A hallucination management system based on adversarial training

Abstract

The present invention provides an adversarial training-based hallucination management system, which relates to the field of electronic digital data processing, and comprises a data management module, an adversarial training module, an interactive evaluation module and a model update module. The data management module is used to manage image data, the adversarial training module is used to perform adversarial training to improve the quality of generated images and the ability to detect image hallucinations, the interactive evaluation module is used to display hallucination detection results and evaluate the detection ability, and the model update module is used to control the update of the image generation model. The system can greatly reduce the probability of hallucinations in generated images and improve the quality of generated images through continuous adversarial upgrades of image generation and hallucination judgment.

Description

Technical Field

[0001] The present invention relates to the field of electronic digital data processing, and in particular to a hallucination management system based on adversarial training. Background Art

[0002] Hallucination refers to the situation where the image output by the generative model contains content that is inconsistent with real semantics, illogical, or visually abnormal. These hallucinations not only reduce the quality of the generated results, but may also lead to misleading information in the application scenario, thus affecting the user experience. Existing generative model optimization methods mostly focus on improving the visual quality and diversity of generated images, but there are relatively few studies and practices on hallucination management. Some studies have attempted to reduce hallucinations by directly improving the model architecture or training objectives, but these methods are often not targeted and cannot accurately identify and control hallucination content. In addition, the generative model relies on a large amount of high-quality data during training and application. In actual scenarios, the cost of data annotation is high, and the quality assessment of the generated results often relies on manual observation, which is inefficient.

[0003] Now many hallucination management systems have been developed. After a lot of searching and reference, it is found that the existing hallucination management systems are such as the system disclosed in the publication number CN117556920B. These systems generally include obtaining a target question-answer pair, wherein the target question-answer pair includes: a target question and an answer to be verified; inputting the target question-answer pair into a target model to obtain a question-answer type corresponding to the target question-answer pair, wherein the target model is obtained by iteratively training the model to be trained with a training sample set; determining a verification rule according to the question-answer type, and verifying the answer to be verified based on the verification rule and the target question-answer pair. If the verification passes, the answer to be verified is determined to be the target answer. However, this system is only applicable to text hallucinations and cannot handle hallucination problems in images. Summary of the invention

[0004] The purpose of the present invention is to propose a hallucination management system based on adversarial training to address the existing deficiencies.

[0005] The present invention adopts the following technical solution:

[0006] A hallucination management system based on adversarial training, including a data management module, an adversarial training module, an interactive evaluation module and a model update module;

[0007] The data management module is used to manage image data, the adversarial training module is used to perform adversarial training to improve the quality of generated images and the ability to detect image hallucinations, the interactive evaluation module is used to display the detection results of hallucinations and evaluate the detection ability, and the model update module is used to control the update of the image generation model;

[0008] The data management module includes an image storage unit, an illusion transformation unit and a classification management unit, wherein the image storage unit is used to store image information, the illusion transformation unit is used to perform illusion processing on normal images, and the classification management unit is used to perform classification management on normal images and illusion images;

[0009] The adversarial training module includes an image generation unit, an illusion judgment unit and a training control unit, wherein the image generation unit is used to generate image information, the illusion judgment unit is used to judge and analyze the illusions in the image, and the training control unit is used to control the adversarial training process;

[0010] The interactive evaluation module includes a judgment display unit, an interactive operation unit and a statistical feedback unit. The judgment display unit is used to display the hallucination information judged to exist, the interactive operation unit is used to perform operation judgment on the accuracy of the hallucination judgment, and the statistical feedback unit is used to collect statistics on the judgment results and feed them back to the adversarial training module.

