Methods, systems, and media for cross-participant image reconstruction in the context of federated learning

Abstract

This invention discloses a cross-participant image reconstruction method, system, and medium under a federated learning framework. The method includes: inferring the label of a target image based on the target gradient uploaded by the client; constructing a conditional denoising diffusion implicit model, using latent variables and labels as input to generate a matching image; determining a loss function, which includes the loss of the matching gradient relative to the target gradient and a regularization loss for generating the matching image. The matching gradient is obtained by feeding the matching image and labels into the federated learning model for one round of learning; initializing latent variables, sequentially optimizing the latent variables for the first n time steps according to the loss function, and then using the conditional denoising diffusion implicit model to obtain the reconstructed image based on the optimized latent variables and the inferred labels. By sequentially optimizing the first n latent variables of the conditional denoising diffusion implicit model through gradient matching loss and regularization loss, the optimization efficiency is improved, and the reconstructed image is more faithful to the target image.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method, system and medium for cross-participant image reconstruction in the context of federated learning. Background Technology

[0002] Federated learning is a distributed learning method aimed at protecting client privacy. It allows multiple clients to collaboratively train a global neural network without sharing their private data, only needing to submit locally updated network gradients to the server. DLG (Deep Gradient Leakage) challenges the privacy protection of federated learning by optimizing random noise through gradient matching loss, ultimately reconstructing the client's image.

[0003] To further protect user privacy, several defense mechanisms have been proposed to prevent privacy leaks through gradient information degradation, such as using additive noise (differential privacy) or gradient compression before sharing with the server. Numerous experiments have demonstrated that DLGs cannot reconstruct meaningful images under these defense mechanisms. However, the privacy protection of federated learning under these defense mechanisms is not absolutely secure. Existing techniques utilize the latent space of a generative adversarial network (GAN) learned from a public image dataset as a prior to compensate for information loss during gradient degradation to reconstruct images. However, GANs are not only costly to train but also extremely difficult to train, often facing problems such as model collapse and non-convergence. Moreover, the distribution of images generated by GANs is limited by the distribution of the latent space, sometimes failing to faithfully reconstruct the image, resulting in issues such as image orientation changes and loss of detail. Based on this, this invention proposes a cross-participant image reconstruction scheme under federated learning, improving the fidelity of the reconstructed image to the target image. This scheme can be used to evaluate the privacy protection performance of federated learning under defense mechanisms, thereby guiding improvements to federated learning defense mechanisms. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a cross-participant image reconstruction method, system, and medium in the context of federated learning. By using gradient matching loss and regularization loss, the first n latent variables are optimized sequentially, enabling these latent variables to reconstruct the target image through a denoising diffusion implicit model, thereby improving the fidelity of the reconstructed image to the target image.

[0005] Firstly, a cross-participant image reconstruction method in the context of federated learning is provided, including the following steps:

[0006] S1: Sample tensors of the same shape as the input of the federated learning model from the standard normal distribution as initial latent variables;

[0007] S2: Obtain the target gradient uploaded by the client, and infer the label of the target image based on the target gradient;

[0008] S3: Use a conditional denoising diffusion implicit model to generate matching images with latent variables and inferred labels as input;

[0009] S4: Input the matching image and the inferred label into the federated learning model. After one round of forward and backward propagation, the matching gradient of the federated learning model is obtained.

[0010] S5: Calculate the loss function, which includes the loss of the matching gradient relative to the target gradient and the regularization loss of the generated matching image;

[0011] S6: Optimize latent variables based on the loss function;

[0012] S7: Repeat steps S3 to S6 until the loss function converges to its minimum or reaches the predetermined number of rounds, and obtain the latent variable z at time step t. t ;

[0013] S8: Optimize the object from the hidden variable z at time step t. t Update the hidden variable z to time step t-1 t-1 , z t-1 The initial value is given by z t Obtained through one-step sampling using a conditional denoising diffusion implicit model;

[0014] S9: Then repeat steps S3 to S8 until the optimization of the latent variables in the first n time steps is completed;

[0015] S10: Based on the latent variables and inferred labels from the previous n time steps, the reconstructed image z0 is obtained using a conditional denoising diffusion implicit model.

