CN-122023873-A - AIGC true and false graph detection method based on frequency domain error analysis

CN122023873ACN 122023873 ACN122023873 ACN 122023873ACN-122023873-A

Abstract

The invention relates to a AIGC true and false graph detection method based on frequency domain error analysis, which comprises the steps of processing target images by a high-frequency image generation model and a low-frequency image generation model respectively, outputting the high-frequency image and the low-frequency image, obtaining errors among the target images, the high-frequency image and the low-frequency image and reconstructed images thereof by reconstruction error model processing, constructing three classification models, respectively processing the probability of the target image corresponding to each combination error relative to a true image label, finally weighting to obtain the comprehensive probability of the target image relative to the true image label, realizing true and false judgment of the target image, effectively capturing inherent defects of AIGC images in terms of frequency spectrum consistency by jointly analyzing reconstruction error responses of the images on the high-frequency and low-frequency components, realizing true and false discrimination capability of high accuracy and intensity, obviously improving detection sensitivity and robustness, and being applicable to various generation models and post-processing scenes.

Inventors

ZHANG SHENG
CHEN FEI
LIN GUANZHOU
ZHAO ZHEHAO
YU MINGYU

Assignees

北京信联数安科技有限公司

Dates

Publication Date: 20260512
Application Date: 20251225

Claims (9)