[0011] The model updating module includes a model docking unit, an effect evaluation unit and a data updating unit, wherein the model docking unit is used to dock with the image generation model, the effect evaluation unit is used to evaluate the improvement effect of the image generation unit, and the data updating unit is used to update the image generation unit data to the image generation model when the effect improvement degree reaches the requirement;

[0012] Further, the image generation unit includes a generation model processor, a generation optimization processor and an image output processor, the generation model processor is used to generate image data, the generation optimization processor is used to optimize the parameters of the generation model, and the image output processor is used to output the generated image data to the training control unit;

[0013] The hallucination judgment unit includes a judgment model processor, a judgment optimization processor and a result output processor, wherein the judgment model processor is used to judge the hallucinations existing in the image, the judgment optimization processor is used to optimize the parameters of the judgment model, and the result output processor is used to output the judgment result to the training control unit;

[0014] The training control unit includes an image mixing processor, a contrast control processor and an optimization feedback processor, wherein the image mixing processor is used to mix the generated image, the normal image and the hallucination image as a sample, the contrast control processor is used to control the contrast process of the sample judgment result, and the optimization feedback processor is used to feed back optimization information to the generation optimization processor and the judgment optimization processor;

[0015] Further, the statistical feedback unit includes a generating statistical processor, a judging statistical processor and a statistical output processor, wherein the generating statistical processor is used to statistically generate data information for optimizing image generation, the judging statistical processor is used to statistically generate data information for optimizing hallucination judgment, and the statistical output processor is used to output the statistical results to the training control unit;

[0016] Furthermore, the optimization feedback processor calculates the optimization selection index P according to the following formula:

[0017] ;

[0018] Among them, R1 is the illusion ratio, R2 is the disguise ratio, R3 is the judgment ratio, and R4 is the correct ratio;

[0019] When P is greater than 0, the generation model processor is optimized, otherwise the judgment model processor is optimized;

[0020] Furthermore, the hallucination management process of this system includes the following steps:

[0021] S1, the image mixing processor sends an instruction to enable the generation model processor to generate a specific number of images;

[0022] S2, an image blending processor obtains an image from an image storage unit and forms a sample with the generated image;

[0023] S3, the judgment model processor performs hallucination judgment on the sample;

[0024] S4, the comparison control processor compares the image attribute with the judgment result and sends it to the judgment display unit;

[0025] S5. judging the judgment result of the generated image through the interactive operation unit;

[0026] S6. The statistical feedback unit collects statistics on the correctness of the judgment result and feeds back the statistics to the optimization feedback processor;

[0027] S7, the optimization feedback processor sends the optimization information to the generation optimization processor or the judgment optimization processor based on the statistical results;

[0028] S8, the generation optimization processor or the judgment optimization processor optimizes the generation model processor or the judgment model processor based on the optimization information;

[0029] S9. Repeat steps S1 to S8 until the comprehensive evaluation effect reaches a threshold, and update the difference information of the generated model processor to the docked model.

[0030] The beneficial effects achieved by the present invention are:

[0031] This system sets up an image generation model and a hallucination judgment model, mixes normal images and hallucination images, and then conducts adversarial training to continuously improve the ability to disguise and recognize hallucinations in generated images, thereby reducing the probability of hallucinations in generated images and improving the quality of generated images. During the training process, the model optimization is selected based on statistical data, which has a higher optimization effect than directly optimizing the generation model.

[0032] To further understand the features and technical contents of the present invention, please refer to the following detailed description and drawings of the present invention. However, the drawings provided are only for reference and description and are not intended to limit the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 It is a schematic diagram of the overall structural framework of the present invention;

[0034] Figure 2 This is a schematic diagram of the data management module of the present invention;

[0035] Figure 3 This is a schematic diagram of the adversarial training module of the present invention;

[0036] Figure 4 This is a schematic diagram of the interactive evaluation module of the present invention;

[0037] Figure 5 This is a schematic diagram of the model update module of the present invention;

[0038] Figure 6 This is a comparison chart of the optimization effects of the model of the present invention. DETAILED DESCRIPTION

[0039] The following is an explanation of the embodiments of the present invention through specific embodiments. Those skilled in the art can understand the advantages and effects of the present invention from the contents disclosed in this specification. The present invention can be implemented or applied through other different specific embodiments, and the details in this specification can also be modified and changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. In addition, the drawings of the present invention are only simple schematic illustrations and are not depicted according to actual sizes. It is stated in advance. The following embodiments will further explain the relevant technical contents of the present invention in detail, but the disclosed contents are not intended to limit the scope of protection of the present invention.

