Method for Generating Multidimensional Realism Scores for Deepfake Videos

Abstract

The invention relates to a method for automatically identifying and separating videos with low realism from datasets by generating intermediate realism scores from a mock video and the source video using different image quality metrics, combining these scores into a multidimensional vector, creating a reference model from high-realism mock videos containing an average representation vector and a covariance matrix, and calculating the multivariate statistical distance of the test video vector from this model as the final video realism score.

Description

1 TARIFF Method for Generating Multidimensional Realism Scores for Deepfake Videos TECHNICAL AREA 5 The invention allows for the computer-assisted enhancement of the visual realism levels of deepfake videos. This involves evaluating and extracting multidimensional quality relationships at the video level. Creating a statistical reference model from highly realistic fake videos and A final video realism of 10 is calculated based on the distance of the analyzed video from the model in question. The generation of scores and the realistic editing of deep sham datasets, low Separating samples with a level of realism and deep forgery detection systems It is related to improving training and evaluation datasets. PREVIOUS TECHNIQUE 15 The rapid advancements in deepfake production technologies in recent years have made them indistinguishable to the human eye. It is quite difficult to produce fake videos with a high degree of visual realism. This situation has led to the spread of misinformation, identity theft, digital manipulation, and media. This brings with it serious security and accuracy problems in areas such as security. (Statement 20) Developing reliable deepfake detection systems against threats is of great importance. It carries. There are numerous studies in the literature for the training and evaluation of such systems. Deepfake datasets have been developed, and these datasets are generally subject to various manipulations. It includes thousands of real and fake video examples created using various methods. However, the current The visual realism levels of the videos included in deepfake datasets differ significantly from one another. The extent of difference can vary. Within the same dataset, some deepfake videos may not be human-like. While some deepfake videos appear quite convincing and natural in terms of visual perception, others... Significant visual distortions, artificial facial transitions, texture inconsistencies, blurring, compression It may have low realism due to artifacts or geometric inconsistencies. These videos, especially those with a low level of visual realism, are used to train deepfake detection models. 30 This involves using trained models in processes that focus only on obvious artifacts. generalization against high-quality deepfake examples in the real world as a reason It can reduce its capabilities. In addition, low-quality deepfake videos, deepfakes by disrupting the visual quality distribution within the video dataset, deepfake detection systems This can also lead to an overestimation of performance. Therefore, deepfake data 35 videos in their clusters according to their level of realism, in a way that is compatible with human visual perception. analysis, automatic identification of low-realism samples, low visual 2 High-quality deepfake videos are extracted from deepfake datasets, resulting in deepfake video data. Organizing these sets in a realism-aware manner is of critical importance. However, studies in the current literature mostly focus on the dataset level (Fréchet (Inception Distance, etc.) performs visual quality assessment at the individual video level. a general, explainable and automated visual realism assessment mechanism 5 It is unable to provide this. Furthermore, a significant portion of the existing methods rely on human evaluations. It requires labeling processes based on or relies on specific deepfake detection models. This situation allows existing visual quality measurement methods to work with different data sets. This also limits its generalizability. There are various studies in the literature regarding the evaluation of the visual realism level of deepfake content. Methods have been proposed. Most of the existing methods involve deepfake datasets. Automatic and reliable realism levels of individual videos that are compatible with human visual perception. There are various technical limitations in determining this method. Below are some examples from the literature on this subject. The proposed solutions to the problem and their technical limitations are summarized below: 15 Heusel and colleagues (Heusel et al., 2017) studied generative artificial intelligence systems. to evaluate the extent to which the generated synthetic images resemble real images For this purpose, Fréchet proposed a distribution-based quality metric called Inception Distance (FID). This method involves embedding representation vectors (EMBs) of images from real and artificial datasets. Comparing the statistical distributions of vectors to identify the statistical differences between two data