AI-Generated Image Detection System and Method Based on Reversible Semantic Reconstruction

Abstract

The present invention relates to a system and method for precise determination of AI-generated images based on reversible semantic restoration, comprising: an image input unit that receives an image to be verified; a first verification module that performs a first determination (PASS, BLOCK, PENDING) by performing a similarity search on a pre-constructed anchor graph based on fingerprints extracted from the input image; a second verification module that analyzes dynamic and geometric differences between an original image and a reconstructed image by performing reversible semantic reconstruction based on a diffusion model utilizing a text prompt on an image for which the judgment is pending (PENDING) as a result of the first determination; and a result analysis and determination unit that performs a final determination on whether an image is generated by synthesizing the analysis results of the second verification module and updates the anchor graph by feeding back the newly identified fingerprints of the forgery pattern. The present invention relates to an AI-generated image precision discrimination system based on reversible semantic restoration, characterized by including an XAI (Explainable Artificial Intelligence) result output module that provides the basis for the final judgment in the form of a visual heatmap and a natural language report.

Description

Technology Field

[0001] The present invention relates to an artificial intelligence-based image analysis technology, and specifically, to a system and method for precisely distinguishing between an AI-generated image and an actual captured image through a reversible semantic reconstruction technique utilizing a diffusion model. More specifically, the invention relates to a system and method for precisely distinguishing between an AI-generated image that hierarchically combines an anchor graph-based first-order high-speed filtering and a diffusion model-based second-order precision analysis, captures the intrinsic characteristics of the generated image through multidimensional analysis techniques such as snap-back dynamics analysis, semantic perception-based reconstruction error (SARE), and Jacobian-spectral indicators, and is equipped with a self-learning mechanism for new patterns and a function to present discrimination grounds based on explainable artificial intelligence (XAI). Background Technology

[0002] The recent rapid advancement of generative AI technology has led to the emergence of high-performance image generation models such as DALL-E, Midjourney, and Stable Diffusion, making it possible to generate high-quality synthetic images that are difficult to distinguish from real photographs. While this technology is being utilized effectively in creative activities and industries, it is simultaneously causing serious social problems, including the dissemination of false information, defamation through deepfakes, copyright infringement, and the manipulation of legal evidence. In particular, as cases of AI-generated images being disguised as real photographs and circulated on social media and news platforms have surged, the demand for technical solutions capable of quickly and accurately determining the authenticity of images has exploded.

[0004] Conventional AI-generated image detection technologies have primarily relied on detecting specific artifacts or frequency patterns that occur during the generation process. For example, techniques that capture checkerboard patterns or abnormal signals in high-frequency ranges appearing in Generative Adversarial Network (GAN)-based images, or supervised learning methods that classify real and generated images by training Convolutional Neural Networks (CNNs), have been widely used. While these approaches demonstrate high detection rates for specific generative models, they have limitations in that detection performance degrades rapidly whenever new generative techniques emerge or existing models are improved. In particular, the latest Diffusion Models generate images that are much more natural than GANs, posing a problem where they are difficult to identify using traditional artifact-based detection methods.

[0006] Furthermore, conventional single-classification model-based systems perform the same high-cost deep learning inference operations on all input images, which limits their computational efficiency in processing large volumes of images in real-time on large-scale platforms. Because they process images uniformly without distinguishing between clearly distinguishable images and ambiguous images on the boundary, they consume unnecessary computational resources and result in longer average processing times. This acts as a serious constraint in practical environments where immediate verification at the time of social media upload or rapid identification in breaking news situations is required.

[0008] Furthermore, most existing systems have limitations in that they present determination results merely as simple probability values ​​(e.g., "87% probability of AI generation") without providing specific grounds for such a judgment. This "black-box" approach makes it difficult to secure user trust and poses a problem in that it is difficult to use as evidence, particularly in situations requiring transparency regarding the basis for determination, such as legal disputes or media reports. Due to the absence of visual and verbal explanations during the determination process regarding which parts of the image are suspicious or which technical characteristics suggest AI generation, it is difficult for even experts—not just non-experts—to verify the validity of the judgment.

[0010] Furthermore, conventional technology relies on static detection models, which presents a limitation in that large amounts of labeled training data must be collected and models retrained whenever a new generative model is released or an existing model is updated. This process typically takes weeks to months, during which time images generated by new generative techniques circulate without being detected. In the current environment where generative AI technology is evolving rapidly, this slow adaptation speed acts as a structural limitation that significantly reduces the effectiveness of detection systems.

[0012] Accordingly, there is an urgent need to develop a next-generation AI-generated image discrimination system equipped with a principle-based discrimination technique capable of capturing the intrinsic characteristics of AI-generated images without relying on artifacts of specific generative models, an efficient hierarchical verification structure that rapidly processes clear cases and precisely analyzes only ambiguous ones, explainability that clearly presents the basis for discrimination visually and linguistically, and a self-learning mechanism that automatically adapts to new generation patterns.

[0014] The present invention aims to solve social problems arising from the rapid spread of AI-generated images, and specifically, may address the following technical challenges.

[0016] First, conventional AI-generated image recognition technology relied on passive methods to detect known artifacts or frequency patterns of specific generative models. Consequently, it suffered from a limitation in that detection performance deteriorated rapidly whenever a new generative model emerged or an existing model was improved. This could lead to problems that limit practicality as the technology fails to keep pace with the rapid evolution of generative technologies.

[0018] Second, because existing systems performed the same high-cost deep learning inference on all input images, there was a problem where computational costs and processing latency increased excessively in social media or news platform environments requiring large-scale real-time processing. This can be a factor that significantly limits applicability in actual industrial settings.

[0020] Third, since conventional technology provided identification results solely as simple probability values ​​or binary classifications, a "black box" problem existed where users could not understand the rationale behind why a particular image was determined to be AI-generated. This can be a serious obstacle to utilizing the technology as legal evidence or securing user trust.

[0022] Fourth, existing systems have a static structure where performance is fixed after deployment, so whenever a new generation technique emerges, it is necessary to collect a large amount of labeled data and retrain the model, which poses a problem requiring several months of time and enormous costs. Prior art literature

[0023] Korean Published Patent KR 10-2022-0065705 Korean Registered Patent KR 10-2896932 B1 US Registered Patent US 12039615 The problem to be solved

[0024] Therefore, the present invention has been devised to solve the aforementioned conventional problems, and aims to establish a universal discrimination system that is robust even in post-processing by precisely analyzing the 'information reversibility difference' existing between a generated image and an actual captured image.

[0025] To this end, the present invention aims to achieve the following key objectives.

[0026] First, it ensures detection reliability through the principle of 'Reversible Semantic Reconstruction.' This is based on the scientific hypothesis that while actual images project the infinite complexity of the real world and suffer significant loss of visual information when summarized into short text captions, AI-generated data exists within a data manifold derived from text information from the outset, thereby exhibiting much higher semantic consistency with the original upon reconstruction.

[0027] Second, the robustness of detection is enhanced through 'Manifold Geometry Analysis.' By introducing mathematical frameworks such as the snap-back dynamics of diffusion models and Jacobian-spectral analysis, false positives are minimized by tracking geometric convergence in high-dimensional space, even if pixel information is partially corrupted.