[0040] Embodiment 1: This embodiment provides a hallucination management system based on adversarial training, combined with Figure 1 , including data management module, adversarial training module, interactive evaluation module and model update module.

[0041] The data management module is used to manage image data, the adversarial training module is used to perform adversarial training to improve the quality of generated images and the ability to detect image hallucinations, the interactive evaluation module is used to display the detection results of hallucinations and evaluate the detection ability, and the model update module is used to control the update of the image generation model.

[0042] The data management module includes an image storage unit, an illusion transformation unit and a classification management unit. The image storage unit is used to store image information. The illusion transformation unit is used to perform illusion processing on normal images. The classification management unit is used to perform classification management on normal images and illusion images.

[0043] The adversarial training module includes an image generation unit, an illusion judgment unit and a training control unit. The image generation unit is used to generate image information, the illusion judgment unit is used to judge and analyze the illusions existing in the image, and the training control unit is used to control the adversarial training process.

[0044] The interactive evaluation module includes a judgment display unit, an interactive operation unit and a statistical feedback unit. The judgment display unit is used to display the hallucination information judged to exist, the interactive operation unit is used to perform operational judgment on the accuracy of the hallucination judgment, and the statistical feedback unit is used to count the judgment results and feed them back to the adversarial training module.

[0045] The model update module includes a model docking unit, an effect evaluation unit and a data update unit. The model docking unit is used to dock with the image generation model, the effect evaluation unit is used to evaluate the improvement effect of the image generation unit, and the data update unit is used to update the image generation unit data to the image generation model when the effect improvement degree reaches the requirement.

[0046] The image generation unit includes a generation model processor, a generation optimization processor and an image output processor. The generation model processor is used to generate image data, the generation optimization processor is used to optimize the parameters of the generation model, and the image output processor is used to output the generated image data to the training control unit.

[0047] The hallucination judgment unit includes a judgment model processor, a judgment optimization processor and a result output processor. The judgment model processor is used to judge the hallucinations existing in the image, the judgment optimization processor is used to optimize the parameters of the judgment model, and the result output processor is used to output the judgment result to the training control unit.

[0048] The training control unit includes an image mixing processor, a contrast control processor and an optimization feedback processor. The image mixing processor is used to mix the generated image, the normal image and the hallucination image as samples. The contrast control processor is used to control the contrast process of the sample judgment results. The optimization feedback processor is used to feed back optimization information to the generation optimization processor and the judgment optimization processor.

[0049] The statistical feedback unit includes a generating statistical processor, a judging statistical processor and a statistical output processor. The generating statistical processor is used to statistically generate data information for optimizing image generation, the judging statistical processor is used to statistically generate data information for optimizing hallucination judgment, and the statistical output processor is used to output the statistical results to the training control unit.

[0050] The optimization feedback processor calculates the optimization selection index P according to the following formula:

[0051] .

[0052] Among them, R1 is the illusion ratio, R2 is the disguise ratio, R3 is the judgment ratio, and R4 is the correct ratio.

[0053] When P is greater than 0, the generation model processor is optimized, otherwise the judgment model processor is optimized.

[0054] The hallucination management process of this system includes the following steps:

[0055] S1. The image mixing processor sends instructions to enable the generation model processor to generate a specific number of images.

[0056] S2. The image mixing processor obtains the image from the image storage unit and forms a sample with the generated image.

[0057] S3. The judgment model processor performs hallucination judgment on the sample.

[0058] S4. The comparison control processor compares the image attribute with the judgment result and sends it to the judgment display unit.

[0059] S5. Evaluate the judgment result of the generated image through the interactive operation unit.

[0060] S6. The statistical feedback unit collects statistics on the correctness of the judgment result and feeds back the statistics to the optimization feedback processor.