1. A AIGC true and false graph detection method based on frequency domain error analysis is characterized in that the following steps A to F are executed to obtain a true and false graph detection model; Step A, constructing a high-frequency image generation model and a low-frequency image generation model, receiving a target image by the high-frequency image generation model, processing and generating a corresponding high-frequency image and outputting the target image, receiving the target image by the low-frequency image generation model, processing and generating a corresponding low-frequency image and outputting the target image, and then entering the step B; step B, constructing a reconstruction error model, respectively receiving a target image, a high-frequency image and a low-frequency image by the reconstruction error model, firstly performing reconstruction to obtain the target reconstruction image, the high-frequency reconstruction image and the low-frequency reconstruction image, and then obtaining an error image between the target reconstruction image and the target image Error image between high frequency reconstructed image and high frequency image Error image between low frequency reconstructed image and low frequency image Outputting, and then entering the step C; Step C, constructing a first classification model, a second classification model and a third classification model, wherein the first classification model is used for receiving the error image 、、 Outputting the probability of the target image about the real image tag; A second classification model for receiving the error image Outputting the probability of the target image about the real image tag; a third classification model for receiving the error image 、 Outputting the probability of the target image about the real image tag; Then enter step D; step D, obtaining error images respectively corresponding to the sample images through the steps A to B based on the sample images respectively corresponding to the real image labels and the non-real image labels in the preset quantity 、、 Training the first classification model, the second classification model and the third classification model according to the real image label or the non-real image label corresponding to each sample image to obtain a first trained classification model, a second trained classification model and a third trained classification model, and then entering the step E; Step E, constructing a weighted judgment model, wherein the weighted judgment model comprises weights corresponding to a first classification model, a second classification model and a third classification model respectively, the weighted judgment model receives the probabilities of target images output by the first classification model, the second classification model and the third classification model respectively on real image labels, weighting is carried out, the comprehensive probabilities of the target images on the real image labels are obtained, a preset probability threshold is combined, whether the target images are judgment results of the real images or not is judged, and then step F is carried out; And F, constructing an authenticity graph checking model, wherein the input end of the high-frequency image generating model, the input end of the low-frequency image generating model and the input end of the reconstruction model are connected to form the input end of the authenticity graph checking model and are used for receiving the target image, the input end of the reconstruction model is simultaneously connected with the output end of the high-frequency image generating model and the output end of the low-frequency image generating model, the output end of the reconstruction model is respectively connected with the input end of the first trained classification model, the input end of the second trained classification model and the input end of the third trained classification model, the first trained classification model, the second trained classification model and the third trained classification model respectively output the probability of the target image on the real image label to the input end of the weighted judgment model, and the output end of the weighted judgment model forms the output end of the authenticity graph checking model and is used for outputting the judgment result of whether the target image is the real image.
2. The method for detecting true and false AIGC patterns based on frequency domain error analysis according to claim 1, wherein in the step A, a high-frequency image generation model is constructed as follows: first for a received target image Performing a two-dimensional discrete fourier transform as follows; ; Obtaining a target image Corresponding first frequency domain image , wherein, , , Representing a target image Is used for the display of the display device, Representing a target image In the number of pixels in the pixel array, Representing a target image In (a) The pixel value of the pixel location, Representing a first frequency domain image In the number of pixels in the pixel array, Representing a first frequency domain image In (a) The pixel value of the pixel location, Representing the mathematical constants of the data set, Representing the units of an imaginary number, Is equal to real number-1; At the same time, for the first frequency domain image The following high frequency mask model: ; obtaining corresponding high-frequency mask images , wherein, The predetermined radius is indicated as such, Representing a high frequency mask image In (a) Mask coefficients for pixel locations; then the following formula is adopted: ; Obtaining high-frequency component images , wherein, Representing a high frequency component image In (a) Frequency coefficients for pixel locations; Finally, aiming at the high-frequency component image The inverse two-dimensional discrete fourier transform is performed as follows: ; Obtaining corresponding high-frequency images , wherein, Representing a high frequency image In (a) Pixel values for pixel locations.
3. The method for detecting AIGC authenticity graph based on frequency domain error analysis according to claim 2, wherein the preset radius is The constraint range of (2) is: 。
4. The method for detecting the true or false AIGC map based on the frequency domain error analysis of claim 1, wherein in the step A, a low-frequency image generation model is constructed as follows: first for a received target image Performing two-dimensional discrete cosine transform according to the following model to obtain a second frequency domain image C, and partitioning the frequency domain image C according to a preset block size; then constructing a low-frequency mask image with the same overall size and block size as those of the frequency domain image C And define a low frequency mask image The mask coefficient of each pixel position in the preset proportional size range of the upper left corner in each block is 1, and the mask coefficient of each other pixel position in each block is defined as 0; Finally, the following formula is adopted: ; Obtaining a low frequency mask image And performing inverse two-dimensional discrete cosine transform to obtain low-frequency image , Representing a term-wise multiplication.
5. The method for detecting the AIGC true or false graph based on the frequency domain error analysis of claim 1, wherein the first classification model, the second classification model and the third classification model are ResNet convolutional neural networks.
6. The method for detecting AIGC authenticity graph based on frequency domain error analysis according to claim 1, wherein in the step D, based on each sample image of which a preset number corresponds to a real image label and a non-real image label respectively, the following image enhancement methods are respectively executed for each sample image to obtain each newly added sample image, and the corresponding real image label and the corresponding non-real image label are respectively associated, so that each sample image and each newly added sample image thereof jointly form each sample image; Randomly cutting, namely randomly cutting 50% -90% of sub-block images with proportional sizes from the sample image to form a newly added sample image; Random rotation and overturn, namely carrying out random plane rotation within a + -preset angle range on a sample image, and carrying out horizontal or vertical overturn according to preset probability to form a newly added sample image; And (3) noise injection, namely adding Gaussian white noise or salt and pepper noise with standard deviation of 0.01-0.05 into the sample image, and simulating signal interference in a real scene to form a newly added sample image.
7. The method for detecting the AIGC authenticity graph based on the frequency domain error analysis of claim 1, wherein in the step D, cross entropy loss is adopted in the training process of the first classification model, the second classification model and the third classification model respectively, an Adam optimizer is selected, the initial learning rate is set to be 0.001, and a learning rate attenuation strategy is adopted.
8. The method for detecting the AIGC true or false graph based on the frequency domain error analysis of claim 1, wherein L2 regularization and Dropout rate are introduced in the training process of the first classification model, the second classification model and the third classification model respectively in the step D.
9. The method for detecting the true or false graph of AIGC based on frequency domain error analysis of claim 1, wherein the weight of the first classification model is greater than the weight of the second classification model by greater than the weight of the third classification model.