sets. It measures numerically. The FID metric is a general visual indicator in the literature for synthetic data generation systems. It is widely used to evaluate performance. However, FID This approach is a method that works at the dataset level and can obtain reliable results. It requires datasets containing a large number of samples (1000 or more). Therefore, 25 The FID metric measures the degree to which deepfake videos within a deepfake dataset are authentic. Detecting videos with low realism levels through video-based analysis. It cannot be used for that purpose. Visual Realism 30 evaluation of deepfake videos in 2023 They organized a competition for Assessment: VRA (Vertical Transfer Assessment). The official name of the competition is DeepFake Game. The competition is called "Competition on Visual Realism Assessment (DFGC-VRA)". Participants in the competition... Peng and his study presented methods proposed by the teams and performance comparisons of these methods. This is presented in detail in the 2023 conference paper by Peng et al. (2023). In the competition, the visual realism scores of the deepfake videos were determined by the organizing team as 35. Targets determined based on real human evaluations and set for each video. Mean Opinion Score (MOS) estimates of reality scores by contestants 3 This has been requested. Similarly, Sun and colleagues (Sun et al., 2023) and Dragar and their work Dragar et al. (2023) conducted a study on deepfake datasets containing a reality score. They have proposed learning-based visual realism prediction models. However, the aforementioned The methods are largely dependent on labeling processes based on human assessments. It works. Generating MOS for each deepfake video is costly, time-consuming, and subjective. 5 Since it is a process, such approaches need to be scaled and reused to different datasets. Problems are encountered in implementing it without the need for training. Furthermore, this... The methods are mostly based on specific datasets and specific labeling protocols. Because it is trained, it performs under different manipulation methods and different compression conditions. They are able to show losses and operate with relatively high error rates. 10 Kim and colleagues (Kim et al., 2024) in their 2024 study, analyzed different deepfake video data. visual quality levels of clusters using referenced and non-referenced image quality metrics (IQM / They analyzed the deepfake dataset using NR-IQM; deepfake detection was based on the visual quality of the deepfake dataset. It examined the relationship between performance and image and video. 15 IQM and NR-IQM methods In the literature, the measurement of visual distortions resulting from compression processes is widely used. However, these metrics are mostly used to measure general image distortions. It was developed for this purpose and to determine the "realism" level of a video. It is not designed in a way that specifically considers compression ratio, resolution, bit-rate, and source video quality. These factors can directly affect IQM scores. Therefore, for visual reality analysis, 20 Using only raw IQM or NR-IQM scores can lead to visual manipulation. It is insufficient for reliably measuring changes in realism. Furthermore, the word... This study focuses on analyzing average visual quality at the dataset level. It does not offer a visual realism assessment mechanism at the video level. Song and colleagues (Song et al., 2024) found low efficiency in training deepfake detection models. The use of high-quality simulated videos negatively impacts model generalization performance. He stated that this has an impact and proposed a quality-focused training approach to address this problem. In the proposed method, the quality scores of the fake videos are derived from the deepfake detection model. By combining the feed outputs with the identity similarity between the source and the fake faces, 30 is calculated. However, this approach is directly related to human visual perception. It relies on facial resemblance and detectability information rather than realism analysis. Especially if the face in the source video is successfully replaced with a different identity, Even deepfake videos with high visual realism have low similarity scores and consequently... This can produce a low visual quality score. Furthermore, the proposed method may not produce a human visual quality score of 35. The extent to which this aligns with perception has not been investigated in this study. For these reasons, Song et al. 4 The method proposed by (2024) is sufficient for the problem covered by this patent and It does not provide a comprehensive solution. CN117591815B, developed by Sun and colleagues (Sun et al., 2024). The patent, however, defines quality 5 as the quality of artificial data produced in different modalities such as image, sound, and video. a holistic quality analysis system that evaluates levels using multiple quality metrics It is recommended that quality metrics such as FID, SSIM, PSNR, and similar metrics be used together in this approach. An overall quality score is created using human-like methods similar to MOS. However, the proposed system uses human-like methods