloud to a latent space based on the latent vector through the point transformation component to obtain an encrypted point cloud.

[0011] In one embodiment, the shape representation component includes a feature extractor and a mixture flow model;

[0012] The step of inputting the original point cloud into the shape representation component and mapping the original point cloud to a latent space based on the shape representation component to obtain a latent vector includes:

[0013] Input the original point cloud into the feature extractor to obtain a one-dimensional vector;

[0014] Input the one-dimensional vector into the mixture flow model, and map the one-dimensional vector to a latent space based on the mixture flow model to obtain a latent vector.

[0015] In one embodiment, the point transformation component includes a conditional flow model;

[0016] The step of inputting each point in the original point cloud and the latent vector into the point transformation component, and mapping each point in the original point cloud to a latent space based on the latent vector through the point transformation component to obtain an encrypted point cloud includes:

[0017] Input each point in the original point cloud and the latent vector into the point transformation component, and map each point in the original point cloud to a latent space based on the point transformation component with the latent vector as a condition to obtain an encrypted point cloud.

[0018] In one embodiment, the point cloud encryption method based on a Gaussian mixture flow model includes:

[0019] Rotate the latent vector through a preset vector orthogonal matrix to obtain a target vector.

[0020] In one embodiment, before the step of mapping the original point cloud to a latent space subject to a mixture of Gaussian distributions based on a preset point cloud protection model to obtain an encrypted point cloud, it further includes:

[0021] Obtain training point clouds, input the training point clouds into a preset feature extractor to obtain first training data, and input the first training data into a preset mixture flow model to obtain second training data;

[0022] Input the second training data into a preset first loss function to obtain a first loss value, where the first loss function includes a negative log-likelihood function and cross-entropy;

[0023] Input each point in the training point clouds and the second training data into a preset conditional flow model to obtain third training data;

[0024] Input the third training data into a preset second loss function to obtain a second loss value, where the second loss function includes a negative log-likelihood function and cross-entropy;

[0025] Based on the first loss value and the second loss value, determine the gradients of the model parameters in the preset feature extractor, the preset mixture flow model, and the preset conditional flow model through the backpropagation algorithm, and use an optimization algorithm to update the model parameters according to the gradients to obtain the point cloud protection model.

[0026] In one embodiment, after the step of rotating each point of the encrypted point cloud through a preset target orthogonal matrix to obtain a target point cloud, the following steps are further included:

[0027] Send the target point cloud to a receiving end, where the receiving end performs an inverse rotation on the target point cloud based on the target orthogonal matrix to obtain the encrypted point cloud, and maps the encrypted point cloud to the data space through the inverse processing process of the point cloud protection model to obtain the original point cloud.

[0028] In addition, to achieve the above object, the present application further proposes a point cloud encryption device, where the device includes: a memory, a processor, and a computer program stored on the memory and executable on the processor, and the computer program is configured to implement the steps of the point cloud encryption method based on a Gaussian mixture flow model as described above.

[0029] In addition, to achieve the above object, the present application further proposes a storage medium, where the storage medium is a computer-readable storage medium, and a computer program is stored on the storage medium, and when the computer program is executed by a processor, the steps of the point cloud encryption method based on a Gaussian mixture flow model as described above are implemented.

[0030] In addition, to achieve the above object, the present application further provides a computer program product, where the computer program product includes a computer program, and when the computer program is executed by a processor, the steps of the point cloud encryption method based on a Gaussian mixture flow model as described above are implemented.

[0031] In this application, the original point cloud is mapped to a latent space subject to a mixture of Gaussian distributions based on a preset point cloud protection model to obtain an encrypted point cloud. The point cloud protection model includes a shape representation component constructed based on a Gaussian mixture flow model and a point transformation component constructed based on a Gaussian mixture flow model. In the process of mapping the original point cloud through the point cloud protection model, the original point cloud is mapped to a latent vector by the shape representation component, and the original point cloud is mapped to an encrypted point cloud based on the latent vector by the point transformation component; each point of the encrypted point cloud is rotated by a preset target orthogonal matrix to obtain a target point cloud.

[0032] In this application, a point cloud protection model is constructed based on a Gaussian mixture flow model, and the original point cloud is mapped to a latent space subject to a mixture of Gaussian distributions through the point cloud protection model, realizing the application of the Gaussian mixture flow model to 3D point cloud data and overcoming the difficulty that traditional flow models cannot be directly applied to point clouds due to the irregularity, sparsity, and disorder of point cloud data. When the user holds a decoder and a target orthogonal matrix, the target point cloud can be restored to the original point cloud, enabling the encrypted point cloud to be used normally in downstream application scenarios, thereby improving the practicality of the encrypted point cloud.

[0033] In addition, in this application, the encrypted point cloud is rotationally encrypted through an orthogonal matrix, realizing a privacy protection framework integrating the point cloud protection model and the rotation key, and realizing the retention of sufficient semantic information to support classification and segmentation tasks without revealing 3D geometric information, improving the security and practicality of the encrypted point cloud. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] The accompanying drawings here are incorporated into the specification and form a part of this specification, showing embodiments consistent with this application, and are used together with the specification to explain the principles of this application.

[0035] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the following will briefly introduce the accompanying drawings required for use in the description of the embodiments or the prior art. Obviously, for those of ordinary skill in the art, other drawings can also be obtained based on these drawings without creative efforts.

[0036] Figure 1 It is a schematic flowchart provided for Embodiment 1 of the point cloud encryption method based on a Gaussian mixture flow model in this application;

[0037] Figure 2 It is a point cloud protection model based on a Gaussian mixture flow model provided for an embodiment of the point cloud encryption method based on a Gaussian mixture flow model in this application;

[0038] Figure 3The flowchart of point cloud encryption and decryption provided by an embodiment of the point cloud encryption method based on the Gaussian mixture flow model in this application;

[0039] Figure 4 The visualization diagram of the application of the encrypted point cloud in the segmentation task and its segmentation result provided by an embodiment of the point cloud encryption method based on the Gaussian mixture flow model in this application;

[0040] Figure 5 The schematic diagram of the camouflage point cloud provided by an embodiment of the point cloud encryption method based on the Gaussian mixture flow model in this application;

[0041] Figure 6 The schematic diagram of the module structure of the point cloud encryption device in the embodiment of this application;

[0042] Figure 7 The schematic diagram of the device structure of the hardware operating environment involved in the point cloud encryption method based on the Gaussian mixture flow model in the embodiment of this application.