[0061] S7. The optimization feedback processor sends the optimization information to the generation optimization processor or the judgment optimization processor based on the statistical results.

[0062] S8. The generation optimization processor or the judgment optimization processor optimizes the generation model processor or the judgment model processor based on the optimization information.

[0063] S9. Repeat steps S1 to S8 until the comprehensive evaluation effect reaches a threshold, and update the difference information of the generated model processor to the docked model.

[0064] Embodiment 2: This embodiment includes all the contents of Embodiment 1, and provides an adversarial training-based hallucination management system, including a data management module, an adversarial training module, an interactive evaluation module, and a model update module.

[0065] The data management module is used to manage image data, the adversarial training module is used to perform adversarial training to improve the quality of generated images and the ability to detect image hallucinations, the interactive evaluation module is used to display the detection results of hallucinations and evaluate the detection ability, and the model update module is used to control the update of the image generation model.

[0066] Combination Figure 2 The data management module includes an image storage unit, an illusion transformation unit and a classification management unit. The image storage unit is used to store image information, the illusion transformation unit is used to perform illusion processing on normal images, and the classification management unit is used to classify and manage normal images and illusion images.

[0067] Combination Figure 3 The adversarial training module includes an image generation unit, an illusion judgment unit and a training control unit. The image generation unit is used to generate image information, the illusion judgment unit is used to judge and analyze the illusions in the image, and the training control unit is used to control the adversarial training process.

[0068] Combination Figure 4 The interactive evaluation module includes a judgment display unit, an interactive operation unit and a statistical feedback unit. The judgment display unit is used to display the hallucination information judged to exist, the interactive operation unit is used to perform operational judgment on the accuracy of the hallucination judgment, and the statistical feedback unit is used to count the judgment results and feed them back to the adversarial training module.

[0069] Combination Figure 5 The model updating module includes a model docking unit, an effect evaluation unit and a data updating unit. The model docking unit is used to dock with the image generation model. The effect evaluation unit is used to evaluate the improvement effect of the image generation unit. The data updating unit is used to update the image generation unit data to the image generation model when the effect improvement degree reaches the requirement.

[0070] The image storage unit includes an image data register, an image information register and an image mapping processor. The image data register is used to store specific image data, the image information register is used to store attribute information of each image, and the image mapping processor is used to map the storage address of the image data with the corresponding attribute information.

[0071] The illusion transformation unit includes an image flow processor, a type locking processor and an illusion execution processor. The image flow processor is used to obtain a normal image and output a transformed illusion image. The type locking processor locks the transformation area in the image based on the illusion type. The illusion execution processor is used to perform illusion transformation processing on the transformation area.

[0072] The classification management unit includes an image encoding processor, a storage control processor and a classification retrieval processor. The image encoding processor is used to encode the image, the storage control processor is used to store the hallucination image in the corresponding area, and the classification retrieval processor is used to retrieve the required image from the image storage unit.

[0073] The image generation unit includes a generation model processor, a generation optimization processor and an image output processor. The generation model processor is used to generate image data, the generation optimization processor is used to optimize the parameters of the generation model, and the image output processor is used to output the generated image data to the training control unit.

[0074] The hallucination judgment unit includes a judgment model processor, a judgment optimization processor and a result output processor. The judgment model processor is used to judge the hallucinations existing in the image, the judgment optimization processor is used to optimize the parameters of the judgment model, and the result output processor is used to output the judgment result to the training control unit.

[0075] The training control unit includes an image mixing processor, a contrast control processor and an optimization feedback processor. The image mixing processor is used to mix the generated image, the normal image and the hallucination image as samples. The contrast control processor is used to control the contrast process of the sample judgment results. The optimization feedback processor is used to feed back optimization information to the generation optimization processor and the judgment optimization processor.

[0076] The judgment display unit includes an automatic display processor and an operation display processor. The automatic display processor is used to display the judgment result of the normal image and the hallucination image, and the operation display processor is used to display the judgment result of the generated image.