similar to MOS. It also uses its assessments in video reality analysis, but this situation is suggested. This restricts the solution from functioning independently of human evaluation. Also, the 10 used... Quality metrics have been largely improved for image and audio data, at the video level. It is unable to model in detail the changes in visual realism caused by manipulation. Specifically, the inability to calculate the FID metric for a single video makes the recommended method video-based. This constitutes a significant technical limitation in terms of realism analysis at this level. As a result of research conducted in the literature, the application number “US20230153973A1” and “Determining A Degree Of Realism Of An Artificially Generated Visual Content The patent application is titled "Determining the Degree of Realism of Produced Visual Content". It has been encountered. The application in question concerns artificially generated material and material captured from the real environment. visual content is presented to people in a mixed manner, and people perceive the content as either artificial or real. 20 It is related to determining the degree of realism in the reactions that are evaluated as such. However Comparison of IQM distributions of source and fake videos in the application, different IQM Generating a multidimensional realism score vector from the distributions and creating a realistic mock video. Regarding the automatic generation of video realism scores based on Mahalanobis distance to the model. No evidence was found. 25 As a result of research conducted in the literature, the application number “US20250220251A1” and “Methods, Systems, And Media For Determining Perceptual Quality Indicators Of Video Content Items - Methods and Systems for Determining Perceptual Quality Indicators of Video Content Elements A patent application titled "And Environments" was found. This application concerns video 30. The frames separately display indicators of content quality, video degradation, and compression sensitivity. by determining the subnets and combining these indicators to determine the overall video quality level It is related to the production of a fake video. However, the application mentions a source video from which a fake video was produced. Comparison of intermediate realism scores obtained from statistical distribution distances, Creating a reference model from highly realistic mock videos and achieving a final score of 35. The calculation of multivariate distance in the model in question is not explained. The studies in the literature summarized above examine visuals in deepfake datasets. While highlighting the importance of the realism problem, these solutions are generally humane. relying on their assessments, it needs specific deepfake detection models. or it can be explained at the video level and an automated realism assessment mechanism It cannot provide this. Therefore, each deepfake video within a deepfake dataset has 5 The automatic analysis of visual information in a way that is compatible with human visual perception; realism. rating according to their levels and videos with low realism quality are considered unreliable. The problem of cleaning the dataset by defining it in this way is still not fully understood in the literature. It has not been resolved. A BRIEF DESCRIPTION OF THE INVENTION The present invention aims to eliminate the aforementioned disadvantages and introduce new technologies to the relevant technical field. In order to gain advantages, a fake video was created using data obtained from the source video from which this fake video was produced. The distances between the multiple image quality distributions obtained are in multidimensional space and video 15 by combining them at this level, a realism that is compatible with human visual perception, explainable and automatic. It relates to the method used to generate the score. The main goal of the invention is to differentiate "different" image quality metrics from original and deepfake videos. Intermediate realism scores generated from generated statistical distributions are multidimensional realism 20 combining the score into a vector, and then performing a realism assessment based on this new vector, is a single... The aim is to reduce measurement inaccuracies that may arise in solutions based on a single quality metric. Another aim of the invention is to provide multi-dimensional quality for "highly realistic" fake videos. Creating a "realistic" pseudo-video reference "model" from the distribution of score vectors, the 25 in question The deepfake will be used to measure and analyze the covariance relationships between the model and quality metrics. The similarity of the video to the reference model can be measured using multivariate statistical distance (e.g. The goal is to produce the final video realism score by measuring it with the mahalanobis distance. Another aim of the invention is to provide a separate opinion for each fake video using data obtained from people. to generate the MOS score or to provide feedback on a specific deep spurious detection model It can be applied to different datasets and manipulation methods without needing to be otherwise