[0028] Third, it maximizes 'computational efficiency and industrial applicability.' Instead of regenerating all areas, it adopts a hierarchical structure consisting of a first stage of high-speed filtering and a second stage of precision inspection. By applying Importance Mask (IML) and Adaptive Routing (CATImage) technologies, it performs high-precision diagnosis using only the minimum resources required for discrimination, thereby drastically reducing operating costs.

[0029] Fourth, there are 'Explainable Detection' and 'System Self-Evolution'. It establishes a virtuous cycle structure in which the system improves its performance by presenting suspected forgery areas within an image as a heatmap, explaining the evidence in natural language, and adding newly discovered forgery patterns to an anchor graph. means of solving the problem

[0030] A reversible semantic restoration-based AI-generated image precision discrimination system according to one aspect of the present invention for solving the above problem may include: an image input unit that receives an image to be verified; a first verification module that performs a first judgment by performing a similarity search in a pre-constructed anchor graph based on feature points extracted from the input image; a second verification module that analyzes dynamic and geometric differences between an original image and a reconstructed image by performing reversible semantic restoration based on a diffusion model using a text prompt on an image for which the judgment of the first judgment result is withheld; a result analysis and judgment unit that performs a final judgment on whether an image is generated by synthesizing the analysis results of the second verification module and updates the anchor graph by feeding back the feature points of the newly identified forgery pattern; and an XAI result output module that provides the basis for the final judgment in the form of a visual heatmap and a natural language report.

[0032] The feature point of the first verification module may be a vectorized representation of at least one of the color histogram, texture gradient, edge curvature, and micro-frequency characteristics unique to the generation model of the image to be verified, and the anchor graph may be an EMR anchor graph having the feature points of the actual captured image and the previously identified AI-generated image as nodes, and the first verification module may classify into immediate trust, immediate blocking, or two-stage transfer based on the distance between the input image and the anchor point through nearest neighbor search on the EMR anchor graph.

[0034] The second verification module above can optimize the analysis area and computation path by generating an importance mask using a deep learning-based protrusion map to select partial areas, and applying an adaptive routing technique that dynamically branches the computation intensity of the reconstruction step of the diffusion model according to the complexity and resolution of the image.

[0036] The second verification module may include an image-adaptive prompt learning engine that minimizes information entropy by fine-tuning optimal prompt tokens for a single input image, and the text prompt extracted therefrom may be used as an input value for the diffusion model.

[0038] The second verification module can analyze at least one multidimensional indicator among snapback dynamics analysis, which confirms the deterministic regression characteristics of the generative model by tracking the trajectory and curvature of the reconstruction quality according to the change in noise intensity of the diffusion model in performing the reversible semantic restoration; semantic cognitive-based reconstruction error calculation, which measures the semantic shift between the original image and the reconstructed image in the CLIP feature space; and Jacobian-Spectral lower bound indicator calculation through singular value decomposition of the Jacobian matrix.

[0040] The above result analysis and judgment unit can self-learn the system by applying a hierarchical proxy loss technique to cluster feature points of newly discovered ambiguous data or new AI generation patterns by similar generation techniques and adding them as new anchors to the anchor graph of the first verification module.

[0042] The XAI result output module described above can provide the user with a SARE-based heatmap that visualizes areas where feature differences measured based on the SARE indicator are concentrated, and an MLLM report that infers the basis for discrimination in natural language based on quantitative data of curvature, SARE values, and Jacobian spectrum values ​​of the snapback dynamics analysis using a large-scale multimodal model.

[0044] A method for precise image discrimination performed in an AI-generated image discrimination system based on reversible semantic restoration according to another aspect of the present invention for solving the above problem may include: receiving an image to be verified and extracting statistical and geometric feature points from the image; searching for the extracted feature points on a pre-established anchor graph to perform a first judgment on the image to be verified; generating a text prompt for an image classified as judgment pending as a result of the first judgment and performing reversible reconstruction using a diffusion model to perform a precise analysis of multidimensional index error with the original image; finally determining whether it is a generated image based on the precisely analyzed multidimensional index and updating the newly identified pattern on the anchor graph; and outputting the final judgment result and the basis for judgment as explainable AI results in the form of a visual heatmap and a natural language report.

[0046] The above-described precision analysis step may include a step of optimizing resources by selecting important regions of the image with judgment pending and allocating computational intensity according to complexity; a step of generating prompt tokens optimized for the image with judgment pending through image adaptive prompt learning; a step of analyzing shrinkage mapping and snapback convergence during the process of reconstructing the image by inputting the generated prompts into a diffusion model; and a step of extracting semantic cognitive-based reconstruction errors and Jacobian spectral indicators using a CLIP feature space.

[0048] The output step may include a step of generating a highlight as a SARE-based heatmap by measuring a specific area where a semantic shift occurred between the original image and the reconstructed image based on distance, and a step of generating a natural language text report with logical findings pointing out structural defects in the image and characteristics of the generating model by analyzing the dynamic trajectory and quantitative values ​​of the multidimensional index error in a multimodal model. Effects of the invention

[0049] According to the present invention, by capturing the deterministic regression phenomenon, which is an intrinsic characteristic of AI-generated images, through a unique mechanism of reversible semantic restoration via a diffusion model, a significant technical effect can be achieved by providing a universal discrimination capability that operates independently of the generation technique, compared to conventional artifact detection methods for specific generation models. In particular, based on a mathematically provable principle that AI-generated images strongly converge to the original when reconstructed by a diffusion model because they are generated in a finite-dimensional latent space, whereas information loss is inevitable when reconstructing actual captured images due to the infinite complexity of nature, this invention can solve the technical challenge of a sharp decline in detection performance upon the emergence of new generation models, which could not be resolved by existing technologies.

[0051] In addition, the two-stage hierarchical verification architecture of the present invention enables an adaptive computational allocation strategy that immediately processes more than 90% of clear cases within an average of 5ms through anchor graph-based primary filtering and transfers only 10% of ambiguous cases to diffusion model-based secondary precision analysis. This reduces the overall average processing time to 0.5-2 seconds compared to the conventional method of applying the same high-cost deep learning inference to all images, thereby providing advantageous effects such as a 5 to 30-fold improvement in processing speed and a 10-fold reduction in GPU resource usage. This can dramatically improve the practical applicability in social media platform environments that require large-scale real-time processing.

[0053] Furthermore, the image adaptive prompt learning engine of the present invention automatically generates an optimal prompt token that minimizes information entropy for each input image, thereby improving reconstruction quality by 15% to 20% compared to conventional methods using fixed prompts, and through this, can provide a synergistic effect that improves final discrimination accuracy by significantly improving the measurement precision of multidimensional indicators such as snapback dynamics analysis, semantic recognition-based reconstruction error, and Jacobian spectral indicator.

[0055] In addition, the hierarchical proxy loss-based self-learning mechanism of the present invention automatically clusters newly discovered AI-generated patterns and adds them to an anchor graph, thereby replacing the manual retraining process, which used to take several months when a new generative model appeared, with an automatic adaptation process of one to two weeks, which has the technical effect of significantly reducing system maintenance costs and continuously improving performance.

