Copyright verification method and device for 3D point cloud data set

Abstract

The invention discloses a copyright verification method and device for a 3D point cloud data set, and the method comprises the steps: carrying out the disturbance processing of a sample set of a target class t based on a pre-training model, and obtaining a sample set after the watermark is implanted; a data set copyright owner is close to each sample in a sample set of a target category t, a preset trigger is embedded in each sample, and a sample set with a watermark implanted is obtained; then, the data set copyright owner opens the source of the sample set implanted with the watermark or authorizes the sample set to a third party; a data set copyright owner obtains an API or an interface of a point cloud deep learning model trained by a third party, and classification results or prediction probability vectors of the point cloud deep learning model trained by the third party on target category samples without and with triggers are obtained; and judging whether the third-party model is trained on the sample set implanted with the watermark or not. According to the method, the use of the data set can be verified through the expression of the point cloud model, and the legitimate right of a copyright owner of the point cloud data set is maintained.

Description

Technical Field

[0001] The present invention belongs to the field of information processing technology, and in particular relates to a copyright verification method and device for a 3D point cloud data set. Background Art

[0002] Deep learning has received widespread attention and heated discussions in recent years. Thanks to its powerful capabilities in many fields, it has greatly changed people's lifestyles and created huge value. A large part of the great achievements of deep learning comes from the existence of high-quality data sets. Deep learning models need to be trained on existing data sets so that the models can learn the potential distribution of data in the data sets. Therefore, the quality of a data set greatly affects the final results of the model. However, the creation of a data set is not an easy task. From data collection, cleaning, and organization to the final release, the entire process requires a lot of manpower, material resources and time. The creators and maintainers of these data sets should undoubtedly get the rewards they deserve. Therefore, copyright protection of data sets has become a very meaningful research topic.

[0003] The existing copyright protection methods for datasets are all for 2D image datasets, and there has been no relevant research on copyright protection for 3D point cloud datasets. Considering the huge difference in data structure between 3D point cloud data and 2D image data, point cloud data is irregular and disordered, while 2D images are regular and ordered, so the processing methods for the two types of data are very different. In addition, due to the difference in intrinsic data structure, the model structure of point cloud-based deep learning models and image-based models is also very different. Considering the differences in data structure and deep model structure, we cannot directly migrate the copyright protection method based on 2D images to 3D point cloud datasets.

[0004] In addition, the existing 3D point cloud backdoor attack that can be used as a potential copyright verification method for 3D point cloud datasets has the following defects:

[0005] 1) Most of the existing point cloud backdoors are based on the poisoned label method, which requires modifying the sample category label and is easily detected by malicious data set theft parties and lacks concealment;

[0006] 2) Existing clean label point cloud backdoor attacks are limited by the number of categories in the dataset. The more categories a dataset contains, the lower the attack success rate, resulting in poor results on large-scale datasets and lack of scalability. Due to the lack of concealment and scalability, existing 3D point cloud backdoor attacks are not suitable as a copyright verification method for point cloud datasets in actual scenarios. Summary of the invention

[0007] The purpose of the present invention is to address the deficiencies of the prior art and provide a copyright verification method and device for a 3D point cloud data set.

[0008] The object of the present invention is achieved through the following technical solution: A copyright verification method for a 3D point cloud data set, comprising the following steps:

[0009] (1) Let the 3D point cloud dataset be D. Select a sample set D of a target category from the dataset D. t And the sample set D of non-target categories s ; For the sample set D of the target category s Each sample in is perturbed based on the pre-trained model to obtain the sample set with watermark implanted

[0010] (2) The sample set of the dataset copyright owner close to the target category t Each sample The pre-set triggers are embedded in the sample points respectively. The embedding method is random replacement, that is, randomly replacing the coordinates of some points in the original point cloud sample with the trigger to obtain the sample set after watermarking. The copyright owner of the dataset will then embed the watermarked sample set Replace the sample set D of the non-target category in the dataset D s , get the watermark data set D w : Then the watermark dataset D w Open source or license to third parties;

[0011] (3) The copyright owner of the dataset obtains the API or interface of the point cloud deep learning model trained by the third party, and determines whether the third-party model is trained on the watermark dataset D through the classification results or predicted probability vectors of the point cloud deep learning model trained by the third party on the target category samples without triggers and with triggers. w superior.