used. The goal is to create a deepfake authenticity assessment system. Another purpose of the invention is to "automatically" create fake videos with a low level of realism. identification, separation from data sets, and higher-level, more realistic samples. The goal is to ensure the preparation of high-quality training and assessment (test) sets. 6 All the purposes mentioned above and those that will emerge from the detailed explanation below. The present invention aims to achieve this by combining source video with mock video for different quality metrics. multiple intervals between statistical distances from the distributions of image quality metrics. Generating realism scores, combining these scores into a multidimensional vector, 5 Average representation vector and covariance matrix of highly realistic fake videos. creating a reference model containing the test video and generating a multidimensional vector from this reference model. the process of generating the statistical distance of the model as the final video realism score It is a method that includes steps. The best way to utilize the advantages of the existing invention, together with its structure and additional elements. For it to be understood, it must be evaluated together with the figures explained below. BRIEF DESCRIPTION OF THE FIGURES Figure 1 is a representative illustration of the intermediate stage realism score analysis module. Figure 2 shows a vector of realism scores based on multiple image quality metrics for fake videos. It is a representative illustration of the account. Figure 3 shows the realism score matrix using highly realistic simulated videos. A representative 20 of the training phase that creates the covariance matrix and the mean representation vector. It is a representation. Figure 4 shows the calculation of the realism score vector for the given pair of mock video and source video. Mahalanobis distance of the vector to the realism model as the final realism score It is a representative illustration of the testing phase in which it is measured. Figure 5 shows the video realism score and the non-referenced image quality metric methods that are the subject of the invention. the area under the curve in terms of the success of detecting videos with a low level of realism It shows the comparison based on AUC (Analytics Universities). The drawings do not necessarily need to be scaled and are necessary for understanding the invention. Details that are missing may have been omitted. Also, the elements that make up the drawings are at least 30 years old. It may be shown schematically to a certain degree. REFERENCE NUMBERS 1001. An interim realism score will be calculated for the fake video based on a single image quality metric (35). The source video from which this video was produced was taken as input to the electronic processing unit. 7 1002. Faces in selected frames from the intro videos (both for the fake and the source video) Preprocessing of regions of interest, based on the selected image quality metric (IQM) for these regions Calculation of IQM scores according to 1003. Selection of the image quality metric method to be used in calculating the interim realism score. 1004. The statistically significant distance between the quality score distributions of the fake and source videos is 5. measurement The statistical distance calculated in steps 1005 and 1004 belongs to the selected image quality metric. generated as an interim realism score (ES-i) 1006. Creating a multidimensional realism score vector (MR) using a fake video and its corresponding data. The source video pair is passed as input to k different reality score analysis modules. 10 1007. Performing the first realism score analysis according to the first image quality metric. 1008. Performing a second realism score analysis based on the second image quality metric. 1009. Performing the analysis of the Kth realism score according to the Kth image quality metric. 1010. Obtaining the first interim realism score (GS-1). 1011. Obtaining the second interim realism score (GS-2) 15 Obtaining the 1012th Kth Intermediate Realism Score (GS-k) 1013. Combining the k intermediate realism scores obtained in vector form. 1014. Generation of a k-dimensional realism score vector (GS vector) for a fake video. 1015. Using n highly realistic fake videos to create a reference model. 20 The input of n source videos belonging to these into the electronic processing unit 1016. k-dimensional realism score for each pair of fake videos and source videos received at the input. calculation of the vector 1017. Realism score vectors in kxn dimension realism score matrix (GSM) unification 1018. kxk dimensional covariance matrix and kx 1 dimensional mean 25 from the realism score matrix. Calculation of the representation vector 1019. Realistic mock video reference including covariance matrix and mean representation vector. creation of the model 1020. The final video realism score will be calculated using a mock test video and its source. The video is received as input to the electronic processing unit. 30 1021. Testing procedure for the first realism score analysis according to the first image quality metric. implementation 1022. Testing procedure