[0077] The interactive operation unit includes a judgment operation processor, a circle operation processor and an operation acquisition processor. The judgment operation processor is used to judge the correctness of the generated image judgment result, the circle operation processor is used to circle the undetected hallucination part in the generated image, and the operation acquisition processor is used to collect all operation information and digitally process it.

[0078] The statistical feedback unit includes a generating statistical processor, a judging statistical processor and a statistical output processor. The generating statistical processor is used to statistically generate data information for optimizing image generation, the judging statistical processor is used to statistically generate data information for optimizing hallucination judgment, and the statistical output processor is used to output the statistical results to the training control unit.

[0079] The model docking unit includes an interface adaptation processor, a core extraction processor and a model reconstruction processor. The interface adaptation processor is used to connect to the model and perform interface adaptation. The core extraction processor is used to extract structural parameter information in the model for generating images. The model reconstruction processor reconstructs the model based on the structural parameter information and sends it to the generation model processor.

[0080] The effect evaluation unit includes an optimization record processor, a single evaluation processor and a comprehensive evaluation processor. The optimization record processor is used to record each optimization information of the generation model processor, the word evaluation processor is used to evaluate the effect of the generation model processor before and after optimization, and the comprehensive evaluation processor is used to evaluate the overall improvement effect of multiple optimizations.

[0081] The data update unit includes a threshold judgment processor, a difference extraction processor and a difference synchronization processor. The threshold judgment processor is used to judge whether the comprehensive improvement effect reaches the update threshold. The difference extraction processor is used to extract the difference information between the generated model processor and the last update. The difference synchronization processor is used to synchronize the difference information to the docked model.

[0082] The hallucination management process of this system includes the following steps:

[0083] S1. The image mixing processor sends instructions to enable the generation model processor to generate a specific number of images.

[0084] S2. The image mixing processor obtains the image from the image storage unit and forms a sample with the generated image.

[0085] S3. The judgment model processor performs hallucination judgment on the sample.

[0086] S4. The comparison control processor compares the image attribute with the judgment result and sends it to the judgment display unit.

[0087] S5. Evaluate the judgment result of the generated image through the interactive operation unit.

[0088] S6. The statistical feedback unit collects statistics on the correctness of the judgment result and feeds back the statistics to the optimization feedback processor.

[0089] S7. The optimization feedback processor sends the optimization information to the generation optimization processor or the judgment optimization processor based on the statistical results.

[0090] S8. The generation optimization processor or the judgment optimization processor optimizes the generation model processor or the judgment model processor based on the optimization information.

[0091] S9. Repeat steps S1 to S8 until the comprehensive evaluation effect reaches a threshold, and update the difference information of the generated model processor to the docked model.

[0092] The data information counted by the generation statistics processor is the illusion ratio and the disguise ratio, wherein the illusion ratio refers to the ratio of the number of illusions in the generated image to the number of generated images, and the disguise ratio is the ratio of the number of illusions in the generated image that are not recognized by the judgment model processor to the number of illusions in the generated image.

[0093] The data information counted by the judgment statistics processor is the judgment ratio and the correct ratio, wherein the judgment ratio refers to the ratio of the number of times the judgment model processor makes judgments to the number of images in the sample, and the correct ratio refers to the ratio of the number of correct judgments made by the judgment model processor to the total number of judgments.

[0094] The optimization feedback processor calculates the optimization selection index P according to the following formula:

[0095] .

[0096] Among them, R1 is the illusion ratio, R2 is the disguise ratio, R3 is the judgment ratio, and R4 is the correct ratio.

[0097] When P is greater than 0, the generation model processor is optimized, otherwise the judgment model processor is optimized.

[0098] The generation optimization processor selects the hallucination part identified in the generated image and maps it to the model node of the generated hallucination part. When there are at least two hallucination parts mapped to the same model node, the common feature information of the hallucination part is extracted, and the parameters in the model node are inversely adjusted to achieve the optimization effect.

[0099] For example, when the parameter When increasing will enhance feature information, reducing the parameter This is called inverse feature adjustment of the feature information.