[0043] The implementation, functional features and advantages of the purpose of this application will be further described with reference to the embodiments and the accompanying drawings. Detailed implementation manners

[0044] It should be understood that the specific embodiments described herein are only used to explain the technical solutions of this application and are not used to limit this application.

[0045] For a better understanding of the technical solutions of this application, the following will be described in detail in combination with the accompanying drawings of the specification and specific implementation manners.

[0046] The main solution of the embodiment of this application is: mapping the original point cloud to a latent space obeying the mixture Gaussian distribution based on a preset point cloud protection model to obtain an encrypted point cloud, where the point cloud protection model includes a shape representation component constructed based on the Gaussian mixture flow model and a point transformation component constructed based on the Gaussian mixture flow model. In the process of mapping the original point cloud through the point cloud protection model, the original point cloud is mapped to a latent vector through the shape representation component, and the original point cloud is mapped to an encrypted point cloud based on the latent vector through the point transformation component; rotating each point of the encrypted point cloud through a preset target orthogonal matrix to obtain a target point cloud.

[0047] In this embodiment, for the convenience of description, the following will be described with the point cloud encryption device as the execution subject.

[0048] Benefiting from the development of 3D point cloud technology, which is due to the progress of sensor devices (such as lidar) and augmented reality and mixed reality technologies, it is widely used in tasks such as classification, segmentation, localization, vehicle positioning, modeling, and 3D vision-language learning. However, due to the irregularity, sparsity, and disorder of point cloud data, achieving privacy protection in the field of 3D vision poses greater challenges, especially in maintaining the reversibility of data. Existing 3D privacy protection research mainly focuses on camera positioning tasks. For devices such as HoloLens, MagicLeap1, and iRobot Roomba, the technology of promoting point clouds to line clouds is used to hide the appearance of objects. However, the original intention of designing such methods is limited to specific positioning tasks, and the improvement process will destroy the geometric information of the point cloud, resulting in irreversibility. Another type of research simplifies the protection problem to an encryption problem and uses a chaotic system to obfuscate the position of the point cloud. However, these encrypted data need to be restored before being used in downstream tasks, increasing the complexity of the application. That is, the main disadvantages of existing 3D point cloud privacy protection are: (1) For the method of promoting point clouds to line clouds, this method is irreversible and the information of the point cloud cannot be restored in specific scenarios; (2) For the point cloud encryption method using a chaotic system, this method only obfuscates the point cloud, and the obfuscated point cloud does not retain its semantic information; (3) Traditional point cloud privacy protection methods do not provide privacy protection for classification, segmentation, and generation tasks.

[0049] This application provides a solution that applies the Gaussian mixture flow model to 3D point cloud data. Compared with one-dimensional trajectories or two-dimensional images, it overcomes the difficulty that traditional flow models cannot be directly applied to point clouds due to the irregularity, sparsity, and disorder of point cloud data. When the user holds the decoder and the target orthogonal matrix, the target point cloud can be restored to the original point cloud, enabling the encrypted point cloud to be used normally in downstream application scenarios, thus improving the practicality of the encrypted point cloud.

[0050] In addition, in this application, the encrypted point cloud is rotationally encrypted through an orthogonal matrix, realizing a privacy protection framework that integrates the point cloud protection model and the rotation key. Without revealing the 3D geometric information, sufficient semantic information is retained to support classification and segmentation tasks, improving the security and practicality of the encrypted point cloud.

[0051] It should be noted that the execution subject of this embodiment can be a computing service device with data processing, network communication, and program running functions, such as a tablet computer, a personal computer, a mobile phone, etc., or an electronic device, a point cloud encryption device, etc. that can implement the above functions. Hereinafter, taking the point cloud encryption device as an example, this embodiment and the following embodiments will be described.

[0052] Based on this, the embodiment of this application provides a point cloud encryption method based on the Gaussian mixture flow model, referring to Figure 1, Figure 1 It is a schematic flowchart of the first embodiment of the point cloud encryption method based on the Gaussian mixture flow model in this application.

[0053] In this embodiment, the point cloud encryption method based on the Gaussian mixture flow model includes steps S10 to S20:

[0054] Step S10: Map the original point cloud to a latent space that follows a mixture Gaussian distribution based on a preset point cloud protection model to obtain an encrypted point cloud. Among them, the point cloud protection model includes a shape representation component constructed based on the Gaussian mixture flow model and a point transformation component constructed based on the Gaussian mixture flow model. During the process of mapping the original point cloud through the point cloud protection model, the original point cloud is mapped to a latent vector by the shape representation component, and the original point cloud is mapped to an encrypted point cloud based on the latent vector by the point transformation component;

[0055] In this embodiment, a point cloud protection model is pre-constructed based on the Gaussian mixture flow model (FlowGMM, Flow Gaussian Mixture Model), which is called the point cloud protection model based on the Gaussian mixture flow (PointFlowGMM, PointFlow Gaussian Mixture Model). Specifically, the point cloud protection model includes a shape representation component constructed based on the Gaussian mixture flow model and a point transformation component constructed based on the Gaussian mixture flow model. Input the original point cloud data to be encrypted into the point cloud protection model. The original point cloud is mapped to a latent vector by the shape representation component, and then the original point cloud is mapped to an encrypted point cloud based on the latent vector by the point transformation component. The encrypted point cloud follows a mixture Gaussian distribution. Among them, different categories of original point clouds are mapped to different Gaussian distributions of the Gaussian mixture, so that different categories of data can be distinguished in the encryption stage, which provides convenience for subsequent data processing and analysis.

[0056] It can be understood that by mapping the point cloud data to a latent space that follows a mixture Gaussian distribution, the complexity and unpredictability of the data are increased, thereby improving the security of the data. And different categories of original point clouds are mapped to different Gaussian distributions of the Gaussian mixture, so that different categories of data can be distinguished in the encryption stage, which provides convenience for subsequent data processing and analysis.

[0057] Step S20: Rotate each point of the encrypted point cloud through a preset target orthogonal matrix to obtain a target point cloud.

