An autoregressive image generation training data detection method and device

Abstract

This invention discloses a method and device for detecting training data in autoregressive image generation, belonging to the field of training data detection technology. It addresses the current technical problems of lacking effective training data detection schemes for autoregressive image generation models, the inability of detection schemes for diffusion models to adapt due to model differences, and the poor performance of detection schemes for autoregressive text generation due to the strong generalization of autoregressive image generation models. The method includes: controlling the output prediction probabilities of a target image token by the autoregressive image generation model using several model prediction probabilities; performing a weighted summation of the prediction probabilities based on preset hyperparameters to obtain an image score for the target image token; and determining whether the target image token is a member sample participating in the training of the autoregressive image generation model based on the image score and a preset judgment threshold. This significantly reduces computational overhead and time costs, enabling rapid sample detection.

Description

Technical Field

[0001] This invention relates to the field of training data detection technology, and in particular to a method and device for detecting training data generated from autoregressive images. Background Technology

[0002] With the continuous evolution of artificial intelligence technology in the field of image generation, image generation paradigms are constantly being updated. Autoregressive image generation technology, due to its outstanding advantages of fast generation speed and high generation quality, has become an important research direction in the current field of image generation. At the same time, the demand for privacy and copyright protection of training data is becoming increasingly urgent, and the importance of training data detection technology (used to determine whether target samples have been illegally used for model training) is becoming increasingly prominent. However, existing technologies still have obvious shortcomings and cannot meet the detection requirements in autoregressive image generation scenarios, as follows: The core principle of autoregressive image generation technology is to decompose an image into multiple image tokens and arrange these tokens into an ordered sequence. The preceding image tokens are used to predict the following image tokens. In the inference stage, the complete image token sequence is predicted step by step and then decoded by a decoder to generate the final image.

[0003] Training data detection techniques are used to determine whether a sample has been used for model training, which is crucial for maintaining privacy and copyright security. Currently, related solutions mainly focus on the following two types of models, but neither can be adapted to autoregressive image generation models: 1. Detection schemes related to diffusion models: These include the Loss scheme, PIA scheme, SecMI scheme, and SSIM score-based schemes. The Loss scheme uses the training loss function as the metric; a lower loss indicates a member variable, while a higher loss indicates a non-member variable. The PIA scheme calculates the loss using prediction noise from a clean image, rather than random noise. The SecMI scheme uses DDIM inversion technology to reverse the image generation process and calculate the prior prediction loss. Another scheme also uses DDIM inversion technology to calculate the SSIM score of the reconstructed image and the original image to distinguish between member and non-member samples. Although research on detection of training data for diffusion models is relatively mature, due to the significant differences between diffusion models and autoregressive image generation models, the above detection schemes for diffusion models cannot be directly applied to autoregressive image generation models.

[0004] 2. Autoregressive text generation related detection schemes: These include Min-k% technique, Min-k%++ technique, and DC-PDD technique. The Min-k% technique assumes that member samples have fewer low-log probability tokens, and achieves detection by selecting the k% of tokens with the lowest log probability and performing a weighted average. The Min-k%++ technique improves upon the Min-k% technique by normalizing the score of each token before weighted averaging, considering the different prediction difficulty of each token. The DC-PDD technique utilizes the difference between the token probability distribution in the text and the token frequency distribution in the reference corpus to identify training data. However, due to the strong generalization ability of autoregressive image generation models, these detection methods for autoregressive text generation cannot be effective on autoregressive image generation models.

[0005] In summary, existing training data detection schemes either cannot be applied to autoregressive image generation models or perform poorly on such models, seriously hindering privacy and copyright protection in the field of autoregressive image generation. Therefore, there is an urgent need for a training data detection scheme for autoregressive image generation models. Summary of the Invention

[0006] This invention provides a method and device for detecting training data for autoregressive image generation, which addresses the following technical problems: Currently, there is a lack of effective training data detection schemes for autoregressive image generation models; detection schemes for diffusion models cannot be adapted due to model differences; and detection schemes for autoregressive text generation are ineffective due to the strong generalization of autoregressive image generation models.