[0057] Furthermore, the multimodal XAI system of the present invention can dramatically improve the transparency and comprehensibility of the basis for judgment compared to conventional black-box judgment methods by visually highlighting areas where semantic changes are concentrated through a SARE-based heatmap and converting quantitative indicators into logical natural language reports using a large-scale multimodal model. This can create a complex effect that secures the potential for use as legal evidence, enhances user trust, and provides educational value by enabling even non-experts to learn the characteristics of AI-generated images.

[0058] Meanwhile, the effects obtainable from the present invention are not limited to those mentioned above, and other unmentioned effects will be clearly understood by those skilled in the art to which the present invention belongs from the description below. Brief explanation of the drawing

[0059] The following drawings attached to this specification illustrate preferred embodiments of the present invention and serve to further enhance understanding of the technical concept of the present invention together with the detailed description of the invention; therefore, the present invention should not be interpreted as being limited only to the matters described in such drawings. FIG. 1 is a table showing the main functions and key technical elements of each stage according to an embodiment of the present invention, illustrating the immediate determination or transfer process through statistical and geometric feature point extraction and anchor graph search in the first stage high-speed filtering stage, and the semantic and structural difference measurement process utilizing resource optimization, IAPL prompt learning, and diffusion model reconstruction dynamics in the second stage precise verification stage. FIGS. 2a and 2b are overall configuration diagrams of an integrated system according to an embodiment of the present invention, illustrating an image input unit, fingerprint extraction and primary judgment based on an EMR anchor graph of a first verification module, reversibility analysis through IAPL prompt learning and diffusion reconstruction engine of a second verification module, final authenticity determination and self-learning feedback of a result analysis and judgment unit, and a configuration for generating a heatmap and natural language report of an XAI result output module. FIGS. 3a and 3b are hierarchical verification integration flowcharts according to an embodiment of the present invention, illustrating the process of extracting feature points and verifying clustering within an anchor graph after receiving an image, immediately terminating the judgment if it belongs to a clear cluster or transferring it to a two-stage precise verification layer if it is ambiguous, performing precise analysis of the SARE index (760) and snapback dynamics through adaptive prompt optimization and image reconstruction, final judgment, and generating an XAI report. FIG. 4 is an example of a SARE-based detection heatmap and report according to an embodiment of the present invention, illustrating an original image, a SARE-based analysis heatmap visualizing the area where a semantic shift occurred between the original and the reconstructed image, SARE values, quantitative data such as snapback curvature and Jacobian spectrum values, and analysis indicators and logical findings including logical discrimination grounds created by MLLM. FIG. 5 is a graph for explaining the snapback dynamics of a diffusion model analyzed in a second verification module according to an embodiment of the present invention, and illustrates the difference in the trajectory of reconstruction quality and convergence that appears when the actual captured image and the AI-generated image each revert to the original according to the change in injected noise intensity. FIG. 6 is a geometric schematic diagram showing the EMR anchor graph structure of a first verification module according to an embodiment of the present invention, visually illustrating the first determination branching criteria based on the nearest neighbor search distance between actual image nodes and AI-generated image nodes clustered in a high-dimensional space and a newly input image to be verified. FIG. 7 is a block diagram illustrating the process of calculating a reversible reconstruction and semantic recognition-based reconstruction error performed in a second verification module according to an embodiment of the present invention, and illustrates the process in which a reconstructed image generated through image adaptive prompt learning and an original image are each embedded into a CLIP feature space to quantify semantic shifts. FIG. 8 is a feedback loop flowchart illustrating the self-learning process of a system according to an embodiment of the present invention, showing the process in which a feature point of OOD data newly identified through a two-stage precision inspection is updated as a new node in the anchor graph of the first verification module through a hierarchical proxy loss technique. Specific details for implementing the invention

[0060] The above objects, other objects, features, and advantages of the present invention will be easily understood through the following preferred embodiments associated with the accompanying drawings. However, the present invention is not limited to the embodiments described herein and may be embodied in other forms. Rather, the embodiments introduced herein are provided to ensure that the disclosed content is thorough and complete and to ensure that the spirit of the invention is sufficiently conveyed to a person skilled in the art.

[0061] In this specification, when a component is described as being on another component, it means that it may be formed directly on the other component or that a third component may be interposed between them. Also, in the drawings, the thicknesses of the components are exaggerated for the effective description of the technical content.

[0062] The embodiments described herein will be explained with reference to cross-sectional and / or plan views, which are exemplary illustrations of the invention. In the drawings, the thicknesses of films and regions are exaggerated for effective explanation of the technical content. Accordingly, the shapes of the exemplary drawings may be modified by manufacturing techniques and / or tolerances, etc. Accordingly, the embodiments of the invention are not limited to the specific shapes depicted but include variations in shape produced according to the manufacturing process. For example, a region depicted as a right angle may be rounded or have a certain curvature. Accordingly, the regions illustrated in the drawings have properties, and the shapes of the regions illustrated in the drawings are intended to illustrate specific shapes of the regions of the device and are not intended to limit the scope of the invention. Although terms such as first, second, etc., have been used to describe various components in the various embodiments of this specification, these components should not be limited by such terms. These terms are used merely to distinguish one component from another. The embodiments described and illustrated herein also include their complementary embodiments.

[0063] The terms used herein are for describing the embodiments and are not intended to limit the invention. In this specification, the singular form includes the plural form unless specifically stated otherwise in the text. As used herein, 'comprises' and / or 'comprising' do not exclude the presence or addition of one or more other components to the mentioned components.

[0064] In describing the specific embodiments below, various specific details have been included to explain the invention more specifically and to aid understanding. However, a reader with sufficient knowledge in the art to understand the invention will recognize that it can be used without these various specific details. In some cases, it is noted in advance that commonly known aspects that are not significantly related to the invention have been omitted to prevent unnecessary confusion in describing the invention.

[0066] The present invention relates to a system and method for precisely determining the authenticity of AI-generated images, and in particular, to a technology for capturing the inherent mathematical characteristics of generated images by utilizing the reversible semantic reconstruction characteristics of a diffusion model.

[0068] Due to the recent rapid advancement of generative AI technology, high-quality AI-generated images that are difficult to distinguish from actual captured images are being produced in large quantities, causing serious social problems such as the spread of false information, copyright infringement, and the falsification of legal evidence. Existing AI image identification technologies have primarily relied on detecting specific artifact patterns left by generative models; however, this approach has limitations, as detection performance degrades rapidly when new generative models emerge or post-processing is applied to images.

[0070] To solve these technical challenges, the present invention proposes an active verification method that moves away from a passive detection method of finding surface traces of generated images and instead verifies essential characteristics by reversing the generation process itself. The core insight is based on the mathematical principle that "while images generated by AI strongly converge to the original when reconstructed by the same generative model, actual captured images cannot be perfectly represented in a finite-dimensional latent space, resulting in semantic loss during reconstruction." Furthermore, the present invention achieves both computational efficiency and discrimination accuracy simultaneously through a hierarchical architecture of first-order filtering that processes clear cases at high speed and second-order verification that precisely analyzes ambiguous cases, and implements a self-evolution mechanism that continuously improves system performance by automatically learning newly discovered forgery patterns.