[0058] It should be noted that the flow model is a model with a symmetric structure and is reversible. That is, when the original point cloud passes through the flow model to obtain the encrypted point cloud, the original point cloud can also be obtained in reverse. The advantage of this is that the protection strategy is reversible, and the point cloud can be fully restored under specific circumstances, which is not available in other point cloud encryption methods (point cloud to line cloud). However, it also poses a threat to security, that is, once the model is leaked, the point cloud can be restored, resulting in the leakage of privacy. To this end, this embodiment proposes a rotation encryption scheme, that is, using an orthogonal matrix as the key to encrypt the point cloud for rotation encryption. After the encrypted point cloud is generated, each point in the encrypted point cloud is rotated by a preset target orthogonal matrix. The characteristic of the orthogonal matrix is that its transpose matrix is equal to its inverse matrix, which can ensure the reversibility of the rotation operation. And through the rotation operation, the complexity of the encrypted point cloud data is increased, and the data security is improved.

[0059] In a feasible embodiment, after the step S20: rotating each point of the encrypted point cloud by a preset target orthogonal matrix to obtain the target point cloud, the following steps are further included:

[0060] Step S30: sending the target point cloud to the receiving end, where the receiving end performs an inverse rotation on the target point cloud based on the target orthogonal matrix to obtain the encrypted point cloud, and maps the encrypted point cloud to the data space through the inverse processing process of the point cloud protection model to obtain the original point cloud.

[0061] Send the rotated target point cloud data to the receiving end. After receiving the target point cloud, the receiving end performs an inverse rotation operation using the same target orthogonal matrix to restore the encrypted point cloud. Then, the receiving end uses the inverse processing process of the point cloud protection model to map the encrypted point cloud back to the data space to obtain the original, unencrypted point cloud data. Through the reversible rotation operation and the inverse process of the point cloud protection processing model, the integrity of the data during the encryption and decryption processes is ensured. During the data transmission process, even if the data is intercepted, due to the complex encryption and rotation operations, it is difficult for attackers to extract useful information without the target orthogonal matrix and the decoder, thereby improving the security of the point cloud data.

[0062] In this embodiment, the Gaussian mixture flow model is applied to 3D point cloud data. Compared with one-dimensional trajectories or two-dimensional images, it overcomes the difficulty that traditional flow models cannot be directly applied to point clouds due to the irregularity, sparsity, and disorder of point cloud data. Moreover, a privacy protection framework is designed using a generative model, and the protection process is integrated with a rotation key. When the user has a decoder and the correct key, the original point cloud can be recovered from the encrypted point cloud, enabling the normal use of the point cloud data downstream and improving the usability of the encrypted point cloud. Additionally, by mapping the point cloud to a latent space using the Gaussian mixture flow model and performing rotational encryption with an orthogonal matrix key, sufficient semantic information is retained to support classification and segmentation tasks without revealing 3D geometric information, further enhancing the usability of the encrypted point cloud.

[0063] Based on the first embodiment of this application, in the second embodiment of this application, for content that is the same as or similar to the above-mentioned first embodiment, reference can be made to the above introduction and will not be elaborated further hereinafter. On this basis, the step S10: mapping the original point cloud to a latent space that follows a mixture Gaussian distribution based on the Gaussian mixture flow model to obtain an encrypted point cloud includes:

[0064] Step S101, mapping the original point cloud to a latent space that follows a mixture Gaussian distribution based on a preset point cloud protection model to obtain an encrypted point cloud;

[0065] Taking the original point cloud data containing three-dimensional coordinate information as input, the original point cloud is mapped to a high-dimensional latent space with specific statistical characteristics based on a shape representation component. In this latent space, the data points are rearranged and organized to form a structure that is easier to process or analyze, namely the latent vector, for subsequent encryption processing of the point cloud by a point transformation component.

[0066] Step S102, inputting each point in the original point cloud and the latent vector into the point transformation component, and mapping each point in the original point cloud to the latent space based on the latent vector through the point transformation component to obtain an encrypted point cloud.

[0067] Each point in the original point cloud and the latent vector are used as input and input into the point transformation component. The point transformation component uses the statistical information contained in the latent vector and the spatial position information of the original point cloud to map each point to a new, encrypted latent space to ensure that the encrypted point cloud data follows a mixture Gaussian distribution in the latent space, and different categories of point clouds are mapped to different Gaussian components to obtain an encrypted point cloud. It should be noted that the encryption process may involve complex mathematical transformations, non-linear mappings, or learning-based transformation methods, which are not limited herein.

[0068] In a feasible embodiment, the shape representation component includes a feature extractor and a mixture-of-flows model; the step S101 of inputting the original point cloud into the shape representation component and mapping the original point cloud to a latent space based on the shape representation component to obtain a latent vector includes:

[0069] Step S1011: Input the original point cloud into the feature extractor to obtain a one-dimensional vector;

[0070] It should be noted that in a traditional flow model, the variable substitution formulas for X and Z are as follows:

[0071]

[0072] The core of the Gaussian mixture-of-flows model lies in mapping a class of labeled data in the original data space to a latent space that follows a Gaussian distribution in the latent Gaussian mixture space. In this way, data from different classes will be mapped to different Gaussian distributions of the Gaussian mixture. The Gaussian mixture-of-flows model uses an invertible network f to transform between the data space X and the latent space Z, that is, f: X → Z f -1 : Z → X, and these two spaces have the same dimension. In the latent space Z, the data follows a mixture of Gaussian distributions Among them, the data of the class with category k follows the corresponding Gaussian distribution p Z (z|y = k) = N(z|μ k , Σ k ), where the mean and variance of each of the k Gaussian spaces are represented by μ k and Σ k respectively.

[0073] In the Gaussian mixture-of-flows model, the variable substitution formulas for X and Z can be rewritten as follows:

[0074]

[0075] Define the loss function according to the negative log-likelihood function and the cross-entropy loss as follows:

[0076]

[0077] Among them, -logp X (x|y) is the negative log-likelihood loss, is the cross-entropy loss, and according to Bayes' formula, we can obtain:

[0078]

[0079] Combining the above formulas 2, 3, and 4, the loss function of the Gaussian mixture-of-flows model is defined as follows:

[0080]

[0081] In this embodiment, an encryption model specifically applied to point clouds, that is, a point cloud protection model, is proposed based on the Gaussian mixture flow model. The model structure is as Figure 2 shown. Given an original point cloud where n represents the number of points in the point cloud, it is converted into an encrypted point cloud to achieve the protection of the geometric privacy information of the original point cloud while retaining its usability in downstream recognition tasks. Specifically, the model proposed in this embodiment includes a shape representation component (that is, the SRC component shown in Figure 2 , Shape Representation Component) and a point transformation component (that is, the PTC component shown in Figure 2 ). The shape representation component consists of a feature extractor and a mixture flow model. The point cloud feature extractor is composed of PointNet (Deep Learning on PointSets for 3D Classification and Segmentation, deep learning on point clouds for 3D classification and segmentation) and PointNet++ (Deep Hierarchical Feature Learning on Point Sets in a MetricSpace, deep hierarchical feature learning of point sets in a metric space). The point cloud feature extraction is represented by the function h. After passing through h, the point cloud X is extracted as a one-dimensional vector w ∈ R 1×m , where m represents the dimension of the vector.