for the analysis of the second realism score according to the second image quality metric. implementation 1023. For the testing process of the Kth realism score analysis according to the Kth image quality metric, 35 implementation 1024. Obtaining the first intermediate realism score (GS-1) (for the test) 8 1025. Obtaining the second interim realism score (GS-2) (for the test) 1026. Obtaining the K-th intermediate realism score (GS-k) for the test) 1027. Combining the k test realism scores obtained in vector form. 1028. kx 1D realism score vector (GS vector) for the test mock video creation 5 1029. Covariance matrix and mean representation of the realistic mock video reference model. taking the vector as input to the testing phase 1030. Test fake video realism score vector with realistic fake video reference model. Calculation of Mahalanobis or similar multivariate statistical distance between them 1031. Generating the calculated distance as the final video realism score (VGS) and electronically 10 recording to the recording medium DETAILED DESCRIPTION OF THE INVENTION This detailed explanation covers the production of a multidimensional realism score of 15 for deepfake videos. The method is explained with illustrative examples that will help to better understand the subject. However, these examples do not limit the scope of protection. The invention is a solution to problems in previous techniques, compatible with human visual perception. Video-20 can automatically determine the level of visual realism of deepfake videos. It relates to an analysis method at this level. Thanks to the proposed invention, a deepfake dataset can be analyzed. The videos inside are automatically categorized into low and high levels of realism. data can be classified, and low-quality and visually artificial samples can be identified. It can be separated from the set. Thus, it is more realistic and more compatible with human perception. Datasets consisting of deepfake samples can be created, resulting in a more reliable, more 25 generalizable and more robust deepfake detection systems to real-world conditions It will be possible to contribute to its development. This invention aims to overcome the limitations found in previous techniques. Deepfake videos are automatically scaled to a level of realism that is compatible with human visual perception (30). A new method capable of analyzing is presented. In the proposed approach, the source is compared with the fake video. The relationship between visual realism in the video and multiple image quality evaluation metrics. They are analyzed using various methods, and the results of each analysis are sent to the introductory videos (fake and original). Multiple (e.g., k) intermediate realism scores are obtained. Then different These scores obtained from realism analyses represent a multidimensional representation of realism (k-dimensional 35 They are combined to form a vector. The invention also includes n high-powered k-dimensional representation (reality) for deepfake and source video pairs with a level of realism 9 vectors are being calculated, and a matrix of nxk dimensions is formed from these representation vectors. A reference realism model is defined (covariance matrix and mean representation vector) and The k-dimensional reality vectors of the deepfake videos to be analyzed in the testing phase are as follows: Statistical distances to the model are calculated. The higher this distance, the better the test. It can be said that the deepfake video produced is so far removed from reality. Thus, deepfake 5 classic methods rely on only a single quality metric for analyzing the realism of videos In contrast to these approaches, multidimensional statistical distributions take into account multidimensional quality relationships. a new deepfake realism assessment based on and compatible with human visual perception The mechanism is obtained. Thanks to the method described in the invention, deepfake videos are made to look realistic. They can be automatically rated according to their level, with 10 having a low level of realism. videos can be reliably identified and more realistic datasets can be created. This can help improve the generalization performance of deepfake detection systems. Figure 1 shows the general overview of the intermediate stage realism score analysis module proposed within the scope of the invention. The working structure is shown. The main purpose of this module is to create a fake video and compare it to 15 other videos. By analyzing the relationship between the visual quality of the source video used in its production and the relevant fake video. The goal is to generate an intermediate realism score for the video. This score is not the final value; it is patented. This contributes to obtaining the final deepfake authenticity score for the method proposed within this scope. It is an intermediate value that will provide. The structure shown in Figure 1 depends on a specific image quality metric. not, different IQM (Image Quality Metric) or NR-IQM (No-Reference Image Quality Metric) 20 It has a modular structure that can work with various methods. Figure 1 shows the deepfake video given as input. And its source video