[0016] Furthermore, the inference of the target image label based on the target gradient specifically includes:

[0017] Analyze the gradients of each weight in the last fully connected classification layer of the target image. If the gradient of the i-th weight in the fully connected classification layer is negative, then the label c = i of the target image.

[0018] Furthermore, the sampling process of the conditional denoising diffusion implicit model is represented as follows:

[0019]

[0020] In the formula, x t and x t-1 Represent the images at time step t and time step t-1, respectively. θ It is a neural network model. Represents the neural network model ∈ θImage x at label c, time step t, and at that time step t Given the input, the output is the same as the image x. t Noise with the same shape; α t and σ t They are all a set of constants related to t, ∈ t This represents random noise.

[0021] Furthermore, the loss function L(z) is expressed as follows:

[0022]

[0023] In the formula, This indicates matching the gradient g with the target gradient. The L2 norm between the two is the loss of the matching gradient relative to the target gradient; R(z) represents the total variation of the generated matching image, which is the regularization loss of the generated matching image; α represents the weight of the regularization loss of the generated matching image.

[0024] Furthermore, the matching gradient g is represented as follows:

[0025]

[0026] In the formula, G represents the conditional denoising diffusion implicit model, z represents the latent variable to be optimized, and c represents the inferred label. This represents the gradient of the federated learning model;

[0027] The total variation R(z) of the generated matching image is expressed as follows:

[0028] R(z)=TV(G(z))

[0029] In the formula, TV(G(z)) represents the total variation of G(z), which is the sum of the absolute values ​​of the differences between adjacent pixel values ​​of G(z) in both the width and height directions.

[0030] Furthermore, the AdamW optimizer is used when optimizing latent variables based on the loss function.

[0031] Furthermore, it also includes:

[0032] Calculate the similarity value between the reconstructed image and the target image, and evaluate the privacy protection performance of federated learning based on the similarity value.

[0033] Furthermore, the similarity value between the reconstructed image and the target image is calculated using one of mean squared error, peak signal-to-noise ratio, and LPIPS.

[0034] Secondly, a cross-participant image reconstruction system in the context of federated learning is provided, including:

[0035] A memory on which computer programs or instructions are stored;

[0036] A processor is used to load and execute the computer program or instructions to implement the cross-participant image reconstruction method in the context of federated learning as described above.

[0037] Thirdly, a computer-readable storage medium is provided having a computer program or instructions stored thereon, which, when executed by a processor, implement the cross-participant image reconstruction method in the context of federated learning as described above.

[0038] This invention proposes a cross-participant image reconstruction method, system, and medium in the context of federated learning, which has the following beneficial effects:

[0039] (1) Compared with the generative adversarial network in the existing technology, the conditional denoising diffusion implicit model has better generalization ability and detail preservation in reconstructing images than the generative adversarial network, and the reconstructed image is more faithful to the target image.

[0040] (2) In this invention, the first n latent variables of the conditional denoising diffusion implicit model are regarded as the latent space of the conditional denoising diffusion implicit model. These latent variables are optimized sequentially through gradient matching loss and regularization loss. Compared with the existing optimization process, the optimization efficiency is improved and higher quality images can be reconstructed under the same number of optimization rounds.

[0041] (3) Under the same dataset conditions, the training cost and training difficulty of the conditional denoising diffusion implicit model are much lower than those of the generative adversarial network.

[0042] (4) Based on this invention, an image reconstruction method is proposed, which provides a new perspective for evaluating the privacy and security issues of federated learning, thereby promoting the development of federated learning defense mechanisms. Attached Figure Description

[0043] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0044] Figure 1 This is a flowchart of a cross-participant image reconstruction method under the context of federated learning provided in this embodiment of the invention;

[0045] Figure 2 This is a framework diagram of a cross-participant image reconstruction method under the context of federated learning provided in this embodiment of the invention. Detailed Implementation

[0046] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be described in detail below. Obviously, the described embodiments are merely some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other implementation methods obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0047] The background of this invention is federated learning networks for processing image multi-classification tasks. The server is honest and curious, directly receiving gradient updates submitted by the client and attempting, and only attempting, to reconstruct the images used by the client to train the model. The gradient updates submitted by the client are generated from a single image through one round of training, and differential privacy is applied to the gradients. The technical solution of this invention will be specifically described below with reference to specific embodiments.