[0082] Step S1012: Input the one-dimensional vector into the mixture flow model, and map the one-dimensional vector to a latent space based on the mixture flow model to obtain a latent vector.

[0083] In this embodiment, the mixture flow model adopts the RealNVP model structure, and the function G is used to represent the mixture flow model. After the feature vector w passes through G, it is mapped to the latent space to obtain That is:

[0084]

[0085] where obeys a Gaussian distribution, that is k s is the category corresponding to each Gaussian space in the mixture Gaussian space. According to Equation 2, the variable substitution formula of the shape representation component can be rewritten as:

[0086]

[0087] In a feasible embodiment, the point conversion component includes a conditional flow model; in step S102, the step of inputting each point in the original point cloud and the latent vector into the point conversion component, and mapping each point in the original point cloud to the latent space based on the latent vector by the point conversion component to obtain the encrypted point cloud includes:

[0088] Step S1021: Input each point in the original point cloud and the latent vector into the point conversion component, and map each point in the original point cloud to the latent space based on the point conversion component with the latent vector as a condition to obtain the encrypted point cloud.

[0089] The point conversion component consists of a conditional flow model, and the input of this conditional flow model is each independent point x in the point cloud i ∈R 1×3 and is mapped into the latent space under the condition of It should be noted that the model structure of the conditional flow model is also RealNVP, but in the neural network in the coupling layer (affine coupling layer), x i and are combined simultaneously. Therefore, a conditional coupling layer is adopted, and the point conversion component is represented by the function F:

[0090]

[0091] Each point x in the point cloud i under the condition of obtains Similarly, obeys a Gaussian distribution, where k p is the category corresponding to each point. After passing all the points in the point cloud through the point conversion component, finally, The variable substitution formula of the point conversion component can be rewritten as:

[0092]

[0093] It should be noted that according to formula 5, it can be analogously obtained that the definition of the loss function of the shape representation component also consists of negative log-likelihood and cross-entropy. The negative log-likelihood function is as follows:

[0094]

[0095] The cross-entropy loss is as follows:

[0096]

[0097] In this embodiment, the loss function of the point conversion component is also defined as the combination of negative log-likelihood and cross-entropy loss. The negative log-likelihood function is as follows:

[0098]

[0099] The cross-entropy loss is:

[0100]

[0101] Finally, the loss function of the overall proposed model is as follows:

[0102]

[0103] Where λs and λp are hyperparameters. In the point conversion component, the loss needs to be calculated for each point. Therefore, the final loss needs to be obtained by accumulating the losses of all points.

[0104] Based on the loss function, the model parameters of the preset shape representation component and the preset point conversion component are optimized through the backpropagation algorithm to obtain the optimized shape representation component and the optimized point conversion component. Based on the optimized shape representation component and the optimized point conversion component, a preset point cloud protection model is constructed.

[0105] In a feasible embodiment, the method further includes:

[0106] Step S40: Rotate the latent vector through a preset vector orthogonal matrix to obtain a target vector.

[0107] In this embodiment, not only each point of the encrypted point cloud is rotated through the target orthogonal matrix to obtain the target point cloud, but also the latent vector is rotated through the vector orthogonal matrix to obtain the target vector. The target orthogonal matrix and the vector orthogonal matrix are used as double keys. Only when these two keys are available can the point cloud be reconstructed. Without a key or in the case of a missing key, the point cloud cannot be reconstructed, thereby further improving the security of the point cloud data.

[0108] That is, the rotation encryption scheme in this embodiment is: that is, the orthogonal matrix is used as the key to rotate and encrypt the latent spaces of the shape representation component and the point conversion component respectively. Specifically, the target orthogonal matrix is defined as A Z and the vector orthogonal matrix is defined as A e , where A Z ∈R 3×3 , A e ∈R m×m , A Z rotates each point in the latent space of the point conversion component (that is, each point of the encrypted point cloud), and A eRotate the latent vector of the shape representation component.

[0109] For each point in the encrypted point cloud, use A Z Rotate each point of the point cloud. Among them, Subject to a Gaussian distribution:

[0110]

[0111] Since Subject to a Gaussian distribution, and A Z Is an orthogonal matrix, so the mean of the product of the two becomes Becomes And for What is used is the product of the identity matrix. Therefore, after running, the variance of the product of the two remains unchanged. It can be found that when the latent space is rotated by an orthogonal matrix, the variance after rotation remains unchanged and is still a Gaussian distribution.

[0112] Similarly, for the latent vector Use the vector orthogonal matrix A e To perform rotation. Among them, e is subject to a Gaussian distribution:

[0113]

[0114] In this embodiment, by using A e And A Z As the dual keys, when having these two keys, the point cloud can be reconstructed. When there is no key or a key is missing, the point cloud cannot be reconstructed because the flow model has a reversible architecture. Thus, as an extension, X = F -1 (z i A Z -1 |eA e -1 )

[0115] In a feasible embodiment, before the step S10 of mapping the original point cloud to a latent space subject to a mixture Gaussian distribution based on a preset point cloud protection model to obtain an encrypted point cloud, it further includes:

[0116] Step S01, obtain training point clouds, input the training point clouds into a preset feature extractor to obtain first training data, and input the first training data into a preset mixture flow model to obtain second training data;

[0117] Obtain training point clouds, input the training point clouds into a preset feature extractor to obtain first training data, and input the first training data into a preset mixture flow model to obtain second training data.

[0118] Step S02: Input the second training data into a preset first loss function to obtain a first loss value. The first loss function includes a negative log-likelihood function and cross-entropy.

[0119] Input the second training data into a preset first loss function to obtain a first loss value. The first loss function includes a negative log-likelihood function and cross-entropy.

[0120] Step S03: Input each point in the training point cloud and the second training data into a preset conditional flow model to obtain third training data.

[0121] Input each point in the training point cloud and the second training data into a preset conditional flow model to obtain third training data.