[0012] Furthermore, the step (1) specifically includes the following sub-steps:

[0013] (1.1) Let the 3D point cloud dataset be D; first, select a sample set D of a target category t from the dataset D t : in, represents any sample of target category t, t = 1, 2, ..., t, ..., M; and selects a sample set D of non-target category t from the dataset D s : in, represents any sample whose target category is not t, s=1,2,…,s,…,N;

[0014] (1.2) The deep learning model g is pre-trained using the dataset D to obtain the pre-trained model g f ;

[0015] (1.3) For the sample set D s Each sample Perform disturbance processing to make the sample After pre-training, the model g f The features on the target category t are close to obtain the optimized perturbation samples as samples close to the target category t.

[0016] For each sample Perform the above processing to obtain a sample set close to the target category t

[0017] Furthermore, the step (1.3) is specifically as follows:

[0018] For the sample set D s Each sample Perform disturbance processing to make the sample After pre-training, the model g f The features on are close to the target category t, and the objective function L is as follows:

[0019]

[0020] Where p(·) represents the perturbation function, represents the perturbation sample; g f (·) represents the sample in the pre-trained model g f The features on ; ε(·,·) represents the Euclidean distance calculation function;

[0021] Based on the sample set D t In the process of optimizing the perturbation samples, the objective function L is minimized, and the optimized perturbation samples are obtained as samples close to the target category t.

[0022] For each sample Perform the above processing to obtain a sample set close to the target category t

[0023] Furthermore, the step (1.3) is specifically as follows:

[0024] For the sample set D s Each sample Perform point cloud rotation operation to make the sample After pre-training, the model gf The features on are close to the target category t, and the objective function L1 is as follows:

[0025]

[0026] Where M(θ) is the rotation matrix with a rotation angle of θ;

[0027] Based on the sample set D t In the process of optimizing the rotation angle, the objective function L1 is minimized, and the optimized rotation angle is obtained by using multiple random initial values ​​and the optimization method based on the chain rule. and will As samples close to the target category t

[0028] For each sample Perform the above processing to obtain a sample set close to the target category t

[0029] Furthermore, the step (1.3) is specifically as follows:

[0030] For the sample set D s Each sample Perform point cloud rotation operation to make the sample After pre-training, the model g f The features on are close to the target category t, and the objective function L1 is as follows:

[0031]

[0032] Where M(θ) is the rotation matrix with a rotation angle of θ;

[0033] Based on the sample set D t In the process of optimizing the rotation angle, the objective function L1 is minimized, and the optimized rotation angle is obtained by using multiple random initial values ​​and the optimization method based on the chain rule.

[0034] Then according to the rotation angle For samples Perform a dithering operation so that the samples After pre-training, the model g f The features on are close to the target category t, and the objective function L2 is as follows:

[0035]

[0036] Where Δx represents the initial jitter size of the coordinates; η is the hyperparameter of the regularization term; ||·||2 represents the L2 norm;

[0037] Based on the sample set D t In the process of optimizing the initial jitter size, the objective function L2 is minimized, and the optimized jitter size (Δx) is obtained by using the momentum-based optimization method. s ; and As samples close to the target category t

[0038] For each sample Perform the above processing to obtain a sample set close to the target category t

[0039] Furthermore, the step (3) specifically includes the following sub-steps:

[0040] (3.1) The copyright owner of the dataset obtains the API or interface of the point cloud deep learning model trained by a third party;

[0041] (3.2) The sample set containing N samples of target category t As a trigger-free sample set, it is input into the point cloud deep learning model trained by a third party to obtain the returned prediction probability vector p b Or the predicted result y1;

[0042] (3.3) Embed the pre-set triggers into the sample set For each sample in , we get a set of samples with triggers And input it into the point cloud deep learning model trained by a third party to obtain the returned prediction probability vector p v Or the predicted result y2;

[0043] (3.4) According to the difference between the two returned prediction probability vectors or prediction results, it is judged whether the point cloud deep learning model trained by the third party is trained on the watermark dataset D w superior.