[0048] like Figure 1 , Figure 2 As shown, this embodiment of the invention provides a cross-participant image reconstruction method in the context of federated learning, including the following steps:

[0049] S1: Sample tensors of the same shape as the input of the federated learning model from the standard normal distribution as initial latent variables; therefore, the mean of the initial latent variables is 0 and the standard deviation is 1.

[0050] S2: Obtain the target gradient uploaded by the client, and infer the label of the target image based on the target gradient. The specific inference process is as follows: Analyze the gradient of each weight of the last fully connected classification layer of the target gradient. If the gradient of the i-th weight of the fully connected classification layer is negative, then the label of the target image is c = i.

[0051] In image classification tasks, images and their labels are used together as input to the model during training. In the DLG method, the labels are randomly generated and optimized along with random noise. Studies have shown that random labels make the optimization problem difficult to converge, so a method to infer the training image labels from the gradients has been proposed.

[0052] Specifically, for a federated learning (FL) model performing a classification task on m classes, the gradient of the i-th (i = 1, 2, ..., m) weight of the final fully connected (FC) classification layer (denoted as...) () is defined by the following formula:

[0053]

[0054] Among them, z i It is the i-th output of the FC layer, f θ This is an FL model, where c is the data label and L is the FL model loss function. It's worth noting the second term on the right-hand side of the equation. The result is the output after activation of the previous layer. If the previous layer uses activation functions such as ReLU or sigmoid, the output is always non-negative. For a neural network using the cross-entropy loss function and training images with one-hot encoded labels, assuming softmax is applied to the last layer, the first term on the right-hand side of the equation is negative if and only if i = c. Therefore, it is possible to identify... The positive and negative signs are used to construct the labels for the target image.

[0055] S3: Use a conditional denoising diffusion implicit model to generate matching images with latent variables and inferred labels as input.

[0056] The conditional denoising diffusion implicit model is an image generation model proposed based on the denoising diffusion implicit model, while the denoising diffusion implicit model is an image generation model proposed based on the denoising diffusion probability model. In order to facilitate understanding of the technical solution of this invention, the denoising diffusion probability model will be introduced first.

[0057] The denoising diffusion probability model starts with noise from a standard normal distribution and generates an image through a series of denoising steps. This process from noise to image is also known as sampling, and its algorithm is described below:

[0058]

[0059] The algorithm extracts pure noise x t Begin by gradually subtracting ∈ θ (x t ,t), and the final result x0 is the generated image. Where ∈ θ It is a neural network model, with time step t and x at that time step. t As input, output is equal to x t Noise with the same shape, this noise means from x t-1 To x t The noise that should be added, where ε represents random noise.

[0060] The process of training the denoising diffusion probability model is actually training ∈ θ The process, described by the algorithm, is as follows:

[0061]

[0062] The algorithm continuously adds noise to the original image x0 until it becomes pure noise x0. T And each time noise is added, use ∈ θ (x t The noise is predicted by t, and the parameter θ is optimized using gradient descent to minimize the error between the predicted noise and the actual noise.

[0063] Therefore, during sampling, ∈θ (x t ,t) can accurately predict from x t-1 To x t The noise that should be added, thus by x t Calculate x t-1 Among them, α t , and σ t These are all constants related to t; ∈ represents the noise added at the current time step t.

[0064] The denoising diffusion implicit model is an improved version of the denoising diffusion probabilistic model. The main improvements are twofold: first, it can control the uncertainty of the generated image; second, it can skip steps during sampling, accelerating the sampling process. The sampling process of the denoising diffusion implicit model is as follows:

[0065]

[0066] In the formula, ∈ t Indicates random noise; Represents the neural network model ∈ θ Image x at time step t. t Given the input, the output is the same as the image x. t Noise with the same shape. Unlike the denoising diffusion probabilistic model, the denoising diffusion implicit model controls the uncertainty of the generated image through the last term of the above equation. When σ t When x is 0, for a given set of random noise (latent variable), a unique image will be generated. This determinism is precisely what this invention requires; it aims to generate the target image deterministically through the optimized latent variable. Furthermore, x in the above equation... t-1 It can be adjusted to x t-n This means skipping some sampling steps. Experiments show that skipping steps can speed up the sampling process by 10 to 50 times without losing image quality.