[0072] Referring to FIG. 1 and FIG. 2, an AI-generated image precision discrimination system (100) based on reversible semantic restoration according to one embodiment of the present invention may include an image input unit (110), a first verification module (120), a second verification module (130), a result analysis and judgment unit (140), and an XAI result output module (150).

[0074] The image input unit (110) is a component that receives an image to be verified into a system and can support various image formats (e.g., JPEG, PNG, WebP, HEIC, etc.) and can perform preprocessing on the received image. Specifically, the image input unit can extract metadata of the image (e.g., EXIF, IPTC information) and use it for analysis, normalize the resolution of the image (e.g., adjust from 512×512 pixels to a range of 2048×2048 pixels), and perform color space conversion (e.g., conversion between RGB, LAB, and HSV) as needed. The preprocessed image data can be transmitted to a subsequent verification module and used as basic data for feature point extraction and discrimination analysis.

[0076] The first verification module (120) is a high-speed filtering layer that performs a first determination by performing a similarity search in a pre-constructed anchor graph based on fingerprints extracted from an input image. The main purpose of this module is to significantly reduce the overall computational load of the system by immediately processing cases that can be clearly distinguished among the entire input image. The first verification module (120) may include a fingerprint extraction engine, an EMR (Efficient Manifold Representation) anchor graph, and a determination logic based on Approximate Nearest Neighbor (ANN).

[0078] The above feature point extraction engine performs the role of compressing the unique visual characteristics of an input image into a multidimensional vector. In one embodiment of the present invention, the feature point may be a vectorized representation of at least one of a color histogram, texture gradient, edge curvature, and fine frequency characteristics unique to the generating model. For example, the color histogram feature may be represented as a total 128-dimensional vector by generating a 32-bin histogram for each RGB channel and a 32-bin histogram for the brightness channel of the LAB color space. The texture gradient feature may be composed of a 64-dimensional vector by applying a Sobel filter to extract the gradient magnitude distribution in 8 directions or by calculating response statistics from a Gabor filter bank (e.g., 4 scales and 8 directions). The edge curvature feature may be represented as a 48-dimensional vector by generating a curvature histogram after Canny edge detection and calculating a density map through Harris corner detection. The unique fine frequency features of the generative model can be constructed into 256 dimensions by analyzing the energy distribution in the high-frequency band (e.g., above 0.4 Nyquist frequency) after a 2D Fast Fourier Transform (FFT) or by extracting coefficient statistics from wavelet transforms (e.g., Haar, Daubechies wavelets), which can capture characteristic frequency patterns (e.g., checkerboard artifact frequencies) generated by GANs or Diffusion models. By combining these various features, a total of 496-dimensional feature vectors can be generated, but are not limited thereto, and the types and dimensions of the features can be adjusted depending on the application field or computational resources.

[0080] The above EMR anchor graph is a graph structure that efficiently represents a high-dimensional feature space and can have feature points of actual captured images and previously identified AI-generated images as nodes. Referring to FIG. 6, the anchor graph can include, for example, 50,000 feature points of actual images and 50,000 feature points of AI-generated images as nodes, and each node can be connected by edges with k-nearest neighbors (e.g., k=15). Each node can have attributes such as a 496-dimensional feature vector, a class label (REAL or FAKE), a generation model type (DALL-E, Midjourney, Stable Diffusion, etc., where applicable), and a confidence score (ranging from 0.0 to 1.0). The above anchor graph can reduce a high-dimensional feature space (e.g., 496 dimensions) to a low-dimensional embedding space (e.g., 64 dimensions) through manifold learning techniques such as the Laplacian Eigenmap, and in this process, a clear manifold separation between the actual image and the generated image can be induced.

[0082] The first verification module (120) can search for the feature vector of the input image on the anchor graph to find the nearest neighbors and perform a first judgment based on the distance and the class distribution of the neighbors. Specifically, k (e.g., k=20) nearest anchors of the input feature vector can be searched using a high-speed ANN algorithm such as the HNSW (Hierarchical Navigable Small World) index. The judgment criteria can be set as follows. A PASS (Real Image) judgment can be made when 18 or more of the 20 neighbors belong to the REAL class and the average distance is less than a threshold (e.g., normalized L2 distance 0.15). A BLOCK (Explicit AI Generated) judgment can be made when 18 or more of the 20 neighbors belong to the FAKE class and the average distance is less than a threshold (e.g., 0.15), and the distance from a specific generative model cluster center is less than a stricter threshold (e.g., 0.10). A PENDING decision can be made in all cases that do not fall under the above conditions, namely when the image is located in a boundary region (e.g., between an average distance of 0.15 and 0.35) or when the ratio of REAL and FAKE neighbors is ambiguous (e.g., between 60:40 and 40:60). Through this initial decision, more than 90% of the entire input image can be processed immediately, and high-speed screening can be achieved with an average processing time of 5ms per image.

[0084] The second verification module (130) is a component that performs reversible semantic restoration based on a diffusion model using text prompts on images classified as PENDING in the first judgment result, and precisely analyzes the dynamic and geometric differences between the original image and the reconstructed image. Referring to FIGS. 3 and FIGS. 7, the second verification module (130) may include an adaptive routing system, an image adaptive prompt learning (IAPL) engine, a diffusion model-based reconstruction engine, and a multidimensional indicator analysis unit.

[0086] The adaptive routing system described above performs the role of optimizing resources by dynamically adjusting the analysis area and computational intensity according to the complexity and resolution of the image. In one embodiment of the present invention, the adaptive routing system can generate an Importance Mask (IML) by utilizing a deep learning-based saliency map. For example, an importance map of the same size as the input image (values ​​ranging from 0.0 to 1.0) can be generated by using a pre-trained ResNet-50 backbone and applying the Grad-CAM++ algorithm. The importance mask selects the top K% of pixels from the saliency map, but the value of K can be adjusted according to the complexity of the image. For example, for a simple image, K can be set to 30 to focus on the central area, and for a complex image, K can be set to 60 to perform extensive analysis. Morphological operations are applied to the selected pixels to expand the connected area, guarantee a minimum patch size (e.g., 128×128 pixels), and up to four independent regions can be selected.

[0088] The adaptive routing system described above can also measure the complexity of an image and branch the computation path. Indicators of complexity may include edge density (Canny edge pixel ratio), color diversity (total number of pixels relative to the number of unique colors), and texture complexity (entropy of the gray level co-occurrence matrix GLCM). The computation path can be determined based on the measured complexity; for example, for a simple image with a complexity of less than 0.3, the processing time may take approximately 2-3 seconds by setting the diffusion step to 20 steps, using DDIM sampling, and analyzing 1-2 patches. For an image of medium complexity with a complexity of 0.3 or higher but less than 0.7, the processing time may take approximately 5-7 seconds by setting the diffusion step to 50 steps, using DDPM sampling, and analyzing 2-3 patches. For complex images with a complexity of 0.7 or higher, setting the diffusion step to 100 steps, adding guidance scale adjustment to DDPM sampling, and analyzing 3-4 patches can take approximately 10-15 seconds. Through this adaptive routing, the average computation time can be reduced by about 60% while maintaining discrimination accuracy.