[0100] The judgment optimization processor selects the region portion of the image that is judged incorrectly, maps it to the model node that judges the region portion, and performs inverse judgment adjustment on the parameters in the model node to achieve an optimization effect.

[0101] For example, if the region is partially hallucinated but not recognized, then if the parameter When increasing the value will reduce the probability of judging it as an illusion, reduce the parameter It is called the reverse judgment adjustment of this judgment.

[0102] The single evaluation processor calculates the optimization effect index U according to the following formula:

[0103] .

[0104] Among them, R 1a is the optimized illusion ratio, R 1b is the illusion ratio before optimization, R 2a is the optimized camouflage ratio, R 2b is the camouflage ratio before optimization.

[0105] The comprehensive evaluation processor calculates the overall effect index Q according to the following formula:

[0106] .

[0107] Where n is the total number of optimizations of the generated model processor after the last model update, and U(i) represents the optimization effect index of the i-th time.

[0108] When Q is greater than the update threshold, the docked model is updated.

[0109] The i appearing in the above text is an ordinal number used to indicate a sequence number and has no actual meaning.

[0110] Some of the code for this system is described below:

[0111] class Generator(nn.Module):

[0112] def __init__(self):

[0113] super(Generator, self).__init__()

[0114] self.main = nn.Sequential(

[0115] nn.Linear(100, 256),

[0116] nn.ReLU(),

[0117] nn.Linear(256, 512),

[0118] nn.ReLU(),

[0119] nn.Linear(512, 28*28),

[0120] nn.Tanh() )

[0122] def forward(self, x):

[0123] return self.main(x.view(-1, 28*28))

[0124] class TrainingController:

[0125] def __init__(self, generator, discriminator, lr=0.0002):

[0126] self.generator = generator

[0127] self.discriminator = discriminator

[0128] self.optimizer_g = optim.Adam(generator.parameters(), lr=lr)

[0129] self.optimizer_d = optim.Adam(discriminator.parameters(), lr=lr)

[0130] self.criterion = nn.BCELoss()

[0131] def train_step(self, real_images, batch_size):

[0132] """Single-step training of the generator and discriminator"""

[0133] # Train the discriminator

[0134] real_labels = torch.ones(batch_size, 1)

[0135] fake_labels = torch.zeros(batch_size, 1)

[0136] real_outputs = self.discriminator(real_images)

[0137] d_loss_real = self.criterion(real_outputs, real_labels)

[0138] z = torch.randn(batch_size, 100)

[0139] fake_images = self.generator(z)

[0140] fake_outputs = self.discriminator(fake_images.detach())

[0141] d_loss_fake = self.criterion(fake_outputs, fake_labels)

[0142] d_loss = d_loss_real + d_loss_fake

[0143] self.optimizer_d.zero_grad()

[0144] d_loss.backward()

[0145] self.optimizer_d.step()

[0146] # Train the generator

[0147] z = torch.randn(batch_size, 100)

[0148] fake_images = self.generator(z)

[0149] fake_outputs = self.discriminator(fake_images)

[0150] g_loss = self.criterion(fake_outputs, real_labels)

[0151] self.optimizer_g.zero_grad()

[0152] g_loss.backward()

[0153] self.optimizer_g.step()

[0154] return d_loss.item(), g_loss.item().

[0155] Now, through actual testing of the training effects of the common system and this system, we can obtain Figure 6 Comparison chart shown.

[0156] The contents disclosed above are only preferred feasible embodiments of the present invention, and do not limit the protection scope of the present invention. Therefore, all equivalent technical changes made using the contents of the present invention specification and drawings are included in the protection scope of the present invention. In addition, the elements therein can be updated as technology develops.

