Regitive digital camouflage generation method, device and equipment

Abstract

The invention provides an adversarial digital camouflage generation method, device and equipment, and the method comprises the steps: carrying out the encoding of an obtained target background image based on an encoder, and forming a first hidden variable; performing denoising processing on the first hidden variable in a hidden space based on a denoising diffusion model to form a second hidden variable; processing the second hidden variable based on a decoder to generate a digital camouflage image; processing the first hidden variable based on an adversarial noise generator to generate an adversarial noise image; and obtaining an adversarial digital camouflage image corresponding to the target background image based on the combination of the camouflage image and the adversarial noise image. According to the adversarial digital camouflage generation method provided by the invention, the adversarial digital camouflage image can be quickly and accurately generated, so that the camouflage effect is good and the adversarial digital camouflage image is not easy to find when facing a neural network classification model.

Description

Technical Field

[0001] The embodiments of the present invention relate to the technical field of digital camouflage, and in particular to a method, device and equipment for generating confrontational digital camouflage. Background Art

[0002] Traditional digital camouflage generation methods often require steps such as background main color extraction, camouflage patch generation, and color filling to complete the final camouflage design. The generation steps are complex and the labor cost is high. In addition, the generated digital camouflage is usually only camouflaged for human vision, and the camouflage effect is poor when facing a neural network classification model. Summary of the invention

[0003] The technical problem to be solved by the present invention is to provide a method, device and equipment for generating adversarial digital camouflage which can achieve good camouflage effect and is not easy to be discovered when facing a neural network classification model.

[0004] The present invention includes a method for generating a confrontational digital camouflage, comprising: Encoding the obtained target background image based on the encoder to form a first latent variable; Performing denoising processing on the first latent variable in latent space based on a denoising diffusion model to form a second latent variable; Processing the second latent variable based on the decoder to generate a digital camouflage image; Processing the first latent variable based on the adversarial noise generator to generate an adversarial noise image; Based on the combination of the digital camouflage image and the adversarial noise image, an adversarial digital camouflage image corresponding to the target background image is obtained.

[0005] In an optional embodiment, the method further includes: Build the initial encoder; The perceptual error is used as the reconstruction error, and the initial encoder is pre-trained based on the adversarial training mechanism and the training background image set to obtain the features for extracting the input background image, and the background image is encoded into the latent space to form the encoder of the latent variable in combination with the features.

[0006] In an optional embodiment, the method further includes: Construct an initial denoising diffusion model; In the latent space, the initial denoising diffusion model is trained based on a training data set to form the denoising diffusion model, wherein the training data set includes a first latent variable set corresponding to a background image data set and a second latent variable set corresponding to a digital camouflage image data set.

[0007] In an optional embodiment, in the latent space, the training process of the initial denoising diffusion model includes a forward diffusion process; The forward diffusion process includes: The hidden variables in the first hidden variable set obtained T times Add the target Gaussian noise to form the corresponding hidden variable Gaussian distribution, the latent variable The Gaussian distribution of is the standard Gaussian distribution.

[0008] In an optional embodiment, the target Gaussian noise is subject to the mean ,variance The Gaussian distribution of is the identity matrix, is a preset value and changes with time step Linear change, the is the corresponding time step The hidden variable, when When it increases, the The Gaussian distribution of is close to the pure noise distribution when When The Gaussian distribution of is the standard Gaussian distribution.

[0009] In an optional embodiment, in the latent space, the training process of the initial denoising diffusion model further includes a denoising inference process; The denoising inference process includes: Learning the inverse process distribution of the Gaussian distribution based on the target neural network, wherein the inverse process distribution is used to restore the latent variable according to the Gaussian distribution, and the target neural network is a neural network that uses an attention mechanism for learning; The denoising process is performed on the first latent variable in the latent space based on the denoising diffusion model to form a second latent variable, including: Predicting the Gaussian distribution of noise in each step of the denoising process based on the denoising diffusion model; After performing the T times of denoising based on the predicted Gaussian distribution of the noise, the second latent variable is formed.