[0090] The image adaptive prompt learning (IAPL) engine described above is a component that automatically generates text prompts optimized for each input image. Referring to FIG. 7, the IAPL engine can first extract 512-dimensional image embeddings from the input image (700) using a CLIP image encoder. Then, a pre-trained image-caption model (e.g., BLIP-2) can be used to generate a basic description, and an initial prompt such as "A photo of a person standing in a park" can be generated. The IAPL engine (710) can optimize learnable prompt tokens based on this initial prompt. Specifically, the prompt structure can be composed of a [CLS] token, eight learnable prefix tokens, a basic description (fixed) generated by BLIP-2, eight learnable suffix tokens, and a [SEP] token.

[0092] Optimization of the aforementioned prompt token can proceed in the direction of minimizing the composite loss function. The loss function can be expressed as L = α × L_reconstruction + β × L_entropy + γ × L_semantic, where L_reconstruction represents the Mean Squared Error (MSE) between the reconstructed image and the original, L_entropy represents uncertainty in the CLIP feature space, and L_semantic represents the loss of semantic consistency. The weights can be set, for example, α=1.0, β=0.3, and γ=0.5, but are not limited thereto. The optimization process can be performed for up to 50 iterations using the AdamW optimizer with a learning rate of 0.001, and may be terminated early if the loss change falls below 0.001 for five consecutive iterations. Through this IAPL technique, reconstruction quality can be improved by approximately 15-20%, and semantic alignment is enhanced, which can increase the accuracy of subsequent analysis.

[0094] The above-described diffusion model-based reconstruction engine (720) receives an optimized prompt as input and performs the role of reconstructing the original image. In one embodiment of the present invention, the reconstruction engine may use a pre-trained diffusion model (e.g., Stable Diffusion), and the noise schedule may be defined as a cosine schedule. The reconstruction process may include a denoising process that gradually adds noise to the original image and then removes it, and in this process, the reconstruction quality may be measured at several time points (e.g., t ∈ {0, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100}). At each time t, the image x_t with added noise can be calculated as √(α_t) × x_0 + √(1-α_t) × ε, where x_0 is the original image, ε is Gaussian noise, and α_t is the noise schedule parameter. The image x_0 reconstructed through the denoising process can be compared with the original image to calculate quality metrics (e.g., PSNR, SSIM, LPIPS).

[0096] The above-described multidimensional indicator analysis unit is a component that quantifies the difference between the original image (710) and the reconstructed image (730) from various perspectives. In one embodiment of the present invention, the analysis unit may perform at least one of snap-back dynamics analysis, semantic perception-based reconstruction error (SARE) calculation, and Jacobian-spectral indicator calculation.

[0097] The above Semantic Recognition-based Reconstruction Error (SARE) is an indicator that measures the semantic shift between the original image and the reconstructed image in the CLIP feature space. Simple pixel-level errors are easily distorted by resizing or compression. To address this, the present invention introduces a reconstruction error measured in the semantic encoder feature space, which can be calculated using the following [Equation 1].

[0098] [Mathematical Formula 1]

[0099]

[0100] Here is a semantic encoder. While real images exhibit significant semantic shifts during reconstruction because captions cannot capture all the details of reality, AI-generated images possess a 'reversible advantage' that allows them to be restored close to the original using only caption information. This metric is highly robust against domain shifts and post-processing.

[0101] Here is a semantic encoder (e.g., CLIP image encoder), I is the original image, and \hat{I} is the reconstructed image. The actual image exhibits a large semantic shift during reconstruction because the caption does not capture all the details of reality, whereas the AI-generated image possesses a 'reversible advantage' that allows it to be restored close to the original using only the caption information. This metric has characteristics that are very robust against domain shift or post-processing. Referring to FIG. 7, the original image I and the reconstructed image \hat{I} are each passed through a CLIP image encoder (740) to be converted into feature vectors (e.g., 768 dimensions when using ViT-L / 14), and then the error can be calculated through the above [Equation 1]. In one embodiment of the present invention, a spatial SARE map can be generated by dividing the image into patch units (e.g., 16×16 pixel patches) and calculating SARE for each patch.

[0102] Referring to Fig. 5, the snapback dynamics analysis is a process of verifying the deterministic regression characteristics of the generative model by tracking the trajectory and curvature of the reconstruction quality according to changes in the noise intensity of the diffusion model. When the reconstruction quality Q(t) is measured at each time point t, the curvature κ of this quality trajectory can be calculated. The curvature can be defined as κ = |Q''(t)| / (1 + Q''(t)²)^(3 / 2), where Q'(t) and Q''(t) are the first and second derivatives of the quality function, respectively. In the case of AI-generated images, the change in curvature is gradual and the average curvature may be above a threshold (e.g., 0.12), and a rapid quality recovery (snapback) may be observed, particularly in the interval between t=40 and 60. On the other hand, in the case of real images, the change in curvature is irregular and the average curvature may be below a threshold (e.g., 0.05). In addition, the contraction mapping characteristics can be verified by calculating the distance between consecutive reconstructions d(t) = ||x_0^(t) - x_0^(t+10)||₂. For AI-generated images, d(t) decreases exponentially and the convergence rate λ may be greater than a threshold (e.g., 0.8), whereas for real images, λ may be less than a threshold (e.g., 0.4).

[0104] The above Semantic Recognition-based Reconstruction Error (SARE) is an indicator that measures the semantic shift between the original image and the reconstructed image in the CLIP feature space. Referring to FIG. 7, the original image x_0 and the reconstructed image x_0 can each pass through a CLIP image encoder (740) to be converted into feature vectors f_orig and f_recon (each 768 dimensions, when using ViT-L / 14). SARE can be calculated as 1 - cosine_similarity(f_orig, f_recon) and can have a value between 0 and 2. The closer to 0, the more semantically identical they are, and the closer to 2, the more completely different they may represent. In one embodiment of the present invention, a spatial SARE map can be generated by dividing the image into patch units (e.g., 16×16 pixel patches) and calculating SARE for each patch. For example, if the image is divided into 32×32 patches, a 32×32 size SARE map can be generated. In addition, multi-layer SARE analysis can be performed by utilizing the multi-layer features of CLIP. For example, low-level (texture, color) changes can be measured with Layer 8 features, mid-level (object shape) changes can be measured with Layer 16 features, and high-level (semantic concept) changes can be measured with Layer 24 features.

[0105] The total SARE can be calculated as a weighted average, such as 0.3×SARE_L8 + 0.3×SARE_L16 + 0.4×SARE_L24, but is not limited thereto. In the case of real images, the SARE may be below a threshold (e.g., 0.15) and spatial variation may be uniform, whereas in the case of AI-generated images, the SARE may be above a threshold (e.g., 0.35) and abrupt variation may occur in specific areas (e.g., object boundaries, complex textures).