[0122] Step S04: Input the third training data into a preset second loss function to obtain a second loss value. The second loss function includes a negative log-likelihood function and cross-entropy.

[0123] Input the third training data into a preset second loss function to obtain a second loss value. The second loss function includes a negative log-likelihood function and cross-entropy.

[0124] Step S05: Based on the first loss value and the second loss value, determine the gradients of the model parameters in the preset feature extractor, the preset hybrid flow model, and the preset conditional flow model through the backpropagation algorithm, and use an optimization algorithm to update the model parameters according to the gradients to obtain the point cloud protection model.

[0125] Based on the first loss value and the second loss value, determine the gradients of the model parameters in the preset feature extractor, the preset hybrid flow model, and the preset conditional flow model through the backpropagation algorithm, and use an optimization algorithm to update the model parameters according to the gradients to obtain an optimized preset feature extractor, an optimized preset hybrid flow model, and an optimized preset conditional flow model. Construct a shape representation component based on the optimized preset feature extractor and the optimized preset hybrid flow model, and construct a point transformation component based on the optimized preset conditional flow model.

[0126] In this embodiment, please refer to Figure 3 , Figure 3 which shows the whole process of point cloud protection, rotation, and reconstruction. In the figure, (a) the point cloud is the original point cloud, which becomes the encrypted point cloud (b) after passing through the point cloud protection model based on the Gaussian mixture flow model. The encrypted point cloud (b) becomes the target point cloud (c) after rotational encryption with the key. Then, the target point cloud (c) is rotated back with the key and the original point cloud is reconstructed through the inverse process of the point cloud protection model based on the Gaussian mixture flow model. It can be found that when having A simultaneously eand A Z When, it can be reconstructed back to the original point cloud as shown in the point cloud of Figure (g). When A is missing e When, what is obtained is (e). When A is missing Z When, what is obtained is (d). When both are missing, what is obtained is (e). It can be found that in any case, it is impossible to reconstruct back to the original point cloud, and the obtained point cloud is also completely irrelevant to the original point cloud, and the privacy information of the point cloud will not be leaked.

[0127] According to the existing research, in this embodiment, the Chamfer Distance (CD) and the Earth Mover's Distance (EMD) are used as similarity measurement indicators to measure the similarity between the original point cloud and the target point cloud. It should be noted that lower CD and EMD scores indicate higher similarity. The method of this embodiment is compared with the differential privacy method, where the Laplace noise intensity is set to different values ∈∈[5, 10]. As shown in Table 1, the CD and EMD values of the target point cloud generated in this embodiment are larger, indicating that the similarity between the original point cloud and the target point cloud is lower, thus achieving better privacy protection.

[0128] In addition, this embodiment also tests whether these point clouds can be recognized by the classifier through simulation experiments. Specifically, two classifiers are trained respectively on the original point cloud dataset (ModelNet10 dataset) and the processed point cloud dataset (processed by this embodiment or the differential privacy method), and are respectively labeled as the original point cloud and the target point cloud. The experimental results show that the classifier trained using the original dataset cannot recognize the point cloud processed by the method of this embodiment or the differential privacy method, as shown in the overall (original point cloud) and average (original point cloud) columns in Table 1. And the classifier trained using the dataset processed by the method of this embodiment can recognize the target point cloud with a relatively high accuracy, while the classifier trained using the dataset processed by the differential privacy method cannot recognize the processed dataset well when the noise is large, as shown in the overall (target point cloud) and average (target point cloud) columns in Table 1.

[0129] It should be noted that the overall accuracy represents the overall accuracy (number of correct / total number), and the average is the weighted average of the accuracy of each category (number of correct in this category / total number in this category). The experimental results show that the method of this embodiment has significant advantages compared with the differential privacy method.

[0130] It should be noted that the ModelNet dataset is a dataset widely used in 3D model classification tasks, containing more than 127,915 CAD hand-drawn point cloud models, which are divided into 55 categories. Two commonly used subsets of this dataset are ModelNet10 and ModelNet40, which contain data of 10 and 40 categories respectively and are mainly used for point cloud classification tasks. In this embodiment, the ModelNet dataset is selected as a specific implementation case to verify the effectiveness of the proposed method in point cloud classification tasks. After encrypting and transforming the point clouds in the ModelNet dataset, it can provide privacy protection while maintaining efficient classification.

[0131] Table 1: Comparison of CD and EMD scores between the original point cloud and the target point cloud

[0132]

[0133] This embodiment provides a point cloud privacy protection strategy, and the encrypted point cloud still retains the semantic information for classification. This embodiment verifies the usability of the target point cloud data generated by the proposed method. This experiment uses the ModelNet and ScanObjectNN datasets for classification tasks and divides the datasets into training sets and test sets in the same way as PointNet and PointNet++. First, train the point cloud protection model based on the Gaussian mixture flow model on the training sets of the ModelNet and ScanObjectNN datasets. Then, input the complete original dataset into the point cloud protection model based on the Gaussian mixture flow model to generate the corresponding target point cloud data.

[0134] In the classification task, in this embodiment, PointNet and PointNet++ are used as classifiers, and they are trained on the training set of the target point cloud and the training set of the original point cloud respectively, and then tested on their respective test sets. Specifically, when the feature extractor is PointNet, then the classifier for the original and encrypted point clouds is PointNet; when the feature extractor is PointNet++, then the classifier is PointNet++. Each classifier is trained for 200 epochs. Referring to Table 2, Table 2 shows the classification effect of the original point cloud and the classification effect of the target point cloud obtained by the method proposed in this embodiment. It can be seen from the table that when PointNet++ is used as the feature extractor, the accuracy rate is higher than that of PointNet, indicating that the improvement of the ability of the feature extractor can improve the overall effect. And when using the PointNet++ feature extractor, the decline range compared with the classification effect of the original point cloud is also the smallest, about a 1% - 2% decline. The overall classification accuracy of the target point cloud is comparable to that of the original point cloud (slightly lower than the original point cloud), indicating that the target point cloud maintains high data availability in the downstream classification task. It should be noted that ScanObjectNN is a real point cloud dataset commonly used in point cloud classification tasks, containing 2902 3D objects from 15 categories, and these objects are all extracted from indoor scan data. It should be noted that since this dataset is from a real scene, it is more challenging than CAD point cloud data. Through the experiments on the ScanObjectNN dataset, this embodiment further verifies the applicability and robustness of the proposed protection method in the real point cloud scenario.