[0007] The embodiments of the present invention adopt the following technical solutions: On one hand, embodiments of the present invention provide a method for detecting training data for autoregressive image generation, the method comprising: controlling an autoregressive image generation model to output several model prediction probabilities of a target image token; Based on preset hyperparameters, the prediction probabilities of the several models are weighted and summed to obtain the image score of the target image token; Based on the image score and a preset judgment threshold, it is determined whether the target image token is a member sample participating in the training of the autoregressive image generation model.

[0008] In one feasible implementation, the autoregressive image generation model is controlled to output several model prediction probabilities for the target image token, specifically including: The conditional input type of each model block in the autoregressive image generation model is set to obtain several combinations of conditional inputs; Based on each combination of conditional inputs, the autoregressive image generation model is controlled to predict the target image token, resulting in several corresponding model prediction probabilities.

[0009] In one feasible implementation, the autoregressive image generation model is controlled to predict the target image token based on each combination of conditional inputs, resulting in several corresponding model prediction probabilities, specifically including: Set all model blocks in the autoregressive image generation model to accept conditional input to obtain the first autoregressive image generation model. The target image token is predicted using the first autoregressive image generation model to obtain the first model prediction probability.

[0010] In one feasible implementation, the autoregressive image generation model is controlled to predict the target image token based on each combination of conditional inputs, resulting in several corresponding model prediction probabilities, specifically including: The conditional input type of all model blocks in the autoregressive image generation model is set to accept unconditional input to obtain the second autoregressive image generation model. The target image token is predicted using the second autoregressive image generation model to obtain the second model prediction probability.

[0011] In one feasible implementation, the autoregressive image generation model is controlled to predict the target image token based on each combination of conditional inputs, resulting in several corresponding model prediction probabilities, specifically including: The conditional input type of all model blocks in the autoregressive image generation model is set to accept conditional input for the first half of the model blocks and accept unconditional input for the second half of the model blocks, thus obtaining the third autoregressive image generation model. The target image token is predicted using the third autoregressive image generation model to obtain the predicted probability of the third model.

[0012] In one feasible implementation, the autoregressive image generation model is controlled to predict the target image token based on each combination of conditional inputs, resulting in several corresponding model prediction probabilities, specifically including: The conditional input type of all model blocks in the autoregressive image generation model is set to accept unconditional input for the first half of the model blocks and conditional input for the second half of the model blocks, thus obtaining the fourth autoregressive image generation model. The target image token is predicted using the fourth autoregressive image generation model to obtain the fourth model prediction probability.

[0013] In one feasible implementation, the image score of the target image token is obtained by weighted summation of the prediction probabilities of the several models based on preset hyperparameters, specifically including: Based on S = P1-P2+k (P3-P4), the prediction probabilities of the several models are weighted and summed to obtain the image score S of the target image token; Wherein, P1 is the prediction probability of the first model, P2 is the prediction probability of the second model, P3 is the prediction probability of the third model, P4 is the prediction probability of the fourth model, and k is the preset hyperparameter.

[0014] In one feasible implementation, the method further includes: After all target image tokens have had their image scores calculated, low-scoring image tokens with image scores below a first preset score threshold are selected, and the average image score of these low-scoring image tokens is calculated as the unified score of these low-scoring image tokens.

[0015] In one feasible implementation, determining whether the target image token is a member sample participating in the training of the autoregressive image generation model is based on the image score and a preset judgment threshold, specifically including: Remove target image tokens whose image scores are higher than the second preset score threshold; If the image score is higher than the preset judgment threshold among the remaining target image tokens, then the target image token is determined to be a member sample participating in the training of the autoregressive image generation model. If the image score is not higher than the preset judgment threshold, the target image token is determined to be a non-member sample and has not participated in the training of the autoregressive image generation model.

[0016] On the other hand, embodiments of the present invention also provide an autoregressive image generation training data detection device, the device comprising: At least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to execute the autoregressive image generation training data detection method.