[0107] The above Jacobian-Spectral metric is an indicator that quantifies the degree to which an image deviates from the data manifold by analyzing the local sensitivity of the diffusion model. The Jacobian matrix J can be defined as ∂x_0 / ∂x_t and can be calculated at a specific time point (e.g., t=50). In practice, since calculating the Jacobian for the entire image is computationally burdensome, a patch-by-patch approximation (e.g., a 128×128 pixel patch) can be used. Singular Value Decomposition (SVD) can be performed on the calculated Jacobian matrix to decompose it into J = U Σ V^T, yielding singular values ​​σ₁ ≥ σ₂ ≥ ... ≥ σ. From this, various spectral analysis metrics can be derived. The condition number can be calculated as κ = σ_max / σ_min, and while real images may exhibit high sensitivity above a threshold (e.g., 100), AI-generated images may exhibit stable reconstruction characteristics below a threshold (e.g., 50). The effective rank can be calculated as r_eff = (Σσ)² / Σ(σ²), and while real images may have rich information of about 70% or more of the total dimension (e.g., r_eff > 0.7×n), AI-generated images may have less than about 40% (e.g., r_eff 0.7), and if the distance from the existing generative model cluster is greater than a threshold (e.g., 0.25), the image may be estimated as a new generative technique.

[0109] The result analysis and judgment unit (140) can cluster newly discovered OOD data by applying a Hierarchical Proxy Loss technique. The Hierarchical Proxy Loss can be defined as L_proxy = L_intra + λ × L_inter, where L_intra is the loss within the class calculated as Σ ||f_i - c_k||²(i ∈ class k), and L_inter is the loss between the classes calculated as -Σ_k Σ_j≠k max(0, m - ||c_k - c_j||²). Here, f_i is the feature vector of image i, c_k is the proxy center of class k, m is the margin (e.g., 0.5), and λ is the balance weight (e.g., 0.3). By minimizing this loss function, the cohesion within the class and the separation between the classes can be optimized simultaneously.

[0111] A clustering algorithm (e.g., HDBSCAN) can be applied to the feature vectors of N images detected by OOD. The parameters of HDBSCAN can be set, for example, min_cluster_size=5, min_samples=3, and metric='euclidean', through which K clusters (new generation technique types) can be identified. For each cluster C_k, the cluster center anchor_k = mean(features in C_k) is calculated, and representative samples can be selected. For example, the 5 samples closest to the cluster center and 3 samples at the cluster boundary (to ensure diversity) can be selected. New nodes can be created for the selected samples and added to the anchor graph, and each node can have attributes such as a feature vector, a label (FAKE), a generation model type (e.g., "Unknown_Type_K"), and an initial confidence (e.g., 0.6, which can be gradually increased). The k nearest neighbors (e.g., k=15) of a new node are connected, and the neighbor relationships of existing nodes can be updated. Additionally, the Laplacian eigenmap can be recalculated to fine-tune the embedding space; this is performed using an incremental update method, which can reduce the cost of rebuilding the entire graph.

[0113] The above anchor graph update can be performed when specific trigger conditions are satisfied. For example, the update may be triggered if 100 or more OOD samples are accumulated or if 7 days have passed. A validation process may be performed prior to the update, and the update may be committed only if performance is measured on a holdout validation set and accuracy improves by a certain level (e.g., 1%) or more and the increase in false positives is within an acceptable range (e.g., 0.5%). Through this self-learning mechanism, the system can automatically adapt without manual retraining when a new generative model emerges, and performance can continuously improve over time.

[0115] The above XAI (Explainable Artificial Intelligence) result output module (150) is a component that provides the basis for the final judgment in the form of a visual heatmap and a natural language report. Referring to FIG. 4, the XAI result output module (150) may include a SARE-based heatmap generator and an MLLM (Large-scale Multimodal Model)-based report generator.

[0117] The SARE-based heatmap generator described above serves to visualize areas where semantic changes between the original image and the reconstructed image are concentrated. First, a 32×32 SARE map can be generated by dividing the image into patches (e.g., 32×32 patches, each 16×16 pixels) and calculating SARE by extracting CLIP features for each patch. The generated SARE map can be visualized through color mapping. For example, after normalizing the SARE values ​​(SARE_norm = (SARE - min) / (max - min)), a Jet colormap is applied, which is significantly lower than blue (SARE 100). The effective dimension is only 35% of the total dimension, and the spectral entropy of 0.46 indicates that singular values ​​are concentrated in a few principal components. Interpretations such as "This strongly suggests that the image was generated in a low-dimensional latent space" may be included. In the Suspicious Areas section, specific location information may be provided, such as "As a result of heatmap analysis, texture inconsistencies were found in the top left (coordinates: 50-200, 50-180), unnatural object boundaries were found in the central area (coordinates: 250-400, 300-450), and fine repeating patterns (suspected generative model artifacts) were found in the background area." In the Conclusion and Recommendations section, a comprehensive opinion and recommendations may be presented, such as "Based on comprehensive analysis, this image is determined to be an AI-generated image with 78% confidence. In particular, numerous characteristic patterns of the diffusion model series (Stable Diffusion estimation) were observed. Further verification is required when using for important purposes (legal evidence, media reports, etc.), and verification of the original metadata and source is recommended."

[0119] Through these XAI results, users can transparently understand the basis of the system's judgment, and even non-experts can learn the characteristics of AI-generated images and obtain logical opinions that can be used as legal evidence.

[0121] Referring to FIG. 3, an AI-generated image precision discrimination method according to one embodiment of the present invention can be performed in the following steps.

[0123] First, a step (a) of receiving an image to be verified and extracting statistical and geometric feature points from the image may be performed. In step (a), the image to be verified is received through an image input unit (110), and preprocessing operations (metadata extraction, resolution normalization, color space conversion, etc.) may be performed. Then, a multidimensional feature vector (e.g., 496 dimensions) including a color histogram, texture gradient, edge curvature, and fine frequency characteristics unique to the generating model may be generated through the feature point extraction engine of the first verification module (120).

[0125] Next, step (b) may be performed to search for the extracted feature points on a pre-constructed anchor graph and perform a first judgment (PASS, BLOCK, PENDING) on the image to be verified. In step (b), k (e.g., k=20) nearest anchors of the input feature vector are searched using the HNSW index, and a judgment of PASS, BLOCK, or PENDING may be made based on the class distribution and average distance of the neighbors. If judged as PASS or BLOCK, the result may be output immediately and the process may be terminated, and if judged as PENDING, the process may proceed to the next step.

[0127] For images classified as PENDING, step (c) may be performed to generate a text prompt and perform reversible reconstruction using a diffusion model to precisely analyze the multidimensional index error with respect to the original image. Step (c) may include several sub-steps.

[0129] First, a step of optimizing resources by selecting the region of interest (ROI) of the image with the judgment pending and assigning computational intensity according to complexity may be performed. In this step, a prominence map may be generated to create an importance mask (IML), and the complexity of the image may be measured to determine a computational path (e.g., 20 steps, 50 steps, or 100 steps).

[0131] Next, a step of generating a prompt token optimized for the image with the judgment pending may be performed through image adaptive prompt learning (IAPL). In this step, an initial prompt is generated using a CLIP image encoder (740) and BLIP-2, and a learnable prefix and suffix token is optimized to minimize a composite loss function that considers reconstruction quality, entropy, and semantic consistency.

[0133] Next, a step of analyzing contraction mapping and snap-back convergence can be performed during the process of reconstructing an image by inputting the generated prompt into a diffusion model. In this step, the trajectory of the reconstruction quality is tracked by adding and removing noise at multiple time points, and the curvature and convergence speed are calculated to verify the deterministic regression characteristics of the generative model.