[0010] In an optional embodiment, the method further includes: Constructing the adversarial noise generator based on convolutional layers, residual blocks, downsampling layers, and attention modules; In the latent space, the adversarial noise generator is trained so that it can perform noise modeling based on the input latent variables, determine the distribution space of the noise disturbance, and then generate adversarial noise corresponding to the latent variables.

[0011] In an optional embodiment, the loss function of the adversarial noise generator is: Said B represents the adversarial perturbation batch size, n' is from [1, B ] is a random integer, and Respectively represent the first n data samples and The adversarial perturbation generated by data samples, represents the distance metric, , Representative classification network Layer output function, Representative n data samples.

[0012] Another embodiment of the present invention also provides a device for generating a confrontational digital camouflage, including: An encoding module, used for encoding the target background image obtained by the encoder to form a first latent variable; A denoising module, configured to perform denoising processing on the first latent variable in a latent space according to a denoising diffusion model to form a second latent variable; A first processing module, configured to generate a digital camouflage image according to the decoder processing the second latent variable; A second processing module, configured to process the first latent variable according to an adversarial noise generator to generate an adversarial noise image; A generating module is used to combine the digital camouflage image with the adversarial noise image to generate an adversarial digital camouflage image corresponding to the target background image.

[0013] Another embodiment of the present invention further provides an electronic device, including: at least one processor; and, a memory communicatively coupled to the at least one processor; The memory stores instructions executable by the at least one processor, and the instructions are configured to execute the method for generating a confrontational digital camouflage as described in any one of the embodiments above.

[0014] The beneficial effects of the present invention include encoding the obtained target background image through an encoder to form a first latent variable, then denoising the first latent variable in the latent space based on a denoising diffusion model to form a second latent variable, using a decoder to process the second latent variable to generate a camouflage image, and then processing the first latent variable based on an adversarial noise generator to generate an adversarial noise image. At this time, an adversarial digital camouflage image corresponding to the target background image can be obtained based on the combination of the camouflage image and the adversarial noise image. This process is fast and accurate, and in this embodiment, the digital camouflage generation task is regarded as an image-to-image translation task from the background image domain to the digital camouflage image domain, and the automatic extraction of background image features and the generation of the corresponding digital camouflage image are achieved by constructing a related model. That is, the method in this embodiment belongs to an end-to-end digital camouflage generation method, which not only does not require background-camouflage pairing training data, saves data collection and manual labeling costs, but also can automatically generate digital camouflage according to the input background image, and can resist the camouflage recognition of the intelligent recognition algorithm, significantly improving the camouflage effect.

[0015] Other features and advantages of the present invention will be set forth in the description which follows, and in part will be apparent from the description, or may be learned by practice of the present invention.

[0016] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0018] Figure 1 It is a schematic flow chart of the confrontation digital camouflage generation method of the present invention.

[0019] Figure 2 It is a schematic application flow chart of the confrontation digital camouflage generation method of the present invention.

[0020] Figure 3 It is a schematic diagram of an application process of the confrontation digital camouflage generation method of the present invention.

[0021] Figure 4 It is a schematic structural block diagram of the confrontation digital camouflage generating device of the present invention. DETAILED DESCRIPTION

[0022] The specific embodiments of the present invention are described in detail below with reference to the accompanying drawings, but are not intended to limit the present invention.

[0023] It should be understood that various modifications may be made to the embodiments disclosed herein. Therefore, the following description should not be considered as limiting, but merely as an example of an embodiment. Other modifications within the scope and spirit of the present disclosure will occur to those skilled in the art.

[0024] The accompanying drawings, which are incorporated in and constitute a part of the specification, illustrate embodiments of the present application and, together with the general description of the present disclosure given above and the detailed description of the embodiments given below, serve to explain the principles of the present invention.

[0025] These and other characteristics of the invention will become apparent from the following description of a preferred form of embodiment given as a non-limiting example, with reference to the accompanying drawings.