[0067] By embedding conditional information into the network structure of the denoising diffusion implicit model, an unconditional denoising diffusion implicit model can be transformed into a conditional denoising diffusion implicit model. When training the conditional denoising diffusion implicit model, based on AlgorithmTraining, a conditional label c is provided, allowing... Become During sampling, condition c is input along with ∈ θ The sampling process of the conditional denoising diffusion implicit model is represented as follows:

[0068]

[0069] In the formula, x t and x t-1Represent the images at time step t and time step t-1, respectively. θ It is a neural network model. Represents the neural network model ∈ θ Image x at label c, time step t, and at that time step t Given the input, the output is the same as the image x. t Noise of the same shape. This allows the conditional denoising diffusion implicit model to generate images with specified label categories.

[0070] S4: The matched image and the inferred label are fed into the federated learning model. After one round of forward and backward propagation, the matching gradient g of the federated learning model is obtained. Specifically, the matching gradient g is represented as follows:

[0071]

[0072] In the formula, G represents the conditional denoising diffusion implicit model, z represents the latent variable to be optimized, and c represents the inferred label. This represents the gradient of the federated learning model.

[0073] S5: Calculate the loss function, which includes the loss of the matching gradient relative to the target gradient and the regularization loss of the generated matching image.

[0074] Specifically, the loss function L(z) is expressed as follows:

[0075]

[0076] In the formula, This indicates matching the gradient g with the target gradient. The L2 norm between the two gradients is the loss of the matching gradient relative to the target gradient. The goal is to make the two gradients gradually approach each other, which means that the generated image gradually approaches the target image. R(z) represents the total variation of the generated matching image, which is the regularization loss of the generated matching image. It measures the realism of the generated image and makes the generated image natural and realistic. α represents the weight of the regularization loss of the generated matching image.

[0077] The total variation R(z) of the generated matching image is expressed as follows:

[0078] R(z)=TV(G(z))

[0079] In the formula, TV(G(z)) represents the total variation of G(z), which is the sum of the absolute values ​​of the differences between adjacent pixel values ​​of G(z) in both the width and height directions.

[0080] S6: Optimize the latent variables based on the loss function so that the value of the loss function approaches the minimum.

[0081] This process can be modeled as an optimization problem:

[0082]

[0083] That is, searching in the latent space for the latent variable z that minimizes L(z). * .

[0084] In this embodiment, the AdamW optimizer is used when optimizing latent variables based on the loss function.

[0085] The Adam series optimizers have the following features:

[0086] (1) Adaptive learning rate: The Adam optimizer uses the first and second moments of the gradient to adjust the learning rate of each weight individually. This adaptive learning rate method can lead to more efficient updates and faster convergence.

[0087] (2) Efficient gradient descent: Unlike SGD, which requires updating all parameters by the same amount, Adam achieves efficient gradient descent by making smaller updates to frequently updated parameters and larger updates to infrequently updated parameters.

[0088] (3) Less dependent on the initial learning rate: The Adam optimizer is less sensitive to the initial learning rate, thus reducing the time spent on hyperparameter tuning.

[0089] (4) Corrected Weight Decay: The main difference between Adam and AdamW lies in how they handle weight decay (a form of regularization). In Adam, weight decay is applied before the gradient is computed, which leads to suboptimal results. AdamW applies weight decay only after the gradient is computed, which is a more correct implementation and thus further improves performance compared to Adam.

[0090] (5) Improved generalization: By correctly applying weight decay, AdamW often performs better than Adam in generalization, especially on larger models or datasets. Better convergence: Empirical results show that AdamW converges faster than Adam and yields better solutions.

[0091] S7: Repeat steps S3 to S6 until the loss function converges to its minimum or reaches the predetermined number of rounds, and obtain the latent variable z at time step t. t .