[0135] Table 2: Comparison of Classification Accuracy

[0136]

[0137] This embodiment provides a point cloud privacy protection strategy, and the encrypted point cloud still retains the segmented semantic information. This embodiment verifies the data availability of the target point cloud generated by the proposed method.

[0138] In the segmentation task, in this embodiment, PointNet and PointNet++ are used as classifiers, and they are trained on the training set of the target point cloud and the training set of the original point cloud respectively, and then tested on their respective test sets. Specifically, if the feature extractor is PointNet, then the classifier for the original and target point clouds is PointNet; when the feature extractor is PointNet++, then the classifier is PointNet++. Each classifier is trained for 200 epochs. The segmentation results of the ShapeNet part dataset are shown in Table 3, and the evaluation metric for the segmentation results is the mean intersection over union (mIoU%) at the points. The target point cloud generated in this embodiment can effectively complete the point cloud segmentation task, and the segmentation accuracy only drops by 0.9% to 1.4% compared with that of the original point cloud, which proves that while protecting the privacy of the target point cloud, this embodiment can still maintain the usability of the data. In addition, it can be observed from Table 3 that the categories with more point clouds have a smaller decline compared with the results of the original point cloud, while the categories with fewer point clouds (such as bags, hats, and headphones) have a more obvious decline. The possible reason is the unbalanced class distribution in the Gaussian space. As Figure 4 shown, it demonstrates the application of the target point cloud in the segmentation task and the visualization of its segmentation results. This embodiment compares the segmentation results of the target point cloud and the original point cloud and their corresponding ground truths, and the results show that the segmentation effect of the target point cloud is comparable to that of the original point cloud.

[0139] Table 3: Segmentation results of the ShapeNet part dataset

[0140] Average Airplane Schoolbag Hat Car Chair Headphone Guitar Sample Quantity 2690 76 55 898 3758 69 787 Original Point Cloud 83.7 82.3 76.0 78.9 73.5 89.8 72.0 90.7 Target Point Cloud 82.3 80.1 70.1 67.3 72.9 87.6 58.4 90.3 Knife Desk Lamp Notebook Motorcycle Quilt Pistol Rocket Table Sample Quantity 392 1547 451 202 184 283 66 5271 Original Point Cloud 86.4 78.4 95.3 65.3 92.7 82.5 51.9 82.3 Target Point Cloud 84.9 79.6 95.3 64.6 90.6 79.2 50.5 81.3

[0141] This embodiment provides a power consumption simulation display for an embedded device. As shown in Table 4, accelerated inference is performed in a deep learning inference accelerator. In point cloud processing, only 16.3 milliseconds are required for the shape representation component, while the point conversion component takes 394.6 milliseconds, and the overall time of the point cloud protection model based on Gaussian mixture flow is 410.5 milliseconds.

[0142] Table 4: Running time and energy consumption of the proposed privacy protection strategy on the embedded device.

[0143]

[0144] This invention case provides a new way to protect the privacy of point clouds. This method can disguise the point cloud as other point clouds to protect the privacy of the point cloud, and at the same time, this method is also reversible, and the original point cloud can be recovered under specific circumstances.

[0145] By rotating the Gaussian components, while protecting the three-dimensional geometric structure of the original point cloud, the feasibility of the recognition task can be maintained. However, during this rotation process, the transformed point cloud will lose its original shape type and become an unordered point cloud. This change may enable eavesdroppers to quickly detect the existence of the privacy protection module. In some application scenarios, if it is not desired for eavesdroppers to notice the privacy protection module, a perturbed point cloud consistent with the original shape type can be generated by modifying the rotation process, which is called a camouflaged point cloud. In this method, the rotation encryption method is different from the previous rotation methods. In this method, it is necessary to ensure the mean invariance of the Gaussian distribution. Therefore, the rotation method is as follows:

[0146]

[0147] The following is the proof of its mean invariance:

[0148]

[0149] It can be seen that the distribution of the point cloud obtained by rotation in this method in the Gaussian space remains unchanged. At the same time, the flow model is a generative model. Therefore, the camouflaged point cloud can be obtained by passing the rotated point cloud through the inverse process. For details, please refer to Figure 5 , Figure 5 which shows different camouflaged point clouds obtained by a point cloud under different A e and A Z keys.

[0150] This application also provides a point cloud encryption device. For details, please refer to Figure 6 , and the point cloud encryption device includes:

[0151] A mapping module 10, configured to map the original point cloud to a latent space that follows a mixture of Gaussian distributions based on a preset point cloud protection model to obtain an encrypted point cloud. Among them, the point cloud protection model includes a shape representation component constructed based on a Gaussian mixture flow model and a point transformation component constructed based on a Gaussian mixture flow model. During the process of mapping the original point cloud through the point cloud protection model, the original point cloud is mapped to a latent vector through the shape representation component, and the original point cloud is mapped to an encrypted point cloud based on the latent vector through the point transformation component;

[0152] A rotation module 20, configured to rotate each point of the encrypted point cloud through a preset target orthogonal matrix to obtain a target point cloud.

[0153] Optionally, the mapping module 10 is configured to:

[0154] Input the original point cloud into the shape representation component, and map the original point cloud to the latent space based on the shape representation component to obtain a latent vector;

[0155] Input each point in the original point cloud and the latent vector into the point transformation component, and through the point transformation component, map each point in the original point cloud to the latent space based on the latent vector to obtain the encrypted point cloud.

[0156] Optionally, the shape representation component includes a feature extractor and a flow model; the mapping module 10 is used for:

[0157] Input the original point cloud into the feature extractor to obtain a one-dimensional vector;

[0158] Input the one-dimensional vector into the hybrid flow model, and map the one-dimensional vector to the latent space based on the hybrid flow model to obtain a latent vector.

[0159] Optionally, the point transformation component includes a conditional flow model; the mapping module 10 is used for:

[0160] Input each point in the original point cloud and the latent vector into the point transformation component, and based on the point transformation component with the latent vector as a condition, map each point in the original point cloud to the latent space to obtain the encrypted point cloud.

[0161] Optionally, the rotation module 20 is used for:

[0162] Rotate the latent vector through a preset vector orthogonal matrix to obtain a target vector.