[0026] It should also be understood that, although the present application has been described with reference to some specific examples, those skilled in the art will be able to realize many other equivalent forms of the present invention that have the characteristics described in the claims and are therefore within the scope of protection defined thereby.

[0027] The above and other aspects, features and advantages of the present invention will become more apparent from the following detailed description when taken in conjunction with the accompanying drawings.

[0028] Specific embodiments of the present invention are described hereinafter with reference to the accompanying drawings; however, it should be understood that the disclosed embodiments are merely examples of the present invention, which may be implemented in a variety of ways. Well-known and / or repeated functions and structures are not described in detail to avoid obscuring the present disclosure with unnecessary or redundant details. Therefore, the specific structural and functional details disclosed herein are not intended to be limiting, but merely serve as a basis and representative basis for the claims to teach those skilled in the art to use the present disclosure in a variety of ways with substantially any suitable detailed structure.

[0029] This specification may use the phrases "in one embodiment," "in another embodiment," "in a further embodiment," or "in other embodiments," all of which may refer to one or more of the same or different embodiments according to the present disclosure.

[0030] The embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0031] like Figure 1 As shown, the present invention provides a method for generating a confrontational digital camouflage, comprising: S1: Encode the obtained target background image based on the encoder to form a first latent variable; S2: performing denoising processing on the first latent variable in latent space based on a denoising diffusion model to form a second latent variable; S3: Based on the decoder, the second latent variable is processed to generate a digital camouflage image; S4: Processing the first latent variable based on the adversarial noise generator to generate an adversarial noise image; S5: Obtaining an adversarial digital camouflage image corresponding to the target background image based on the combination of the digital camouflage image and the adversarial noise image.

[0032] The method in this embodiment actually maps the image to the latent space through the pre-trained encoder to extract the structural content and other features of the background image, and then uses the diffusion model to process the encoded space to form the denoised latent variables, and then the decoder generates a digital camouflage image based on the latent variables, and combines this digital camouflage image with the result generated by the adversarial noise generator to obtain an adversarial digital camouflage image corresponding to the background image. Based on this image, the user can combine it with any target that needs to be camouflaged in the background corresponding to the target background image to obtain an adversarial digital camouflage image when the target is located in the background, which can successfully avoid the intelligent recognition algorithm and has a good camouflage effect.

[0033] Therefore, based on the above content, it can be seen that the method process of this embodiment is fast and accurate. By considering the digital camouflage generation task as an image-to-image translation task from the background image domain to the digital camouflage image domain, and by building a related model, the automatic extraction of background image features and the generation of corresponding digital camouflage images are realized. That is, the method in this embodiment belongs to an end-to-end digital camouflage generation method. This method not only does not require background-camouflage pairing training data, saving the cost of data collection and manual labeling, but also can automatically generate digital camouflage according to the input background image, and can resist the camouflage recognition of the intelligent recognition algorithm, significantly improving the camouflage effect.

[0034] Furthermore, if Figure 2 As shown, the method of this embodiment actually takes the neural network classification model (the above-mentioned intelligent recognition algorithm) as the attack target, transforms the camouflage background to the latent space through the encoder, performs a diffusion process in the latent space, and then generates digital camouflage through the decoder. Then, adversarial noise is generated in the latent space through the deconvolution layer, and the noise is combined with the camouflage generated after decoding to generate an adversarial digital camouflage image. When applied later, the digital camouflage can be covered on the camouflage target through digital image processing technology. In order to ensure the camouflage effect, the camouflage target image covered with digital camouflage can also be input into the discriminator (i.e., the attack target) for discrimination, and the model in this embodiment is continuously trained based on the discrimination result to achieve a more accurate attack effect. This method is also used in the verification process of the method in this embodiment.

[0035] Specifically, the method described in this embodiment also includes: S6: construct the initial encoder; S7: Using the perceptual error as the reconstruction error, and pre-training the initial encoder based on the adversarial training mechanism and the training background image set, to obtain features for extracting the input background image, and combining the features to encode the background image into the latent space to form the encoder of the latent variable.