[0017] Compared with the prior art, the autoregressive image generation training data detection method and device provided in this embodiment of the invention have the following beneficial effects: 1. Highly targeted and adaptable to autoregressive image generation models: Based on the core characteristics of autoregressive image generation models, namely "image token prediction + multi-scale encoding", this invention designs a prediction probability acquisition method of "conditional / unconditional input control model block", which is specifically adapted to autoregressive image generation models such as VAR, VAR-CLIP, and M-VAR, filling the technical gap in training data detection in this field.

[0018] 2. High detection efficiency and low overhead: This invention only requires four forward inference steps to obtain four prediction probabilities, without the need for additional training of the feature extraction model, which greatly reduces computational overhead and time cost, and can quickly achieve sample detection.

[0019] 3. Excellent detection accuracy: Experimental verification on the ImageNet dataset and VAR model shows that the AUROC (area under the receiver operating characteristic curve, the higher the score, the stronger the detection accuracy) score of the present invention is significantly higher than that of the traditional scheme, which solves the problem of poor detection performance of the existing scheme.

[0020] 4. Strong robustness: This invention avoids the interference of abnormally high token scores in non-member samples by using a scoring strategy that removes tokens with high outlier scores and takes the average score of the tokens with the lowest scores. This ensures that it can stably distinguish between member samples and non-member samples under different model depths and different dataset scenarios, and its robustness is significantly better than existing solutions. Attached Figure Description

[0021] 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 recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings: Figure 1 This is a flowchart of an autoregressive image generation training data detection method provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of a model forward propagation prediction embodiment provided by an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of an autoregressive image generation training data detection device provided in an embodiment of the present invention. Detailed Implementation

[0022] To enable those skilled in the art to better understand the technical solutions of this invention, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this invention.

[0023] This invention provides a method for detecting training data for autoregressive image generation, such as... Figure 1 As shown, the autoregressive image generation training data detection method specifically includes steps S101-S103: S101. Control the prediction probability of several models for the target image token output by the autoregressive image generation model.

[0024] The training data detection process of the autoregressive image generation model constructed in this invention consists of three stages. First, the prediction probability acquisition stage obtains the prediction probabilities of four different models. Next, the scoring stage merges the four probabilities using a scoring mechanism, assigning a final score. Finally, the judgment stage identifies samples with higher scores as member samples and those with lower scores as non-member samples, thus achieving training data detection.

[0025] During the training of an autoregressive image generation model, an image is segmented into multiple tokens. The autoregressive image generation model can predict subsequent tokens using preceding tokens and given conditions. To use classifier-free guidance, the training process randomly replaces the conditions of model blocks with unconditional ones, directly using preceding tokens to predict subsequent tokens. Therefore, in the use of the autoregressive image generation model, when predicting image tokens, the model input passes through multiple model blocks to obtain the output, and each model block accepts either conditional or unconditional input, such as... Figure 2 As shown, different model output probabilities for image tokens can be obtained depending on the different conditional inputs to the model block.

[0026] Specifically, you first define the conditional input type for each model block in the autoregressive image generation model, resulting in several combinations of conditional inputs. Based on each combination of conditional inputs, you control the autoregressive image generation model to predict the target image token, obtaining several corresponding model prediction probabilities.

[0027] As a feasible implementation method, the conditional input type of all model blocks in the autoregressive image generation model is set to accept conditional input to obtain the first autoregressive image generation model; then, the target image token is predicted through the first autoregressive image generation model to obtain the first model prediction probability.

[0028] The conditional input type of all model blocks in the autoregressive image generation model is set to accept unconditional input to obtain the second autoregressive image generation model; the target image token is predicted using the second autoregressive image generation model to obtain the second model prediction probability.

[0029] The conditional input type of all model blocks in the autoregressive image generation model is set to accept conditional input for the first half of the model blocks and unconditional input for the second half of the model blocks, resulting in the third autoregressive image generation model. The target image token is predicted using the third autoregressive image generation model to obtain the prediction probability of the third model.