[0163] Optionally, the device further includes a training module, which is used for:

[0164] Obtain training point clouds, input the training point clouds into a preset feature extractor to obtain first training data, and input the first training data into a preset hybrid flow model to obtain second training data;

[0165] Input the second training data into a preset first loss function to obtain a first loss value, where the first loss function includes a negative log-likelihood function and cross-entropy;

[0166] Input each point in the training point cloud and the second training data into a preset conditional flow model to obtain third training data;

[0167] Input the third training data into a preset second loss function to obtain a second loss value, where the second loss function includes a negative log-likelihood function and cross-entropy;

[0168] Based on the first loss value and the second loss value, determine the gradients of the model parameters in the preset feature extractor, the preset hybrid flow model, and the preset conditional flow model through the backpropagation algorithm, and use an optimization algorithm to update the model parameters according to the gradients to obtain the point cloud protection model.

[0169] Optionally, the device further includes a decryption module for:

[0170] Sending the target point cloud to a receiving end, where the receiving end performs inverse rotation on the target point cloud based on the target orthogonal matrix to obtain the encrypted point cloud, and maps the encrypted point cloud to the data space through the inverse processing process of the point cloud protection model to obtain the original point cloud.

[0171] The point cloud encryption device provided in this application adopts the point cloud encryption method based on the Gaussian mixture flow model in the above embodiment, and can solve the technical problem of low practicality of the encrypted point cloud in current three-dimensional point cloud privacy protection. Compared with the prior art, the beneficial effects of the point cloud encryption device provided in this application are the same as those of the point cloud encryption method based on the Gaussian mixture flow model provided in the above embodiment, and other technical features in the point cloud encryption device are the same as the features disclosed in the above embodiment method, and will not be elaborated here.

[0172] This application provides a point cloud encryption 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 point cloud encryption method based on the Gaussian mixture flow model in the above embodiment.

[0173] Next, refer to Figure 7 , which shows a schematic structural diagram of a point cloud encryption device suitable for implementing the embodiments of this application. The point cloud encryption device in the embodiments of this 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), vehicle-mounted terminals (such as vehicle-mounted navigation terminals), etc., and fixed terminals such as digital TVs, desktop computers, etc. Figure 7 The point cloud encryption device shown is only an example and should not impose any limitations on the functions and usage scope of the embodiments of this application.

[0174] As Figure 7As shown, the point cloud encryption device may include a processing device 1001 (such as a central processing unit, a graphics processing unit, etc.), which may perform various appropriate actions and processes according to a program stored in a read-only memory (ROM: Read Only Memory) 1002 or a program loaded from a storage device 1003 into a random access memory (RAM: Random Access Memory) 1004. In the RAM 1004, various programs and data required for the operation of the point cloud encryption device are also stored. The processing device 1001, the ROM 1002, and the RAM 1004 are connected to each other through a bus 1005. An input / output (I / O) interface 1006 is also connected to the bus. Generally, the following systems may be connected to the I / O interface 1006: an input device 1007 including, for example, a touch screen, a touchpad, a keyboard, a mouse, an image sensor, a microphone, an accelerometer, a gyroscope, etc.; an output device 1008 including, for example, a liquid crystal display (LCD: Liquid Crystal Display), a speaker, a vibrator, etc.; a storage device 1003 including, for example, a magnetic tape, a hard disk, etc.; and a communication device 1009. The communication device 1009 may allow the point cloud encryption device to communicate with other devices wirelessly or wiredly to exchange data. Although the figure shows a point cloud encryption device having various systems, it should be understood that it is not required to implement or have all the shown systems. More or fewer systems may be implemented or had alternatively.

[0175] In particular, according to the embodiments disclosed in the present application, the processes described above with reference to the flowcharts may be implemented as computer software programs. For example, the embodiments disclosed in the present application include a computer program product, which includes a computer program carried on a computer-readable medium, and the computer program includes program codes for executing the methods shown in the flowcharts. In such an embodiment, the computer program may be downloaded and installed from a network through the communication device, or installed from the storage device 1003, or installed from the ROM 1002. When the computer program is executed by the processing device 1001, the above functions defined in the methods of the embodiments disclosed in the present application are executed.

[0176] The point cloud encryption device provided by the present application adopts the point cloud encryption method based on the Gaussian mixture flow model in the above embodiments, and can solve the technical problem of low practicability of the encrypted point cloud in current three-dimensional point cloud privacy protection. Compared with the prior art, the beneficial effects of the point cloud encryption device provided by the present application are the same as those of the point cloud encryption method based on the Gaussian mixture flow model provided by the above embodiments, and other technical features in the point cloud encryption device are the same as the features disclosed in the method of the previous embodiment, and will not be elaborated here.

[0177] It should be understood that each part disclosed in this application can be implemented by hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in a suitable manner in any one or more embodiments or examples.

[0178] As described above, the above are only specific embodiments of this application, but the protection scope of this application is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed in this application, and all of them should be covered by the protection scope of this application. Therefore, the protection scope of this application should be subject to the protection scope of the claims.

[0179] This application provides a computer-readable storage medium having computer-readable program instructions (i.e., computer programs) stored thereon, and the computer-readable program instructions are used to execute the point cloud encryption method based on the Gaussian mixture flow model in the above embodiments.

[0180] The computer-readable storage medium provided by this application can 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 of the above. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections with one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM) or flash memory, optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the above. In this embodiment, the computer-readable storage medium can be any tangible medium that contains or stores a program, and this program can be used by or in combination with an instruction execution system or device. The program code contained on the computer-readable storage medium can be transmitted by any appropriate medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination of the above.

[0181] The above computer-readable storage medium can be included in the point cloud encryption device; it can also exist separately without being assembled into the point cloud encryption device.

[0182] The above computer-readable storage medium carries one or more programs, and when the one or more programs are executed by the point cloud encryption device, the point cloud encryption device realizes the point cloud encryption method based on the Gaussian mixture flow model in the above embodiments.

[0183] Computer program code for performing the operations of this application can be written in one or more programming languages or combinations thereof. The above-mentioned programming languages include object-oriented programming languages such as Java, Smalltalk, C++, and also include conventional procedural programming languages such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, executed as an independent software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the case of a remote computer, the remote computer can be connected to the user's computer through any kind of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computer (for example, by connecting through the Internet using an Internet service provider).