[0036] That is, the encoder formed by pre-training can effectively extract the content, structure, texture and other features of the original background image, and after encoding the background image and its features into the latent space, it can significantly reduce the amount of calculation and accelerate the digital camouflage generation process. The encoder used in this embodiment is Vector Quantization Generative Adversarial Network (VQGAN), which uses perceptual error (perceptual loss) instead of mean square error as the reconstruction error of VQGAN during training, and introduces an adversarial training mechanism. For example, the original background image is , the encoder is , then the encoded latent space representation The (hidden variables) are: .

[0037] Furthermore, the method further comprises: S8: construct the initial denoising diffusion model; S9: In the latent space, the initial denoising diffusion model is trained based on a training data set to form the denoising diffusion model, wherein the training data set includes a first latent variable set corresponding to the background image data set and a second latent variable set corresponding to the digital camouflage image data set.

[0038] Wherein, in the latent space, the training process of the initial denoising diffusion model includes a forward diffusion process; The forward diffusion process includes: S10: Hidden variables in the first hidden variable set obtained T times Add the target Gaussian noise to form the corresponding hidden variable Gaussian distribution, the latent variable The Gaussian distribution of is the standard Gaussian distribution.

[0039] The target Gaussian noise is subject to the mean ,variance The Gaussian distribution of is the identity matrix, is a preset value and changes with time step Linear change, the is the corresponding time step The hidden variable, when When it increases, the The Gaussian distribution of is close to the pure noise distribution when When The Gaussian distribution of is a standard Gaussian distribution. In this embodiment, each step of the forward diffusion process can be expressed as: .

[0040] In the latent space, the training process of the initial denoising diffusion model also includes a denoising inference process; The denoising inference process includes: S11: learning the inverse process distribution of the Gaussian distribution based on the target neural network, wherein the inverse process distribution is used to restore the latent variable according to the Gaussian distribution, and the target neural network is a neural network that uses an attention mechanism for learning; The denoising process is performed on the first latent variable in the latent space based on the denoising diffusion model to form a second latent variable, including: S12: predicting the Gaussian distribution of noise in each denoising process based on the denoising diffusion model; S13: After performing the T-times denoising based on the predicted Gaussian distribution of the noise, the second latent variable is formed.

[0041] For example, all time steps t The sum of the number is T, and the inverse process distribution is expressed as In this embodiment, the U-Net network model based on the attention mechanism is adopted to learn the inverse process distribution. The learned distribution is recorded as : Among them, the mean You can write: Gaussian noise for each step of the denoising process for neural network predictions.

[0042] You can write: .

[0043] In this embodiment, the loss function of the denoising diffusion model is: .

[0044] Based on the above, the process of denoising inference / training the first latent variable includes: 1) Input background image at each time step and , the neural network prediction output is Gaussian noise , substitute into the above mean formula to calculate the mean ; 2) Calculate the variance based on the above variance formula ; 3) According to the calculated mean and variance, we get , use the re-parameter technique to get the next background image .

[0045] Furthermore, the method of this embodiment also includes: S14: constructing the adversarial noise generator based on the convolutional layer, the residual block, the downsampling layer and the attention module; S15: In the latent space, the adversarial noise generator is trained so that it can perform noise modeling based on the input latent variables, determine the distribution space of the noise disturbance, and then generate adversarial noise corresponding to the latent variables.

[0046] When calculating and determining the loss function of the adversarial noise generator, the method used in this embodiment is as described above, treating the target attack network as a discriminator, judging whether the input disguised image is an adversarial sample or a real sample, and determining whether the disguise is successful based on the discrimination result; the discriminator can be any trained classification network. This method uses ResNet50, and the network parameters are no longer updated.