is primarily processed frame by frame. At this stage, from the videos Face region detection, alignment, cropping, and preprocessing on the obtained image frames. The procedures can be applied. Then, the selected image quality evaluation method number i can be used. Using (IQM), selected frames of interest are found in both the source video and the mock video. Quality scores are calculated for each region. Thus, for each deepfake and source video... A score set for the relevant quality metric (IQM) is obtained. Used within the scope of the invention. Quality assessment methods are not limited to any single metric. For example... BRISQUE (Mittal et al., 2012), LIQE (Zhang et al., 2023), MANIQA (Yang et al., 2022), MUSIQ (Ke (Zhang et al., 2021), DBCNN (Zhang et al., 2020), TReS (Golestaneh et al., 2022) or similar referenced sources or 30 Non-referenced (NR-IQM) image quality assessment methods can also be used. In addition, Different deep learning-based quality analytics tools that can measure quality related to human visual perception. These methods can also be included in the method. Quality scores obtained for the source video and the fake video. The clusters are then statistically compared. At this stage, the aim is only to simulate This is not about measuring the raw quality level of the video, but about the source 35 resulting from the fake video production process. The goal is to analyze the changes occurring in the visual quality structure of the video. Therefore, the invention... within this scope, the statistical difference or amount of quality change between quality score distributions. is calculated. For this analysis, Cohen's d or similar statistical comparison is used. Methods can be used. As a result of this analysis, quality number i was selected. An intermediate realism score is generated corresponding to the evaluation method. This score is shown as GS-i in Figure 1 at 1005 process steps. The GS-i score is the relevant quality metric. 5 represents the extent to which the fake video differs visually from the source video. This allows us to analyze not only the quality value of a single image, but also source and fake videos. A video-level measure of realism is obtained, representing the quality relationship between them. It should be noted that different IQM methods are used for the deepfake and original video pairs to be analyzed. Multiple GS-i values ​​can be obtained using this method. This flow involves operations 1001-1005. It fulfills its functions; allowing the input of a single fake video-source video pair, face or interest 10 Selecting and preprocessing frames containing the region, selecting the relevant area with the selected image quality metric. quality scores and distributions from the regions separately for source and fake video. calculation, measurement of the statistical distance between the distributions in question and this This involves generating the distance as a GS-i intermediate realism score. Figure 2 shows how the multidimensional realism score vector proposed within the scope of the invention is calculated. It is shown that it was created. Figure shows the deepfake video taken as input and the source video. Using the method described in section 1 and different IQM methods, k intermediate realism scores were obtained. (GS-1, GS-2, ... , GS-k) are produced. Realism achieved using a single quality metric. The score alone is 20, which compensates for the visual distortions and manipulation effects that can occur in deepfake videos. It may not represent the image adequately. This is primarily due to different image quality evaluation criteria. The difference is that their metrics show different sensitivities to different types of visual distortion. For example, some Quality assessment metrics include compression artifacts, blurring, or contrast degradation. While working with greater precision, some methods can distort facial texture, geometric inconsistencies, or It is able to analyze detail losses caused by manipulation more successfully. This 25 Therefore, realism assessments based on a single quality metric differ depending on the deepfake production. It can produce inconsistent results in its methods. This invention was created to overcome this problem. Within this scope, multiple IQM or NR-IQM methods are used together. (Figure 2 shows this). In the blocks, which are called realism score analysis modules, each different quality assessment The analysis described in Figure 1 is performed for the metric, and a separate “intermediate stage 30” is conducted for each metric. A "realism score" (GS-i) is obtained. The k realism scores obtained are shown in Figure 2. In this stage, the deepfake video is combined in vector form and a k-dimensional image is created for the video being analyzed. A realism score vector (1014 elements) is generated. This can be used within the scope of the invention. Quality assessment metrics are not limited to a specific set of metrics. Referenced image. Quality metrics, image quality metrics without reference, deep learning-based quality analysis 35 metrics or different quality assessment metrics related to human visual perception are included in the method. This allows the proposed structure to be adapted to different deepfake production methods and different compression ratios. 