[0092] S8: Optimize the object from the hidden variable z at time step t. t Update the hidden variable z to time step t-1 t-1 , z t-1 The initial value is the optimized z t It was obtained through sampling using a conditional denoising diffusion implicit model.

[0093] S9: Then repeat steps S3 to S8 until the hidden variable z of the first n time steps t is completed. T ,z T-1 ,…,z T-n Optimization.

[0094] S10: Based on the latent variables and inferred labels from the previous n time steps t, the reconstructed image z0 is obtained using a conditional denoising diffusion implicit model.

[0095] Unlike existing methods that only optimize the initial latent variables, this approach does not optimize only the initial latent variables. Optimizing only the initial latent variables would gradually weaken their impact on the generated image as the optimization process progresses, affecting the quality of the final reconstructed image and resulting in low optimization efficiency. Therefore, this method sequentially optimizes the first n latent variables of the conditional denoising diffusion implicit model, preserving the impact of changes in the latent variables on the generated image.

[0096] The cross-participant image reconstruction method provided in the above embodiments under the background of federated learning has the following advantages:

[0097] (1) Compared with the generative adversarial network in the existing technology, the conditional denoising diffusion implicit model has better generalization ability and detail preservation in reconstructing images than the generative adversarial network, and the reconstructed image is more faithful to the target image.

[0098] (2) In this invention, the first n latent variables of the conditional denoising diffusion implicit model are regarded as the latent space of the conditional denoising diffusion implicit model. These latent variables are optimized sequentially through gradient matching loss and regularization loss. Compared with the existing optimization process, the optimization efficiency is improved and higher quality images can be reconstructed under the same number of optimization rounds.

[0099] (3) Under the same dataset conditions, the training cost and training difficulty of the conditional denoising diffusion implicit model are much lower than those of the generative adversarial network.

[0100] (4) The present invention uses the gradient-based AdamW optimizer, which makes the optimization problem easier to solve.

[0101] In some embodiments, the method further includes:

[0102] S11: Calculate the similarity value between the reconstructed image and the target image, and evaluate the privacy protection performance of federated learning based on the similarity value. In practice, the similarity value range can be discretized into multiple intervals, each interval corresponding to a privacy protection performance level for evaluating federated learning. Thus, the privacy protection performance level of federated learning can be obtained based on the calculated similarity value.

[0103] In practice, the similarity value between the reconstructed image and the target image is calculated using one of the following: mean squared error (MSE), peak signal-to-noise ratio (PSNR), and LPIPS (Learned Perceptual Image Patch Similarity).

[0104] When using mean squared error to calculate similarity, the pixel-level mean squared error between the reconstructed image and the target image is calculated. The closer the similarity value is to 0, the more similar the two images are. In other words, the closer the similarity value is to 0, the worse the privacy protection performance of federated learning is.

[0105] When using peak signal-to-noise ratio (PSNR) to calculate similarity, it is the ratio of the square of the maximum possible value of the image pixels to the mean square error between the reconstructed image and the target image. The higher the similarity value, the more similar the two images are. In other words, the higher the similarity value, the worse the privacy protection performance of federated learning.

[0106] When using LPIPS to calculate similarity, the VGG model is used to extract high-level features of the image and evaluate the perceptual difference between the reconstructed image and the target image. The closer the similarity value is to 0, the more similar the two images are. In other words, the closer the similarity value is to 0, the worse the privacy protection performance of federated learning is.

[0107] Based on this embodiment, an image reconstruction method is proposed, providing a new perspective for evaluating the privacy and security issues of federated learning, thereby promoting the development of federated learning defense mechanisms.

[0108] This invention also provides a cross-participant image reconstruction system in the context of federated learning, comprising:

[0109] A memory on which computer programs or instructions are stored;

[0110] A processor is used to load and execute the computer program or instructions to implement the cross-participant image reconstruction method in the context of federated learning as described above.

[0111] This invention also provides a computer-readable storage medium storing a computer program or instructions thereon, which, when executed by a processor, implements the cross-participant image reconstruction method in the context of federated learning as described above.

[0112] It is understood that the same or similar parts in the above embodiments can be referred to each other, and the contents not described in detail in some embodiments can be referred to the same or similar contents in other embodiments.