[0047] Specifically, Figure 3 As shown in the figure, the Segmentation Anything pre-trained model (SAM segmentation model) is used to segment the target image (aircraft or tank) to be camouflaged and generate a mask image. Then, the digital camouflage image C corresponding to the background image generated previously is compared with the adversarial noise Fill in the mask image In the example above, we get adversarial sample image A. We send adversarial sample image A to the trained ResNet50 network and get the true / false discrimination result. At this time, the system can compare the discrimination result with the real label and calculate the loss function, and then train again through the back propagation algorithm. The loss function is defined as follows: .

[0048] in, Represents the classification loss for a sample without adding noise , the output of the classifier is c, and the output of the corresponding last layer softmax is q. When adding adversarial noise After that, the output of the softmax layer is recorded as Ideally, the added noise Can change the output of the classifier, so The probability value of the cth class should decrease, and the corresponding loss function is: .

[0049] However, It only enables the generator to have a better ability to fool the target network, but due to the instability of generative adversarial network training, it may cause the training to fall into the local minimum, and the generated adversarial noise does not have generalization and universality. Therefore, this embodiment introduces a new loss function term to enable the generator to better model the noise and explore the distribution space of noise perturbations. For a small batch of generated adversarial perturbations, it is hoped that their feature vectors on the classifier can be separated as much as possible. Specifically, two adversarial perturbations are randomly selected from multiple adversarial perturbations generated from a batch of data samples to form a pair of adversarial perturbations. and , in order to meet the increase in the classifier network The distance between the layer output and the noise generator is defined as follows: Said B represents the adversarial perturbation batch size, n' is from [1, B ] is a random integer, and Respectively represent the first n data samples and n' The adversarial perturbation generated by data samples, represents the distance metric, which can be specifically the Euclidean distance, , Representative classification network Layer output function, Representative n data samples.

[0050] like Figure 4 As shown, another embodiment of the present invention also provides a confrontational digital camouflage generating device 100, including: An encoding module, used for encoding the target background image obtained by the encoder to form a first latent variable; A denoising module, configured to perform denoising processing on the first latent variable in a latent space according to a denoising diffusion model to form a second latent variable; A first processing module, configured to generate a digital camouflage image according to the decoder processing the second latent variable; A second processing module, configured to process the first latent variable according to an adversarial noise generator to generate an adversarial noise image; A generating module is used to combine the digital camouflage image with the adversarial noise image to generate an adversarial digital camouflage image corresponding to the target background image.

[0051] In an optional embodiment, the device further comprises: A first building block, for building an initial encoder; The first training module is used to use the perceptual error as the reconstruction error and pre-train the initial encoder based on the adversarial training mechanism to obtain the features for extracting the input background image, and encode the background image into the latent space to form the encoder of the latent variable in combination with the features.

[0052] In an optional embodiment, the device further comprises: The second building module is used to build an initial denoising diffusion model; The second training module is used to train the initial denoising diffusion model in the latent space based on the training data set to form the denoising diffusion model. The training data set includes a first latent variable set corresponding to the background image data set and a second latent variable set corresponding to the digital camouflage image data set.

[0053] In an optional embodiment, in the latent space, the training process of the initial denoising diffusion model includes a forward diffusion process; The forward diffusion process includes: The hidden variables in the first hidden variable set obtained T times Add the target Gaussian noise to form the corresponding hidden variable Gaussian distribution, the latent variable The Gaussian distribution of is the standard Gaussian distribution.

[0054] In an optional embodiment, the target Gaussian noise is subject to the mean ,variance The Gaussian distribution of is the identity matrix, is a preset value and changes with time step Linear change, the is the corresponding time step t The hidden variable, when When it increases, the The Gaussian distribution of is close to the pure noise distribution when When The Gaussian distribution of is the standard Gaussian distribution.