[0135] Finally, a step of extracting semantic cognitive-based reconstruction error (SARE) and Jacobian-spectral indicators using the CLIP feature space can be performed. In this step, CLIP feature vectors of the original image and the reconstructed image are extracted to calculate cosine similarity (750), patch-specific SARE maps are generated, and SVD of the Jacobian matrix is ​​performed to calculate the condition number, effective dimension, and spectral entropy.

[0137] Next, step (d) can be performed to finally determine whether an image is generated based on the above-detailed multidimensional indicators and to update the newly identified pattern to the anchor graph. In step (d), the final score S_fake is calculated by weighted summing of the snapback, SARE, and Jacobian-Spectral indicators, and a determination of FAKE, REAL, or UNCERTAIN can be made based on this. Additionally, a hierarchical proxy loss technique can be applied to the new pattern detected by OOD to cluster it, and representative samples can be selected and added as new nodes to the anchor graph.

[0139] Finally, step (e) of outputting the above final judgment result and judgment basis as an explainable artificial intelligence (XAI) result in the form of a visual heatmap and a natural language report may be performed. Step (e) may include two sub-steps.

[0141] First, a step may be performed to generate a highlight using a SARE-based heatmap by measuring specific areas where a semantic shift occurred between the original image and the reconstructed image based on distance. In this step, SARE values ​​per patch may be calculated, visualized by applying color mapping, overlaid with the original image, and bounding boxes and labels may be added to major areas of change.

[0143] Next, a step may be performed in which a multimodal model (MLLM) analyzes the dynamic trajectory and quantitative values ​​of the multidimensional indicator error to generate a natural language text report containing logical findings pointing out structural defects within the image and characteristics of the generative model. In this step, the original image, the reconstructed image, the SARE heatmap, and all quantitative indicators are input into the MLLM, and a structured report including an overall judgment, key findings, detailed analysis, suspected areas, conclusions, and recommendations can be generated according to pre-designed prompts.

[0145] According to some other embodiments of the present invention, the components of the system may be implemented or modified in various ways.

[0147] Regarding feature point extraction, in one embodiment of the present invention, a 496-dimensional feature vector was used, but in another embodiment, it can be reduced to 128 dimensions by applying PCA (Principal Component Analysis) for computational efficiency. In yet another embodiment, high-dimensional features can be extracted using a deep learning-based feature extractor (e.g., a pre-trained ResNet, EfficientNet, Vision Transformer). In yet another embodiment, Discrete Cosine Transform (DCT) coefficients or Discrete Wavelet Transform (DWT) coefficients may be additionally included to enhance frequency domain features.

[0149] Regarding the construction of the anchor graph, in one embodiment of the present invention, a Laplacian eigenmap was used, but in other embodiments, other manifold learning techniques such as t-SNE, UMAP, and Isomap may be used. In another embodiment, a graph neural network (GNN) may be used to learn the structure of the anchor graph and generate node embeddings. In another embodiment, the number of anchor nodes may be adjusted according to the application field, for example, reduced to 10,000 nodes in a lightweight version for edge devices, or expanded to more than 500,000 nodes in a large-scale server system.

[0151] Regarding the diffusion model, Stable Diffusion was used in one embodiment of the present invention, but other diffusion models such as DALL-E, Midjourney, and Imagen may be used in other embodiments. In yet another embodiment, a lightweight version with the number of parameters reduced to one-fourth by applying a model distillation technique may be used. In yet another embodiment, the noise schedule may be changed to various forms, such as a linear schedule or a squared schedule. In yet another embodiment, the number of diffusion steps may be adjusted to a wider range from 10 steps to 200 steps depending on the importance of the image.

[0153] Regarding the IAPL engine, in one embodiment of the present invention, 16 trainable tokens (8 prefixes + 8 suffixes) were used, but in another embodiment, the number of tokens can be adjusted from 4 to 32. In yet another embodiment, the weights (α, β, γ) of the composite loss function can be adjusted according to the application, for example, if semantic consistency is more important, γ can be increased. In yet another embodiment, the optimization algorithm can be changed to Adam, SGD, RMSprop, etc. In yet another embodiment, other multimodal models such as LLaVA, InstructBLIP, and Flamingo can be used instead of BLIP-2 for initial prompt generation.

[0155] Regarding multidimensional metric analysis, in one embodiment of the present invention, three metrics—snapback, SARE, and Jacobian-Spectral—were used, but in other embodiments, additional metrics may be included. For example, pixel-level reconstruction error (MSE, SSIM, LPIPS), frequency domain error (FFT spectral difference), perceptual loss, and adversarial loss may be added. In another embodiment, weights (w₁, w₂, w₃) may be dynamically adjusted according to the application field or image type when calculating the final score. In yet another embodiment, SARE may be calculated using other vision-language models such as ALIGN, BLIP, and Florence instead of CLIP.

[0157] Regarding the self-learning mechanism, in one embodiment of the present invention, HDBSCAN clustering was used, but in other embodiments, other clustering algorithms such as K-means, DBSCAN, Spectral Clustering, and Agglomerative Clustering may be used. In another embodiment, the separation between clusters may be controlled by adjusting the margin (m) and balance weight (λ) of the hierarchical proxy loss. In another embodiment, the anchor graph update trigger conditions may be changed, for example, to perform more frequent updates after 50 OOD samples have accumulated or 3 days have passed, or to perform less frequent updates after 200 samples have accumulated or 14 days have passed.

[0159] Regarding the output of XAI results, in one embodiment of the present invention, a Jet color map was used, whereas in other embodiments, other color maps such as Viridis, Plasma, and Inferno may be used. In another embodiment, the resolution of the heat map may be adjusted, for example, a 16×16 patch (low resolution, high speed) or a 64×64 patch (high resolution, precision) may be used. In another embodiment, other models such as Claude, Gemini, and LLaVA-1.6 may be used as the MLLM instead of GPT-4V. In another embodiment, the report format may be HTM

[0160] It can be output in various formats such as L, PDF, and JSON.

[0162] Regarding hardware configuration, in one embodiment of the present invention, an NVIDIA A100 GPU was used, but in other embodiments, other GPUs such as V100, H100, and RTX 4090 may be used. In yet another embodiment, a dedicated AI accelerator such as a TPU (Tensor Processing Unit) or NPU (Neural Processing Unit) may be used. In yet another embodiment, model quantization, pruning, and knowledge distillation may be applied to make it executable on edge devices (smartphones, IoT cameras) to achieve lightweight performance.

[0164] Regarding applications, the system of the present invention can be applied to various industrial fields. In social media platforms, user-uploaded images can be verified in real time to automatically attach warning labels to AI-generated images. In news media, the authenticity of photos submitted by readers or images used in articles can be verified to ensure the credibility of reporting. In the legal and forensic fields, the authenticity of digital evidence can be verified to utilize it as legal evidence. In the field of copyright protection, AI-generated content can be automatically identified and distinguished from original creations. In e-commerce platforms, the authenticity of product images can be verified to contribute to consumer protection. In the medical imaging field, medical safety can be enhanced by verifying whether diagnostic images have been forged.