[0044] Furthermore, in step (3.4), when |p b -p v |Greater than threshold , it means that the point cloud deep learning model trained by the third party is trained on the watermark dataset D w On; or when |y1-y2| is greater than the threshold , it means that the point cloud deep learning model trained by the third party is trained on the watermark dataset D w superior.

[0045] Furthermore, in step (3.4), based on the hypothesis testing method, it is assumed that pb -p v ≤τ, where τ is the predetermined difference, 0<τ<1; calculate the corresponding p-value; when the p-value is greater than the threshold , it means that the point cloud deep learning model trained by the third party is not trained on the watermark dataset D w superior.

[0046] The present invention also includes a copyright verification device for a 3D point cloud dataset, including a memory and one or more processors, wherein the memory stores executable code, and when the one or more processors execute the executable code, they are used for the above-mentioned copyright verification method for a 3D point cloud dataset.

[0047] The present invention also includes a computer-readable storage medium on which a program is stored. When the program is executed by a processor, the above-mentioned copyright verification method for a 3D point cloud data set is implemented.

[0048] The beneficial effects of the present invention are as follows: the present invention designs a method for adding a watermark to a 3D point cloud dataset, which can verify whether a 3D point cloud classification model is trained on the watermark dataset in a black box scenario. The invention can not only protect the legitimate rights and interests of the point cloud dataset owner, thereby encouraging more people to create higher-quality datasets, but also prevent malicious use and illegal abuse of the dataset. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 is a flowchart of a copyright verification method for a 3D point cloud dataset;

[0050] Figure 2 is a basic flow chart of a copyright verification method for 3D point cloud datasets;

[0051] Figure 3 A schematic diagram of the structure of a copyright verification device for a 3D point cloud dataset. DETAILED DESCRIPTION

[0052] In order to make the purpose, technical scheme and advantages of the present invention more clear, the present invention is further described in detail in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0053] The present invention is based on the clean label backdoor attack method, and protects the legitimate rights and interests of the dataset owner by adding a watermark to the protected 3D point cloud dataset. After setting the target category, perturbation samples and triggers, the dataset owner releases the 3D point cloud dataset as open source or authorizes it to be sold to a third party for model training; for a third-party point cloud model that is suspected of illegal or unauthorized use of the dataset, the dataset copyright owner verifies whether it uses the protected dataset through the backdoor performance of the third-party model on the target category. The purpose of the present invention is to protect the legitimate rights and interests of the copyright owner of the 3D point cloud dataset, and at the same time prevent the dataset from being illegally abused by third parties.

[0054] Example 1

[0055] like Figure 1 and Figure 2 As shown, the present invention provides a copyright verification method for a 3D point cloud dataset, comprising the following steps:

[0056] (1) Let the 3D point cloud dataset be D. Select a sample set D of a target category from the dataset D. t And the sample set D of non-target categories s ; For the sample set D of the target category s Each sample in is perturbed based on the pre-trained model to obtain the sample set with watermark implanted

[0057] The step (1) specifically includes the following sub-steps:

[0058] (1.1) Let the 3D point cloud dataset be D; first, select a sample set D of a target category t from the dataset D t : in, represents any sample of target category t, t = 1, 2, ..., t, ..., M; and selects a sample set D of non-target category t from the dataset D s : in, represents any sample whose target category is not t, s=1,2,…,s,…,N.

[0059] (1.2) The deep learning model g is pre-trained using the dataset D to obtain the pre-trained model g f .

[0060] (1.3) For the sample set D s Each sample Perform disturbance processing to make the sample After pre-training, the model g f The features on the target category t are close to obtain the optimized perturbation samples as samples close to the target category t.