[0055] In an optional embodiment, in the latent space, the training process of the initial denoising diffusion model further includes a denoising inference process; The denoising inference process includes: Learning the inverse process distribution of the Gaussian distribution based on the target neural network, wherein the inverse process distribution is used to restore the latent variable according to the Gaussian distribution, and the target neural network is a neural network that uses an attention mechanism for learning; The denoising process is performed on the first latent variable in the latent space based on the denoising diffusion model to form a second latent variable, including: Predicting the Gaussian distribution of noise in each step of the denoising process based on the denoising diffusion model; After performing the T times of denoising based on the predicted Gaussian distribution of the noise, the second latent variable is formed.

[0056] In an optional embodiment, the device further comprises: A third building module is used to build the adversarial noise generator according to the convolution layer, the residual block, the downsampling layer and the attention module; The third training module is used to train the adversarial noise generator in the latent space so that it can perform noise modeling based on the input latent variables, determine the distribution space of the noise disturbance, and then generate adversarial noise corresponding to the latent variables.

[0057] In an optional embodiment, the loss function of the adversarial noise generator is: Said B represents the adversarial perturbation batch size, n' is from [1, B ] is a random integer, and Respectively represent the first n data samples and The adversarial perturbation generated by the data samples is represents the distance metric, , Representative classification network Layer output function, Representative n data samples.

[0058] Another embodiment of the present invention further provides an electronic device, including: at least one processor; and, a memory communicatively coupled to the at least one processor; The memory stores instructions executable by the at least one processor, and the instructions are configured to execute the method for generating a confrontational digital camouflage as described in any one of the embodiments above.

[0059] Another embodiment of the present invention further provides a storage medium, which includes a stored program, wherein when the program is running, a device including the storage medium is controlled to execute the confrontation digital camouflage generation method as described in any of the embodiments above.

[0060] The embodiment of the present invention also provides a computer program product, which is tangibly stored on a computer-readable medium and includes computer-readable instructions, and when the computer-executable instructions are executed, at least one processor executes the method for generating adversarial digital camouflage patterns such as in the above-described embodiment. It should be understood that each solution in this embodiment has the corresponding technical effects in the above-described method embodiment, which will not be repeated here.

[0061] It should be noted that the computer storage medium of the present application may be a computer-readable signal medium or a computer-readable storage medium or any combination of the above two. The computer-readable medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or device, or any combination of the above. More specific examples of computer-readable storage media may include, but are not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access storage medium (RAM), a read-only storage medium (ROM), an erasable programmable read-only storage medium (EPROM or flash memory), an optical fiber, a portable compact disk read-only storage medium (CD-ROM), an optical storage medium, a magnetic storage medium, or any suitable combination of the above. In the present application, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in combination with an instruction execution system, device or device. In the present application, a computer-readable signal medium may include a data signal propagated in a baseband or as part of a carrier wave, which carries a computer-readable program code. This propagated data signal may take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. A computer-readable signal medium may also be any computer-readable medium other than a computer-readable storage medium, which may send, propagate or transmit a program configured to be used by or in conjunction with an instruction execution system, apparatus or device. The program code contained on the computer-readable medium may be transmitted using any appropriate medium, including but not limited to: wireless, antenna, optical cable, RF, etc., or any suitable combination of the above.

[0062] A person skilled in the art should understand that the discussion of any of the above embodiments is merely illustrative and is not intended to imply that the scope of protection of the present application is limited to these examples. In line with the concept of the present application, the technical features in the above embodiments or different embodiments may be combined, the steps may be implemented in any order, and there are many other variations of different aspects of one or more embodiments of the present application as described above, which are not provided in detail for the sake of simplicity.

[0063] One or more embodiments of the present application are intended to cover all such substitutions, modifications and variations that fall within the broad scope of the present application. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of one or more embodiments of the present application should be included in the protection scope of the present application.

Claims

1. A method for generating adversarial digital camouflage, characterized in that: include: Encoding the obtained target background image based on the encoder to form a first latent variable; Performing denoising processing on the first latent variable in latent space based on a denoising diffusion model to form a second latent variable; Processing the second latent variable based on the decoder to generate a digital camouflage image; Processing the first latent variable based on the adversarial noise generator to generate an adversarial noise image; Based on the combination of the digital camouflage image and the adversarial noise image, an adversarial digital camouflage image corresponding to the target background image is obtained.