Description

AIGC true and false graph detection method based on frequency domain error analysis Technical Field The invention relates to a AIGC true and false graph detection method based on frequency domain error analysis, and belongs to the technical field of artificial intelligent detection. Background In recent years, with the rapid development of the technology of generating artificial intelligence, the generated content of artificial intelligence (AI-GENERATED CONTENT, AIGC for short) has made a remarkable breakthrough in various fields of images, audio, texts and the like. Depth generation Models represented by Diffusion Models, generation countermeasure networks (GANs), and variational self-encoders (VAEs) have been able to generate highly realistic, indistinguishable image content to the naked eye. The techniques are widely applied to scenes such as digital art creation, virtual reality, film and television special effects, advertisement design and the like, and greatly improve the efficiency and diversity of content production. However, abuse of AIGC technology also presents serious social and security challenges. The phenomena of counterfeiting images (Deepfake), false news patterns, malicious synthesis of identity information and the like are increasingly inundated, and seriously threaten information authenticity, personal privacy protection and public opinion security. Particularly in the key fields of social media, judicial evidence obtaining, news spreading and the like, how to efficiently and accurately distinguish real images from AI generated images becomes a technical problem to be solved urgently. To address this challenge, researchers have proposed a number of AIGC image detection methods. Early methods rely mainly on microscopic marks in the spatial domain, such as noise distribution inconsistencies, color distortions, illumination mismatches, etc., for classification decisions. Subsequently, some deep learning based detection models have been proposed to automatically extract artifact features in the generated image by training Convolutional Neural Networks (CNNs) or vision Transformer (ViT). However, as the generated model is continuously optimized, visual flaws in the spatial domain of the output image thereof have been greatly reduced, resulting in a significant decrease in the detection performance of the above method. Notably, while modern generative models are capable of generating visually realistic images in the spatial domain, statistical deviations that are difficult to completely eliminate may remain in the frequency domain (i.e., the fourier transform domain of the image). This is because most generative models are more concerned with perceptual loss and pixel level reconstruction errors during training, while lack of consistency modeling of spectral characteristics. This results in a systematic difference of the generated image from the real image in high frequency detail, phase distribution or frequency domain energy attenuation pattern. Although the differences are invisible in vision, the method has strong stability and generalization across models, and provides a new break for true and false image identification. Currently, there have been preliminary studies attempting AIGC detection using frequency domain analysis, such as classification by analyzing the fourier spectral energy distribution or local band response of the image. However, these methods are limited to static threshold judgment or shallow feature analysis, lack of fine modeling of frequency domain errors, and are still insufficient in robustness and generalization capability especially in the face of various generation models (e.g., stable dispersion, midjourney, DALL ·e) and post-processing operations (e.g., compression, scaling, filtering). Therefore, a detection mechanism capable of deeply mining AIGC the inherent deviation of the image in the frequency domain is needed to construct a detection frame sensitive to the frequency domain error and having a strong generalization capability. Most of the prior art focuses on the local artifact in the spatial domain, ignores the more stable and universal abnormal mode in the frequency domain, and is difficult to meet the high-precision and high-robustness authenticity identification requirements in practical application. Under the background, a AIGC image authenticity detection scheme based on frequency domain errors is provided, and whether the picture is generated by the AI can be detected more accurately. The image authenticity detection methods AIGC in the prior art mostly rely on spatial domain features or shallow frequency domain analysis, and the methods gradually expose some limitations when facing to the rapidly developed generative model. Although they can provide a degree of accuracy in a specific scenario, in practical applications, particularly when processing images after various generative models and post-processing operations, their generalization ability