[0061] For each sample Perform the above processing to obtain a sample set close to the target category t

[0062] In the first method, the step (1.3) is specifically:

[0063] For the sample set D s Each sample Perform disturbance processing to make the sample After pre-training, the model g f The features on are close to the target category t, and the objective function L is as follows:

[0064]

[0065] Where p(·) represents the perturbation function, represents the perturbation sample; g f (·) represents the sample in the pre-trained model g f The features on ; ε(·,·) represents the Euclidean distance calculation function;

[0066] Based on the sample set D t In the process of optimizing the perturbation samples, the objective function L is minimized, and the optimized perturbation samples are obtained as samples close to the target category t.

[0067] For each sample Perform the above processing to obtain a sample set close to the target category t

[0068] In the second method, the step (1.3) is specifically as follows:

[0069] For the sample set D s Each sample Perform point cloud rotation operation to make the sample After pre-training, the model g f The features on are close to the target category t, and the objective function L1 is as follows:

[0070]

[0071] Where M(θ) is the rotation matrix with a rotation angle of θ;

[0072] Based on the sample set D tIn the process of optimizing the rotation angle, the objective function L1 is minimized, and the optimized rotation angle is obtained by using multiple random initial values ​​and the optimization method based on the chain rule. and will As samples close to the target category t

[0073] For each sample Perform the above processing to obtain a sample set close to the target category t

[0074] In the third method, the step (1.3) is specifically as follows:

[0075] For the sample set D s Each sample Perform point cloud rotation operation to make the sample After pre-training, the model g f The features on are close to the target category t, and the objective function L1 is as follows:

[0076]

[0077] Where M(θ) is the rotation matrix with a rotation angle of θ;

[0078] Based on the sample set D t In the process of optimizing the rotation angle, the objective function L1 is minimized, and the optimized rotation angle is obtained by using multiple random initial values ​​and the optimization method based on the chain rule.

[0079] Then according to the rotation angle For samples Perform a dithering operation so that the samples After pre-training, the model g f The features on are close to the target category t, and the objective function L2 is as follows:

[0080]

[0081] Where Δx represents the initial jitter size of the coordinates; η is the hyperparameter of the regularization term; ||·||2 represents the L2 norm;

[0082] Based on the sample set D t In the process of optimizing the initial jitter size, the objective function L2 is minimized, and the optimized jitter size (Δx) is obtained by using the momentum-based optimization method. s ; and As samples close to the target category t

[0083] For each sample Perform the above processing to obtain a sample set close to the target category t

[0084] (2) The sample set of the dataset copyright owner close to the target category t Each sample The pre-set triggers are embedded in the sample points respectively. The embedding method is random replacement, that is, randomly replacing the coordinates of some points in the original point cloud sample with the trigger to obtain the sample set after watermarking. The copyright owner of the dataset will then embed the watermarked sample set Replace the sample set D of the non-target category in the dataset D s , get the watermark data set D w : Then the watermark dataset D w Open source or licensed to third parties.

[0085] (3) The copyright owner of the dataset obtains the API or interface of the point cloud deep learning model trained by the third party, and determines whether the third-party model is trained on the watermark dataset D through the classification results or predicted probability vectors of the point cloud deep learning model trained by the third party on the target category samples without triggers and with triggers. w superior.

[0086] The step (3) specifically includes the following sub-steps:

[0087] (3.1) The copyright owner of the dataset obtains the API or interface of the point cloud deep learning model trained by a third party;

[0088] (3.2) The sample set containing N samples of target category t As a trigger-free sample set, it is input into the point cloud deep learning model trained by a third party to obtain the returned prediction probability vector p b Or the predicted result y1;

[0089] (3.3) Embed the pre-set triggers into the sample set For each sample in , we get a set of samples with triggers And input it into the point cloud deep learning model trained by a third party to obtain the returned prediction probability vector p v Or the predicted result y2;

[0090] (3.4) According to the difference between the two returned prediction probability vectors or prediction results, it is judged whether the point cloud deep learning model trained by the third party is trained on the watermark dataset D w superior.

[0091] In step (3.4), when |p b -p v |Greater than threshold , it means that the point cloud deep learning model trained by the third party is trained on the watermark dataset D w On; or when |y1-y2| is greater than the threshold , it means that the point cloud deep learning model trained by the third party is trained on the watermark dataset D w superior.