[0030] By setting the conditional input type of all model blocks in the autoregressive image generation model to accept unconditional input for the first half of the model blocks and conditional input for the second half, a fourth autoregressive image generation model is obtained. The fourth autoregressive image generation model is then used to predict the target image token, yielding the predicted probability of the fourth model.

[0031] In one embodiment, Figure 2 This is a schematic diagram of a model forward propagation prediction embodiment provided by an embodiment of the present invention, as shown below. Figure 2 As shown, different model output probabilities for image tokens can be obtained depending on the different input conditions: (1) “Conditional” prediction probability: All model blocks accept conditional inputs, and the prediction probability is denoted as P1.

[0032] (2) "Unconditional" prediction probability: All model blocks accept unconditional input, and the prediction probability is denoted as P2.

[0033] (3) “Conditional-unconditional” prediction probability: The first half of the model block accepts conditional input, and the second half of the model block accepts unconditional input. The prediction probability is denoted as P3.

[0034] (4) "Unconditional-conditional" prediction probability: The first half of the model block accepts unconditional input, and the second half of the model block accepts conditional input. The prediction probability is denoted as P4.

[0035] S102. Based on preset hyperparameters, the prediction probabilities of several models are weighted and summed to obtain the image score of the target image token.

[0036] Specifically, in this stage, the present invention introduces a scoring mechanism that performs a weighted summation of the four prediction probabilities obtained in the previous step to obtain a score S for each target image token: S = P1-P2+k (P3-P4). Wherein, P1 is the prediction probability of the first model, P2 is the prediction probability of the second model, P3 is the prediction probability of the third model, P4 is the prediction probability of the fourth model, and k is a preset hyperparameter.

[0037] The reason for using this form is as follows: (1) During the model training process, the training under conditions is more frequent than the training under no conditions, and the specific conditions make the model fit better. Therefore, the difference between the conditional prediction probability and the unconditional prediction probability used for training samples is greater.

[0038] (2) Research on large language models indicates that the accuracy of the prediction of the final result is most critical in the first few layers of model blocks. Therefore, in the case of "conditional-unconditional", the first half of the model blocks accepts conditional input and the second half of the model blocks accepts unconditional input, and the prediction probability is higher than in the case of "unconditional-conditional". The difference between the two can be used as an auxiliary verification.

[0039] Furthermore, after all target image tokens have had their image scores calculated, low-scoring image tokens with image scores below a first preset score threshold are selected, and the average image score of these low-scoring image tokens is calculated as the unified score for all low-scoring image tokens.

[0040] As a feasible implementation method, in practice, the present invention selects the m% of target image tokens with the lowest scores, averages their scores, and obtains the final score of these low-scoring image tokens, where m is a hyperparameter.

[0041] S103. Based on the image score and the preset judgment threshold, determine whether the target image token is a member sample participating in the training of the autoregressive image generation model.

[0042] Specifically, target image tokens with image scores higher than a second preset score threshold are removed.

[0043] Furthermore, among the remaining target image tokens, if the image score is higher than a preset judgment threshold, the target image token is determined to be a member sample participating in the training of the autoregressive image generation model. If the image score is not higher than the preset judgment threshold, the target image token is determined to be a non-member sample and has not participated in the training of the autoregressive image generation model.

[0044] As a feasible implementation method, since the prediction difficulty of different image tokens is different, some image tokens with abnormally high scores may appear in non-member samples, but most image tokens will still follow the low score, so it is necessary to remove some high-scoring tokens.

[0045] To verify the efficiency of this invention in autoregressive image generation models, experiments were conducted on the commonly used ImageNet dataset and advanced VAR models to examine the AUROC scores. Experiments were performed on models of different sizes, with the training data detection method of the autoregressive text generation model used as a baseline. The hyperparameters k and m were set to 0.25 and 80, respectively. The specific results are shown in Table 1. The results in the table show that the AUROC scores of this invention are significantly higher than the baseline method at different model depths. When the model depth is low, the AUROC score of the baseline method is close to 0.5, indicating random guessing, which is almost meaningless for determining whether the data is training data. In contrast, the solution of this invention is 0.15 and 0.26 higher than the best baseline in model depths of 16 and 20, respectively. At larger model depths, such as 24 and 30, the AUROC scores reach 0.96 and 1.00, respectively, demonstrating that the solution of this invention is extremely accurate in detecting whether the data is training data.