2. The method for generating adversarial digital camouflage according to claim 1, characterized in that: The method further comprises: Build the initial encoder; The perceptual error is used as the reconstruction error, and the initial encoder is pre-trained based on the adversarial training mechanism and the training background image set to obtain the features for extracting the input background image, and the background image is encoded into the latent space to form the encoder of the latent variable in combination with the features.

3. The method for generating adversarial digital camouflage according to claim 1, characterized in that: The method further comprises: Construct an initial denoising diffusion model; In the latent space, the initial denoising diffusion model is trained based on a training data set to form the denoising diffusion model, wherein the training data set includes a first latent variable set corresponding to a background image data set and a second latent variable set corresponding to a digital camouflage image data set.

4. The method for generating adversarial digital camouflage according to claim 3, characterized in that: In the latent space, the training process of the initial denoising diffusion model includes a forward diffusion process; The forward diffusion process includes: The hidden variables in the first hidden variable set obtained T times Add the target Gaussian noise to form the corresponding hidden variable Gaussian distribution, the latent variable The Gaussian distribution of is the standard Gaussian distribution.

5. The method for generating confrontational digital camouflage according to claim 4, characterized in that: The target Gaussian noise is subject to the mean ,variance Gaussian distribution, the is the identity matrix, is a preset value and changes with time step Linear change, the is the corresponding time step The hidden variable, when When it increases, the The Gaussian distribution of is close to the pure noise distribution when When The Gaussian distribution of is the standard Gaussian distribution.

6. The method for generating adversarial digital camouflage according to claim 4, characterized in that: In the latent space, the training process of the initial denoising diffusion model also includes a denoising inference process; The denoising inference process includes: Learning the inverse process distribution of the Gaussian distribution based on the target neural network, wherein the inverse process distribution is used to restore the latent variable according to the Gaussian distribution, and the target neural network is a neural network that uses an attention mechanism for learning; The denoising process is performed on the first latent variable in the latent space based on the denoising diffusion model to form a second latent variable, including: Predicting the Gaussian distribution of noise in each step of the denoising process based on the denoising diffusion model; After performing the T times of denoising based on the predicted Gaussian distribution of the noise, the second latent variable is formed.

7. The method for generating adversarial digital camouflage according to claim 1, characterized in that: The method further comprises: Constructing the adversarial noise generator based on convolutional layers, residual blocks, downsampling layers, and attention modules; In the latent space, the adversarial noise generator is trained so that it can perform noise modeling based on the input latent variables, determine the distribution space of the noise disturbance, and then generate adversarial noise corresponding to the latent variables.

8. The method for generating confrontational digital camouflage according to claim 7, characterized in that: The loss function of the adversarial noise generator is: Said B represents the adversarial perturbation batch size, n' is from [1, B ] is a random integer, and Respectively represent the first n data samples and The adversarial perturbation generated by data samples, represents the distance metric, , Representative classification network Layer output function, Representative n data samples.

9. A device for generating confrontational digital camouflage, characterized in that: include: An encoding module, used for encoding the target background image obtained by the encoder to form a first latent variable; A denoising module, configured to perform denoising processing on the first latent variable in a latent space according to a denoising diffusion model to form a second latent variable; A first processing module, configured to generate a digital camouflage image according to the decoder processing the second latent variable; A second processing module, configured to process the first latent variable according to an adversarial noise generator to generate an adversarial noise image; A generating module is used to combine the digital camouflage image with the adversarial noise image to generate an adversarial digital camouflage image corresponding to the target background image.

10. An electronic device, characterized in that: include: at least one processor; as well as, a memory communicatively coupled to the at least one processor; The memory stores instructions executable by the at least one processor, and the instructions are configured to execute the confrontational digital camouflage generation method as described in any one of claims 1-8.

Images