[0092] In step (3.4), based on the hypothesis testing method, it is assumed that p b -p v ≤τ, where τ is the predetermined difference, 0<τ<1; calculate the corresponding p-value; when the p-value is greater than the threshold , it means that the point cloud deep learning model trained by the third party is not trained on the watermark dataset D w superior.

[0093] Figure 2 This is a basic flow chart of the copyright verification method for 3D point cloud datasets. The entire process is divided into two stages: watermark implantation and watermark verification. In the watermark implantation stage, after selecting the target category and a part of the non-target category samples, the selected non-target category samples are first moved close to the target category in the feature space by perturbation, and then embedded with the pre-set trigger. After completing the above watermark implantation operation, the dataset owner will open source the dataset or authorize it to be sold to a third-party person or organization. After obtaining the watermarked point cloud dataset, the third party trains a 3D point cloud deep learning model based on the dataset. In the watermark verification stage, the dataset copyright owner only needs to obtain the API or interface of the point cloud deep learning model trained by the third party, and judge whether the third-party model is trained on the watermarked dataset through the classification results or predicted probability vectors of the model on the target category samples without triggers and with triggers.

[0094] Example 2

[0095] The present embodiment relates to a copyright verification device for a 3D point cloud dataset, comprising a memory and one or more processors, wherein the memory stores executable code, and when the one or more processors execute the executable code, a copyright verification method for a 3D point cloud dataset in the above-mentioned embodiment 1 is used; the device embodiment can be applied to any device with data processing capabilities, and the any device with data processing capabilities can be a device or apparatus such as a computer.

[0096] like Figure 3At the hardware level, the knowledge distillation device includes a processor, an internal bus, a network interface, a memory, and a non-volatile memory, and may also include other hardware required for the business. The processor reads the corresponding computer program from the non-volatile memory into the memory and then runs it to achieve the above Figure 1 Of course, in addition to the software implementation, the present invention does not exclude other implementations, such as logic devices or a combination of software and hardware, etc., that is, the execution subject of the following processing flow is not limited to each logic unit, but can also be hardware or logic devices.

[0097] For the improvement of a technology, it can be clearly distinguished whether it is a hardware improvement (for example, improvement of the circuit structure of diodes, transistors, switches, etc.) or a software improvement (improvement of the method flow). However, with the development of technology, many improvements of the method flow today can be regarded as direct improvements of the hardware circuit structure. Designers almost always obtain the corresponding hardware circuit structure by programming the improved method flow into the hardware circuit. Therefore, it cannot be said that the improvement of a method flow cannot be implemented with a hardware entity module. For example, a programmable logic device (PLD) (such as a field programmable gate array (FPGA)) is such an integrated circuit whose logical function is determined by the user's programming of the device. Designers can "integrate" a digital system on a PLD by programming themselves, without having to ask chip manufacturers to design and make dedicated integrated circuit chips. Moreover, nowadays, instead of manually making integrated circuit chips, this kind of programming is mostly implemented by "logic compiler" software, which is similar to the software compiler used when developing and writing programs, and the original code before compilation must also be written in a specific programming language, which is called hardware description language (HDL). There is not only one HDL, but many kinds, such as ABEL (Advanced Boolean Expression Language), AHDL (Altera Hardware Description Language), Confluence, CUPL (Cornell University Programming Language), HDCal, JHDL (Java Hardware Description Language), Lava, Lola, MyHDL, PALASM, RHDL (Ruby Hardware Description Language), etc. The most commonly used ones are VHDL (Very-High-Speed ​​Integrated Circuit Hardware Description Language) and Verilog. Those skilled in the art should also know that it is only necessary to program the method flow slightly in the above-mentioned hardware description languages ​​and program it into the integrated circuit, and then it is easy to obtain the hardware circuit that implements the logic method flow.