[0046] Table 1

[0047] In addition, embodiments of the present invention also provide an autoregressive image generation training data detection device, such as... Figure 3 As shown, the autoregressive image generation training data detection device specifically includes: At least one processor; and a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by at least one processor, so that at least one processor can perform the following: Several model prediction probabilities are used to control the output of the target image token by the autoregressive image generation model. Based on preset hyperparameters, the prediction probabilities of the several models are weighted and summed to obtain the image score of the target image token; Based on the image score and a preset judgment threshold, it is determined whether the target image token is a member sample participating in the training of the autoregressive image generation model.

[0048] This invention addresses the training data detection needs of autoregressive image generation models. Through an innovative three-stage technical architecture of "prediction probability acquisition-scoring-judgment," it overcomes many limitations of existing technologies, achieving significant improvements in adaptability, efficiency, accuracy, and robustness. 1. This invention does not simply transfer the detection logic of other models (such as diffusion models and autoregressive text models), but rather customizes the design based on the core characteristics of autoregressive image generation models. It fully considers the essential features of these models—using image token sequences as generation units, multi-model block collaborative inference as the core architecture, and conditional / unconditional input guiding the generation direction—and specifically designs a technical path to control different model blocks to accept conditional or unconditional input to obtain prediction probabilities. This design precisely matches the inference logic of mainstream autoregressive image generation models such as VAR, VAR-CLIP, and M-VAR, completely solving the problem of existing solutions being incompatible with model characteristics and unable to effectively integrate into the detection process. It successfully fills the technical gap in training data detection in the field of autoregressive image generation, providing a dedicated and suitable technical solution for privacy and copyright protection in this field.

[0049] 2. Compared to existing related patents that require additional training of member / non-member sample feature extraction models, consuming significant computational resources and time, this invention simplifies the detection process into a lightweight architecture of model forward inference + score calculation: four core prediction probabilities can be directly obtained through only a limited number of model forward inferences, without the need to build an additional training framework, prepare a large amount of auxiliary training data, or consume resources for feature learning. This design significantly reduces the computational overhead of the detection process, eliminates the need for high-performance hardware to run training tasks for extended periods, and significantly shortens the detection cycle. Training preparation, which previously required hours to days, is now compressed to complete a single sample detection within hundreds of milliseconds. This enables rapid response to the needs of batch sample detection and real-time detection in real-world scenarios, better meeting the actual needs of enterprises, research institutions, and other users for low-cost, high-efficiency detection tools.

[0050] 3. This invention effectively improves the accuracy of detection results through multi-dimensional probability fusion and a refined scoring strategy. On the one hand, by simultaneously introducing the difference between conditional and unconditional prediction probabilities, and the probability difference under different model block input combinations, it captures the essential differences between member samples and non-member samples from two core dimensions. Member samples, having participated in model training, exhibit higher prediction fit under conditional guidance and more regular probability changes under different model block input combinations, while non-member samples lack this regularity. On the other hand, by eliminating abnormally high-scoring tokens and averaging the scores of low-scoring tokens, it avoids interference from individual tokens in non-member samples that may have high scores due to accidental factors, further strengthening the reliability of the judgment criteria. Ultimately, this scheme can clearly distinguish between member samples that have participated in training and non-member samples that have not, preventing copyright protection failure due to mistakenly classifying member samples as non-members, and avoiding disputes caused by misjudgment due to mistakenly classifying non-member samples as members, thus achieving reliable detection of training data.