Claims

1. A hallucination management system based on adversarial training, characterized in that: It includes a data management module, an adversarial training module, an interactive evaluation module and a model update module; the data management module is used to manage image data, the adversarial training module is used to perform adversarial training to improve the quality of generated images and the ability to detect image hallucinations, the interactive evaluation module is used to display the detection results of hallucinations and evaluate the detection ability, and the model update module is used to control the update of the image generation model; the data management module includes an image storage unit, an illusion transformation unit and a classification management unit, the image storage unit is used to store image information, the illusion transformation unit is used to perform hallucination processing on normal images, and the classification management unit is used to classify and manage normal images and hallucination images; the adversarial training module includes an image generation unit, an illusion judgment unit and a training control unit, the image generation unit is used to generate image information, the illusion judgment unit is used to judge and analyze the illusions in the image, and the training control unit is used to control the adversarial training process; the image generation unit includes a generation model processor, the generation model processor is used to generate image data; the illusion judgment unit includes a judgment model processor, the judgment model processor is used to judge the illusions in the image; The interactive evaluation module includes a judgment display unit, an interactive operation unit and a statistical feedback unit. The judgment display unit is used to display the hallucination information judged to exist, the interactive operation unit is used to perform operation judgment on the accuracy of the hallucination judgment, and the statistical feedback unit is used to collect statistics on the judgment results and feed them back to the adversarial training module. The adversarial training module controls the adversarial training process based on the evaluation result; The model updating module includes a model docking unit, an effect evaluation unit and a data updating unit, wherein the model docking unit is used to dock with the image generation model, the effect evaluation unit is used to evaluate the improvement effect of the image generation unit, and the data updating unit is used to update the image generation unit data to the image generation model when the effect improvement degree reaches the requirement; Controlling the adversarial training process involves optimizing the generative model processor according to the optimization selection index P: ； Among them, R1 is the illusion ratio, R2 is the disguise ratio, R3 is the judgment ratio, and R4 is the correct ratio; When P is greater than 0, the generation model processor is optimized, otherwise the judgment model processor is optimized.

2. A hallucination management system based on adversarial training as claimed in claim 1, characterized in that: The image generation unit further comprises a generation optimization processor and an image output processor, wherein the generation optimization processor is used to optimize parameters of the generation model, and the image output processor is used to output the generated image data to the training control unit; The hallucination judgment unit further includes a judgment optimization processor and a result output processor, wherein the judgment optimization processor is used to optimize the parameters of the judgment model, and the result output processor is used to output the judgment result to the training control unit; The training control unit includes an image mixing processor, a contrast control processor and an optimization feedback processor. The image mixing processor is used to mix the generated image, the normal image and the hallucination image as samples. The contrast control processor is used to control the contrast process of the sample judgment results. The optimization feedback processor is used to feed back optimization information to the generation optimization processor and the judgment optimization processor.

3. A hallucination management system based on adversarial training as claimed in claim 2, characterized in that: The statistical feedback unit includes a generating statistical processor, a judging statistical processor and a statistical output processor. The generating statistical processor is used to statistically generate data information for optimizing image generation, the judging statistical processor is used to statistically generate data information for optimizing hallucination judgment, and the statistical output processor is used to output the statistical results to the training control unit.

4. A hallucination management system based on adversarial training as claimed in claim 3, characterized in that: The hallucination management process of this system includes the following steps: S1, the image mixing processor sends an instruction to enable the generation model processor to generate a specific number of images; S2, an image blending processor obtains an image from an image storage unit and forms a sample with the generated image; S3, the judgment model processor performs hallucination judgment on the sample; S4, the comparison control processor compares the image attribute with the judgment result and sends it to the judgment display unit; S5. judging the judgment result of the generated image through the interactive operation unit; S6. The statistical feedback unit collects statistics on the correctness of the judgment result and feeds back the statistics to the optimization feedback processor; S7, the optimization feedback processor sends the optimization information to the generation optimization processor or the judgment optimization processor based on the statistical results; S8, the generation optimization processor or the judgment optimization processor optimizes the generation model processor or the judgment model processor based on the optimization information; S9. Repeat steps S1 to S8 until the comprehensive evaluation effect reaches a threshold, and update the difference information of the generated model processor to the docked model.

Images