[0098] The controller can be implemented in any appropriate manner, for example, the controller can take the form of a microprocessor or processor and a computer-readable medium storing a computer-readable program code (such as software or firmware) that can be executed by the (micro)processor, a logic gate, a switch, an application-specific integrated circuit (ASIC), a programmable logic controller, and an embedded microcontroller. Examples of controllers include, but are not limited to, the following microcontrollers: ARC 625D, Atmel AT91SAM, Microchip PIC18F26K20, and Silicone Labs C8051F320. The memory controller can also be implemented as part of the control logic of the memory. Those skilled in the art also know that in addition to implementing the controller in a purely computer-readable program code manner, the controller can be implemented in the form of a logic gate, a switch, an application-specific integrated circuit, a programmable logic controller, and an embedded microcontroller by logically programming the method steps. Therefore, this controller can be considered as a hardware component, and the devices included therein for implementing various functions can also be regarded as structures within the hardware component. Or even, the devices for implementing various functions can be regarded as both software modules for implementing the method and structures within the hardware component.

[0099] The systems, devices, modules or units described in the above embodiments may be implemented by computer chips or entities, or by products with certain functions. A typical implementation device is a computer. Specifically, the computer may be, for example, a personal computer, a laptop computer, a cellular phone, a camera phone, a smart phone, a personal digital assistant, a media player, a navigation device, an email device, a game console, a tablet computer, a wearable device, or a combination of any of these devices.

[0100] It should also be noted that the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, commodity or device. In the absence of more restrictions, the elements defined by the sentence "comprises a ..." do not exclude the existence of other identical elements in the process, method, commodity or device including the elements.

[0101] It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, systems or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.

[0102] The present invention may be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc. that perform specific tasks or implement specific abstract data types. The present invention may also be practiced in distributed computing environments where tasks are performed by remote processing devices connected through a communication network. In a distributed computing environment, program modules may be located in local and remote computer storage media, including storage devices.

[0103] Example 3

[0104] An embodiment of the present invention further provides a computer-readable storage medium having a program stored thereon. When the program is executed by a processor, a copyright verification method for a 3D point cloud data set according to the above-mentioned embodiment 1 is implemented.

[0105] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A copyright verification method for a 3D point cloud dataset, characterized in that: The following steps are involved: (1) Let the 3D point cloud dataset be D. Select a sample set D of a target category from the dataset D. t And the sample set D of non-target categories s ; For the sample set D of the target category s Each sample in is perturbed based on the pre-trained model to obtain the sample set with watermark implanted (2) The sample set of the dataset copyright owner close to the target category t Each sample The pre-set triggers are embedded in the sample points respectively. The embedding method is random replacement, that is, randomly replacing the coordinates of some points in the original point cloud sample with the trigger to obtain the sample set after watermarking. The copyright owner of the dataset will then embed the watermarked sample set Replace the sample set D of the non-target category in the dataset D s , get the watermark data set D w : Then the watermark dataset D w Open source or license to third parties; (3) The copyright owner of the dataset obtains the API or interface of the point cloud deep learning model trained by the third party, and determines whether the third-party model is trained on the watermark dataset D through the classification results or predicted probability vectors of the point cloud deep learning model trained by the third party on the target category samples without triggers and with triggers. w superior.