[0184] The flowcharts and block diagrams in the accompanying drawings illustrate the possible architectures, functions, and operations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in the flowchart or block diagram can represent a module, a program segment, or a part of the code, and this module, program segment, or part of the code contains one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than that marked in the accompanying drawings. For example, two consecutively represented blocks can actually be executed substantially in parallel, and they can sometimes be executed in the reverse order, depending on the functions involved. It should also be noted that each block in the block diagram and / or flowchart, as well as the combination of blocks in the block diagram and / or flowchart, can be implemented by a dedicated hardware-based system for performing the specified functions or operations, or can be implemented by a combination of dedicated hardware and computer instructions.

[0185] The modules described in the embodiments of this application can be implemented in software or in hardware. Among them, the name of the module does not constitute a limitation on the unit itself in some cases.

[0186] The readable storage medium provided in this application is a computer-readable storage medium. The computer-readable storage medium stores computer-readable program instructions (i.e., computer programs) for performing the above-mentioned point cloud encryption method based on the Gaussian mixture flow model, and can solve the technical problem of low practicality of the encrypted point cloud in current three-dimensional point cloud privacy protection. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as those of the point cloud encryption method based on the Gaussian mixture flow model provided in the above embodiments, and will not be elaborated here.

[0187] The present application also provides a computer program product, including a computer program which, when executed by a processor, implements the steps of the above-mentioned point cloud encryption method based on the Gaussian mixture flow model.

[0188] The computer program product provided by the present application can solve the technical problem of low practicality of the encrypted point cloud in current three-dimensional point cloud privacy protection. Compared with the prior art, the beneficial effects of the computer program product provided by the present application are the same as those of the point cloud encryption method based on the Gaussian mixture flow model provided in the above embodiments, and will not be elaborated here.

[0189] The above are only partial embodiments of the present application, and thus do not limit the patent scope of the present application. Any equivalent structural transformation made by using the content of the specification and drawings of the present application under the technical concept of the present application, or any direct / indirect application in other related technical fields is included in the patent protection scope of the present application.

Claims

1. A point cloud encryption method based on a Gaussian mixture flow model, characterized in that: The point cloud encryption method based on the Gaussian mixture flow model includes: Based on a preset point cloud protection model, an original point cloud is mapped to a latent space obeying a mixed Gaussian distribution to obtain an encrypted point cloud, wherein the point cloud protection model includes a shape representation component constructed based on a Gaussian mixture flow model and a point conversion component constructed based on a Gaussian mixture flow model. In the process of mapping the original point cloud through the point cloud protection model, the original point cloud is mapped to a latent vector through the shape representation component, and the original point cloud is mapped to an encrypted point cloud based on the latent vector through the point conversion component; Each point of the encrypted point cloud is rotated by a preset target orthogonal matrix to obtain a target point cloud.

2. The point cloud encryption method based on the Gaussian mixture flow model according to claim 1, characterized in that: The step of mapping the original point cloud to a latent space obeying a mixed Gaussian distribution based on a preset point cloud protection model to obtain an encrypted point cloud comprises: Inputting the original point cloud into the shape representation component, and mapping the original point cloud to a latent space based on the shape representation component to obtain a latent vector; Each point in the original point cloud and the latent vector are input into the point conversion component, and the point conversion component maps each point in the original point cloud to a latent space based on the latent vector to obtain an encrypted point cloud.

3. The point cloud encryption method based on the Gaussian mixture flow model according to claim 2, characterized in that: The shape representation component includes a feature extractor and a mixed flow model; The step of inputting the original point cloud into the shape representation component and mapping the original point cloud to a latent space to obtain a latent vector based on the shape representation component comprises: Inputting the original point cloud into the feature extractor to obtain a one-dimensional vector; The one-dimensional vector is input into the mixed flow model, and the one-dimensional vector is mapped to a latent space based on the mixed flow model to obtain a latent vector.

4. The point cloud encryption method based on the Gaussian mixture flow model according to claim 2, characterized in that: The point conversion component includes a conditional flow model; The step of inputting each point in the original point cloud and the latent vector into the point conversion component, and mapping each point in the original point cloud to a latent space based on the latent vector by the point conversion component to obtain an encrypted point cloud comprises: Each point in the original point cloud and the latent vector are input into the point conversion component, and based on the point conversion component and taking the latent vector as a condition, each point in the original point cloud is mapped to the latent space to obtain an encrypted point cloud.

5. The point cloud encryption method based on the Gaussian mixture flow model according to claim 4, characterized in that: The point cloud encryption method based on the Gaussian mixture flow model includes: The latent vector is rotated by a preset vector orthogonal matrix to obtain a target vector.

6. The point cloud encryption method based on the Gaussian mixture flow model according to claim 5, characterized in that: Before the step of mapping the original point cloud to a latent space obeying a mixed Gaussian distribution based on a preset point cloud protection model to obtain an encrypted point cloud, the method further includes: Acquire a training point cloud, input the training point cloud into a preset feature extractor to obtain first training data, and input the first training data into a preset mixed flow model to obtain second training data; Inputting the second training data into a preset first loss function to obtain a first loss value, wherein the first loss function includes a negative log-likelihood function and a cross entropy; Input each point in the training point cloud and the second training data into a preset conditional flow model to obtain third training data; The third training data and a preset second loss function are input to obtain a second loss value, wherein the second loss function includes a negative log-likelihood function and a cross entropy; Based on the first loss value and the second loss value, the gradients of the model parameters in the preset feature extractor, the preset mixed flow model and the preset conditional flow model are determined by a back propagation algorithm, and an optimization algorithm is used to update the model parameters according to the gradient to obtain the point cloud protection model.

7. The point cloud encryption method based on the Gaussian mixture flow model according to any one of claims 1 to 6, characterized in that: After the step of rotating each point of the encrypted point cloud by a preset target orthogonal matrix to obtain a target point cloud, the method further includes: The target point cloud is sent to a receiving end, wherein the receiving end inversely rotates the target point cloud based on the target orthogonal matrix to obtain the encrypted point cloud, and maps the encrypted point cloud to the data space through an inverse processing of the point cloud protection model to obtain the original point cloud.

8. A point cloud encryption device, characterized in that: The device comprises: 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 point cloud encryption method based on the Gaussian mixture flow model as described in any one of claims 1 to 7.

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 point cloud encryption method based on the Gaussian mixture flow model as described in any one of claims 1 to 7 are implemented.

10. A computer program product, characterized in that The computer program product comprises a computer program, which, when executed by a processor, implements the steps of the point cloud encryption method based on the Gaussian mixture flow model as described in any one of claims 1 to 7.

Images