[0051] 4. The technical solution of this invention possesses excellent robustness, maintaining stable detection performance across various application scenarios. Whether dealing with autoregressive image generation models of varying depths and architectures (such as shallow VAR models and deep M-VAR models), or processing image data of different types and complexities (such as simple scene images and complex texture images), this solution can output stable results through a standardized probability acquisition-scoring-judgment process. It will not experience a sudden drop in detection performance due to model structure adjustments, nor will it generate numerous false positives due to differences in image data features. This robustness stems from the solution's grasp of the general logic of autoregressive image generation models: it does not rely on specific model-specific parameters or image-specific features, but rather designs detection metrics based on the token prediction logic and conditional guidance mechanisms common to this type of model. Therefore, it can flexibly adapt to diverse models and data scenarios, providing users with a stable detection tool usable in multiple scenarios, reducing the additional cost of adjusting the detection solution due to scenario changes.

[0052] The various embodiments in this invention are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments of apparatus, devices, and non-volatile computer storage media are basically similar to the method embodiments, so the descriptions are relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0053] The foregoing has described specific embodiments of the present invention. Furthermore, the processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired results. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0054] The above description is merely an embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, various modifications and variations can be made to the embodiments of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the embodiments of the present invention should be included within the protection scope of the present invention.

Claims

1. An autoregressive image generation training data detection method, characterized in that, The method includes: Several model prediction probabilities are used to control the output of the target image token by the autoregressive image generation model, specifically including: The conditional input type of each model block in the autoregressive image generation model is set to obtain several combinations of conditional inputs; Based on each combination of conditional inputs, the autoregressive image generation model is controlled to predict the target image token, resulting in several corresponding model prediction probabilities, specifically including: The conditional input type of all model blocks in the autoregressive image generation model is set to accept conditional input for the first half of the model blocks and accept unconditional input for the second half of the model blocks, to obtain the third autoregressive image generation model; the target image token is predicted using the third autoregressive image generation model to obtain the third model prediction probability. The conditional input type of all model blocks in the autoregressive image generation model is set to accept unconditional input for the first half of the model blocks and conditional input for the second half of the model blocks, thus obtaining the fourth autoregressive image generation model. The target image token is predicted using the fourth autoregressive image generation model to obtain the predicted probability of the fourth model. Based on preset hyperparameters, the predicted probabilities of the several models are weighted and summed to obtain the image score of the target image token, specifically including: Based on S = P1-P2+k (P3-P4), the prediction probabilities of the several models are weighted and summed to obtain the image score S of the target image token; Wherein, P1 is the prediction probability of the first model, P2 is the prediction probability of the second model, P3 is the prediction probability of the third model, P4 is the prediction probability of the fourth model, and k is the preset hyperparameter; Based on the image score and a preset judgment threshold, it is determined whether the target image token is a member sample participating in the training of the autoregressive image generation model, specifically including: Remove target image tokens whose image scores are higher than the second preset score threshold; If the image score is higher than the preset judgment threshold among the remaining target image tokens, then the target image token is determined to be a member sample participating in the training of the autoregressive image generation model. If the image score is not higher than the preset judgment threshold, the target image token is determined to be a non-member sample and has not participated in the training of the autoregressive image generation model.

2. The method of claim 1, wherein, Based on each combination of conditional inputs, the autoregressive image generation model is controlled to predict the target image token, resulting in several corresponding model prediction probabilities, specifically including: Set all model blocks in the autoregressive image generation model to accept conditional input to obtain the first autoregressive image generation model. The target image token is predicted using the first autoregressive image generation model to obtain the first model prediction probability.

3. The method of claim 1, wherein, Based on each combination of conditional inputs, the autoregressive image generation model is controlled to predict the target image token, resulting in several corresponding model prediction probabilities, specifically including: The conditional input type of all model blocks in the autoregressive image generation model is set to accept unconditional input to obtain the second autoregressive image generation model. The target image token is predicted using the second autoregressive image generation model to obtain the second model prediction probability.

4. The method of claim 1, wherein, The method further includes: After all target image tokens have had their image scores calculated, low-scoring image tokens with image scores below a first preset score threshold are selected, and the average image score of these low-scoring image tokens is calculated as the unified score of these low-scoring image tokens.

5. An autoregressive image generation training data detection device, characterized by, The device includes: At least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions executable by the at least one processor to enable the at least one processor to perform an autoregressive image generation training data detection method according to any one of claims 1-4.

Images