2. A copyright verification method for a 3D point cloud dataset according to claim 1, characterized in that: The step (1) specifically includes the following sub-steps: (1.1) Let the 3D point cloud dataset be D; first, select a sample set D of a target category t from the dataset D t : in, represents any sample of target category t, t = 1, 2, ..., t, ..., M; and selects a sample set D of non-target category t from the dataset D s : in, represents any sample whose target category is not t, s=1,2,…,s,…,N; (1.2) The deep learning model g is pre-trained using the dataset D to obtain the pre-trained model g f ; (1.3) For the sample set D s Each sample Perform disturbance processing to make the sample After pre-training, the model g f The features on the target category t are close to obtain the optimized perturbation samples as samples close to the target category t. For each sample Perform the above processing to obtain a sample set close to the target category t 3. A copyright verification method for a 3D point cloud data set according to claim 2, characterized in that: The step (1.3) is specifically: For the sample set D s Each sample Perform disturbance processing to make the sample After pre-training, the model g f The features on are close to the target category t, and the objective function L is as follows: Where p(·) represents the perturbation function, represents the perturbation sample; g f (·) represents the sample in the pre-trained model g f The features on ; ε(·,·) represents the Euclidean distance calculation function; Based on the sample set D t In the process of optimizing the perturbation samples, the objective function L is minimized, and the optimized perturbation samples are obtained as samples close to the target category t. For each sample Perform the above processing to obtain a sample set close to the target category t 4. A copyright verification method for a 3D point cloud data set according to claim 2, characterized in that: The step (1.3) is specifically: For the sample set D s Each sample Perform point cloud rotation operation to make the sample After pre-training, the model g f The features on are close to the target category t, and the objective function L1 is as follows: Where M(θ) is the rotation matrix with a rotation angle of θ; Based on the sample set D t In the process of optimizing the rotation angle, the objective function L1 is minimized, and the optimized rotation angle is obtained by using multiple random initial values ​​and the optimization method based on the chain rule. and will As a sample close to the target category t For each sample Perform the above processing to obtain a sample set close to the target category t 5. A copyright verification method for a 3D point cloud data set according to claim 2, characterized in that: The step (1.3) is specifically: For the sample set D s Each sample Perform point cloud rotation operation to make the sample After pre-training, the model g f The features on are close to the target category t, and the objective function L1 is as follows: Where M(θ) is the rotation matrix with a rotation angle of θ; Based on the sample set D t In the process of optimizing the rotation angle, the objective function L1 is minimized, and the optimized rotation angle is obtained by using multiple random initial values ​​and the optimization method based on the chain rule. Then according to the rotation angle For samples Perform a dithering operation so that the samples After pre-training, the model g f The features on are close to the target category t, and the objective function L2 is as follows: Where Δx represents the initial jitter size of the coordinates; η is the hyperparameter of the regularization term; ||·||2 represents the L2 norm; Based on the sample set D t In the process of optimizing the initial jitter size, the objective function L2 is minimized, and the optimized jitter size (Δx) is obtained by using the momentum-based optimization method. s ; and As a sample close to the target category t For each sample Perform the above processing to obtain a sample set close to the target category t 6. A copyright verification method for a 3D point cloud data set according to claim 1, characterized in that: The step (3) specifically includes the following sub-steps: (3.1) The copyright owner of the dataset obtains the API or interface of the point cloud deep learning model trained by a third party; (3.2) The sample set containing N samples of target category t As a trigger-free sample set, it is input into the point cloud deep learning model trained by a third party to obtain the returned prediction probability vector p b Or the predicted result y1; (3.3) Embed the pre-set triggers into the sample set For each sample in , we get a set of samples with triggers And input it into the point cloud deep learning model trained by a third party to obtain the returned prediction probability vector p v Or the predicted result y2; (3.4) According to the difference between the two returned prediction probability vectors or prediction results, it is judged whether the point cloud deep learning model trained by the third party is trained on the watermark dataset D w superior.

7. A copyright verification method for a 3D point cloud data set according to claim 6, characterized in that: In step (3.4), when |p b -p v |Greater than threshold , it means that the point cloud deep learning model trained by the third party is trained on the watermark dataset D w On; or when |y1-y2| is greater than the threshold , it means that the point cloud deep learning model trained by the third party is trained on the watermark dataset D w superior.

8. The copyright verification method for a 3D point cloud data set according to claim 6, characterized in that: In step (3.4), based on the hypothesis testing method, it is assumed that p b -p v ≤τ, where τ is the predetermined difference, 0<τ<1; calculate the corresponding p-value; when the p-value is greater than the threshold , it means that the point cloud deep learning model trained by the third party is not trained on the watermark dataset D w superior.

9. A copyright verification device for a 3D point cloud data set, characterized in that: It comprises a memory and one or more processors, wherein the memory stores executable code, and when the one or more processors execute the executable code, they are used to implement a copyright verification method for a 3D point cloud data set as described in any one of claims 1-8.

10. A computer-readable storage medium, characterized in that: A program is stored thereon, and when the program is executed by a processor, a copyright verification method for a 3D point cloud data set as described in any one of claims 1-8 is implemented.

Images