[0113] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0114] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0115] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0116] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0117] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A cross-participant image reconstruction method in the context of federated learning, characterized in that, Includes the following steps: S1: Sample tensors of the same shape as the input of the federated learning model from the standard normal distribution as initial latent variables; S2: Obtain the target gradient uploaded by the client, and infer the label of the target image based on the target gradient; S3: Use a conditional denoising diffusion implicit model to generate matching images with latent variables and inferred labels as input; S4: Input the matching image and the inferred label into the federated learning model. After one round of forward and backward propagation, the matching gradient of the federated learning model is obtained. S5: Calculate the loss function, which includes the loss of the matching gradient relative to the target gradient and the regularization loss of the generated matching image; S6: Optimize latent variables based on the loss function; S7: Repeat steps S3 to S6 until the loss function converges to its minimum or reaches the predetermined number of rounds, and obtain the latent variable z at time step t. t ; S8: Optimize the object from the hidden variable z at time step t. t Update the hidden variable z to time step t-1 t-1 , z t-1 The initial value is given by z t Obtained through one-step sampling using a conditional denoising diffusion implicit model; S9: Then repeat steps S3 to S8 until the optimization of the latent variables in the first n time steps is completed; S10: Based on the latent variables and inferred labels from the previous n time steps, the reconstructed image is obtained using a conditional denoising diffusion implicit model.

2. The cross-participant image reconstruction method in the context of federated learning according to claim 1, characterized in that, The label inferred from the target gradient specifically includes: Analyze the gradients of each weight in the last fully connected classification layer of the target image. If the gradient of the i-th weight in the fully connected classification layer is negative, then the label c = i of the target image.

3. The cross-participant image reconstruction method in the context of federated learning according to claim 1, characterized in that, The sampling process of the conditional denoising diffusion implicit model is represented as follows: In the formula, x t and x t-1 Represent the images at time step t and time step t-1, respectively. θ It is a neural network model. Represents the neural network model ∈ θ Image x at label c, time step t, and at that time step t Given the input, the output is the same as the image x. t Noise with the same shape; α t and σ t They are all a set of constants related to t, ∈ t This represents random noise.

4. The cross-participant image reconstruction method in the context of federated learning according to claim 1, characterized in that, The loss function L(z) is expressed as follows: In the formula, This indicates matching the gradient g with the target gradient. The L2 norm between the two is the loss of the matching gradient relative to the target gradient; R(z) represents the total variation of the generated matching image, which is the regularization loss of the generated matching image; α represents the weight of the regularization loss of the generated matching image.

5. The cross-participant image reconstruction method in the context of federated learning according to claim 4, characterized in that, The matching gradient g is represented as follows: In the formula, G represents the conditional denoising diffusion implicit model, z represents the latent variable to be optimized, and c represents the inferred label. This represents the gradient of the federated learning model; The total variation R(z) of the generated matching image is expressed as follows: R(z)=TV(G(z)) In the formula, TV(G(z)) represents the total variation of G(z), which is the sum of the absolute values ​​of the differences between adjacent pixel values ​​of G(z) in both the width and height directions.

6. The cross-participant image reconstruction method in the context of federated learning according to claim 1, characterized in that, The AdamW optimizer is used when optimizing latent variables based on the loss function.

7. The cross-participant image reconstruction method in the context of federated learning according to claim 1, characterized in that, Also includes: Calculate the similarity value between the reconstructed image and the target image, and evaluate the privacy protection performance of federated learning based on the similarity value.

8. The cross-participant image reconstruction method in the context of federated learning according to claim 7, characterized in that, The similarity value between the reconstructed image and the target image is calculated using one of the following: mean squared error, peak signal-to-noise ratio, and LPIPS.

9. A cross-participant image reconstruction system in the context of federated learning, characterized in that, include: A memory on which computer programs or instructions are stored; A processor for loading and executing the computer program or instructions to implement the cross-participant image reconstruction method in a federated learning context as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program or instructions stored thereon, characterized in that, When the computer program or instructions are executed by a processor, they implement the cross-participant image reconstruction method in the context of federated learning as described in any one of claims 1 to 8.

Images