[0166] Regarding performance optimization, various optimization techniques may be applied in some embodiments of the present invention. GPU utilization can be increased by processing multiple images simultaneously through batch processing. Processing speed can be improved by distributing large models across multiple GPUs through model parallelization. Search speed can be increased by storing frequently accessed anchor nodes or feature vectors in memory through a caching mechanism. Overall throughput can be increased by processing first-order filtering and second-order verification in a pipeline manner through asynchronous processing. Resources can be used efficiently by automatically adjusting the batch size according to system load through dynamic batch size adjustment.

[0168] Regarding security and privacy, some embodiments of the present invention may include features for protecting user privacy. Faces or personally identifiable information can be automatically masked before analysis through image anonymization. Analysis can be performed on edge devices without transmitting images to a server through federated learning. Noise can be added through differential privacy to prevent information about individual images from being exposed when updating the anchor graph. Data can be protected during image transmission and result reception through encrypted communication.

[0170] The embodiments described above are merely illustrative of the technical concept of the present invention, and those skilled in the art to which the present invention pertains will be able to make various modifications and variations within the scope of the essential characteristics of the present invention. Accordingly, the embodiments disclosed in the present invention are intended to explain, not limit, the technical concept of the present invention, and the scope of the technical concept of the present invention is not limited by these embodiments. The scope of protection of the present invention shall be interpreted by the claims below, and all technical concepts within an equivalent scope shall be interpreted as being included within the scope of rights of the present invention. Explanation of the symbols

[0174] 100: AI-Generated Image Precision Discrimination System Based on Reversible Semantic Restoration 110: Image input section 120:1st Verification Module 130: 2nd Verification Module 140: Board of Adjudication 150: Output module 700: Original Image 710:IAPL engine 720: Diffusion Model 730: Reconstructed image 740:CLIP Feature Encoder 750: Cosine Similarity Calculator 760: SARE Indicator

Claims

Claim 1 An image input unit that receives an image to be verified; a first verification module that performs a similarity search in an EMR (Efficient Manifold Representation) anchor graph having feature points of an actual captured image and a previously determined AI-generated image as nodes, based on fingerprints extracted from the input image, and makes a first determination as immediate PASS, immediate BLOCK, or PENDING; and a second verification module that, for the image with a PENDING result from the first determination, generates a text prompt optimized to minimize information entropy through an Image Adaptive Prompt Learning (IAPL) engine, inputs the prompt into a Diffusion Model to perform reconstruction, and analyzes whether the image's semantic restoration is reversible by calculating a Semantic cognition-based reconstruction error (SARE) that measures the semantic shift between the original image and the reconstructed image in a CLIP feature space. A reversible semantic restoration-based AI-generated image precision discrimination system characterized by comprising: a result analysis and judgment unit that performs a final determination on whether an image is generated based on the analysis results of the second verification module, and clusters by similar generation method by applying a Hierarchical Proxy Loss technique to newly identified unlearned forgery patterns (OOD) based on the final determination results, and then updates the anchor graph with new anchor nodes; and an XAI (Explainable Artificial Intelligence) result output module that provides the basis for the final determination in the form of a visual heatmap and a natural language report. Claim 2 A reversible semantic restoration-based AI-generated image precision discrimination system according to claim 1, wherein the fingerprint of the first verification module is a vectorized representation of at least one of the color histogram, texture gradient, edge curvature, and micro-frequency characteristics unique to the generation model of the image to be verified, and the anchor graph is an EMR (Efficient Manifold Representation) anchor graph having the fingerprints of the actual captured image and the previously determined AI-generated image as nodes, and the first verification module classifies into immediate PASS, immediate BLOCK, or two-stage PENDING based on the distance between the input image and the anchor point through nearest neighbor search (ANN) on the EMR anchor graph. Claim 3 The AI-generated image precision discrimination system based on reversible semantic restoration according to claim 1, wherein the second verification module generates an Importance Mask (IML) by utilizing a deep learning-based Saliency Map to select partial regions, and optimizes the analysis region and computation path by applying Adaptive Routing (CATImage) technology that dynamically branches the computation intensity of the reconstruction step of the diffusion model according to the complexity and resolution of the image. Claim 4 delete Claim 5 delete Claim 6 A reversible semantic restoration-based AI-generated image precision discrimination system according to claim 1, wherein the result analysis and judgment unit applies a Hierarchical Proxy Loss technique to cluster feature points of newly discovered Out-of-Distribution (OOD) data or new AI generation patterns by similar generation techniques, and self-learns the system by adding them as new anchors to the anchor graph of the first verification module. Claim 7 A reversible semantic restoration-based AI-generated image precision discrimination system according to claim 1, wherein the XAI result output module provides to the user a SARE-based heatmap visualizing a region where feature differences measured based on the SARE index are concentrated; and an MLLM report inferring the basis for discrimination in natural language based on quantitative data of curvature, SARE values, and Jacobian spectrum values ​​of snapback dynamics analysis, which utilizes a large-scale multimodal model (MLLM) to track the trajectory and curvature of reconstruction quality according to changes in noise intensity of the diffusion model to confirm the deterministic regression characteristics of the generative model. Claim 8 A method for precise image discrimination performed in an AI-generated image discrimination system based on reversible semantic restoration according to claim 1, comprising: (a) receiving an image to be verified and extracting statistical and geometric feature points from the image; (b) searching the extracted feature points on an EMR anchor graph including actual captured images and previously identified AI-generated image nodes to perform a first judgment as immediate PASS, immediate BLOCK, or PENDING; (c) for images classified as PENDING as a result of the first judgment, generating an image-adaptive prompt token trained under an information entropy minimization condition, performing reconstruction using a diffusion model, and extracting a Semantic-Aware Reconstruction Error (SARE) index, which is the semantic distance between the original and the reconstructed image in a CLIP feature space, to perform precise analysis; (d) finally determining whether the image is generated based on the precisely analyzed multidimensional index, and clustering the finally determined new forgery pattern based on hierarchical proxy loss to update the anchor graph as a new node, thereby enabling the system to self-learn. and (e) a step of outputting the final judgment result and the basis for judgment as an explainable artificial intelligence (XAI) result in the form of a visual heatmap and a natural language report; comprising an AI-generated image precision judgment method. Claim 9 In claim 8, the above step (c) comprises: a step of optimizing resources by selecting a region of interest (ROI) of the image with judgment pending and allocating computational intensity according to complexity; a step of generating a prompt token optimized for the image with judgment pending through image adaptive prompt learning (IAPL); a step of analyzing contraction mapping and snap-back convergence in the process of reconstructing the image by inputting the generated prompt into a diffusion model; and a step of extracting semantic cognitive-based reconstruction error (SARE) and Jacobian-spectral indicators using a CLIP feature space; characterized in that it comprises an AI-generated image precision discrimination method. Claim 10 In claim 8, the above step (e) comprises: a step of measuring a specific area where a semantic shift between the original image and the reconstructed image has occurred based on distance and generating a highlight using a SARE-based heatmap; and a step in which a multimodal model (MLLM) analyzes the dynamic trajectory and quantitative values ​​of the error of the multidimensional indicator to generate a natural language text report of logical findings pointing out structural defects within the image and characteristics of the generative model; characterized in that it is a method for precise discrimination of AI-generated images.

Images