11 It offers a flexible architecture that can be adapted to different levels and datasets. (Figure) Thanks to the multidimensional realism scoring structure defined in section 2, the analyzed deepfake video Not just from a single quality metric, but from different quality assessment metrics. The information gathered is used and evaluated together. Thus, the visual aspects of the fake video are analyzed. The structure of realism can be analyzed in a more comprehensive and reliable manner. This 5 In this stage, process step number 1006 differs from single analysis entry number 1001 in that it is the same. k different realism score analysis modules using k different quality metrics of the video pair It refers to the transfer of intermediate scores 1010-1012 from analyses 1007-1009. The scores are obtained using steps 1013-1014 for a single kx 1-dimensional system. They are combined into a vector. 10 Figure 3 shows the reference realism model for deepfake videos proposed within the scope of the invention. The creation process flow is shown. The main purpose of this structure is to be aesthetically pleasing to human visual perception. deepfake videos with a high level of realism represent realism behavior The aim is to create a reference model. The reference realistic mock-up videos used within the scope of the invention are 15. They can be determined using different methods. For example, visual assessments based on human evaluations. Realism analyses, MOS (Mean Opinion Score) scores, expert evaluations, or Using existing quality assessment methods in the literature to achieve a high level of accuracy. Deepfake videos can be selected. Thus, the system can target a specific dataset or a particular type of video. It has a flexible structure that can operate without being dependent on the labeling method. 20 An important point here is that the deepfake videos selected must have both a high degree of realism. The key difference is that it is produced using very different deepfake creation methods. This allows... This helps to avoid the generalization problem that would arise from a limited number of deepfake methods. The number of videos to be used for the reference model is denoted by n, and in terms of number, it is a... There are no limitations; videos can be selected manually or automatically, up to 25 selected videos. The videos have a low level of manipulation and are difficult to visually discern as fake. This is considered the main principle. Each of the selected deepfakes (n units) and source videos (n units) is used to create the reference model. For a pair, n k-dimensional realism score vectors are generated as described in Figure 2. 30 At this stage, the k n-dimensional realism score vectors obtained are added side-by-side to form a kxn vector. It creates a three-dimensional reference representation matrix. This matrix represents the realism of n deepfake videos. It shows how the score vectors are distributed in k-dimensional space. Then, for this matrix... A kxk element covariance matrix and a k-element score mean vector (centroid) are calculated. The calculated covariance matrix and the mean score vector are a parametric reference deepfake 35 It will be used to create a realism model. Thus, it will have a high level of realism. A parametric statistical model representing the multidimensional realism characteristics of deepfake videos. 12 A reference model is obtained. The elements of the reference realism model shown in Figure 3. (covariance matrix and mean score vector) deepfake videos in later stages It will be used in evaluating levels of realism. In this way, visual evaluation will be possible. A deepfake video analyzed at this stage has a high level of realism. The similarity relationship with the videos can be taken into consideration. As a result, different quality 5 A reliable system where information obtained from evaluation metrics is evaluated together. A realism analysis mechanism can be obtained. Figure 4 shows the final video realism score (VGS) calculation process proposed within the scope of the invention. The flow is shown. Figure 3 shows the training phase for the proposed invention, and Figure 4 shows the testing (final 10). Figure 4 shows the evaluation phase. The input (1020 elements) of Figure 4 shows the reality score determination. It consists of a deepfake video and its source video. This video pair is then... k different quality methods are used for analysis, resulting in k different intermediate stage realities. A score (GS-1, GS-2, ... , GS-k) is generated (elements 1021-1026). Then, k items are generated. The intermediate stage reality score value is a k-element score vector (GS 15) representing the input video. The vector is transformed (with 1027 elements). The steps applied so far are shown in Figure 1 and Figure 2. The processes are the same as those presented in section 2. The resulting GS vector (1028 elements) is used in the next stage. The covariance matrix and mean score vector created during the training process described in Figure 3. It is compared with the statistical model it represents (1029 elements). During this comparison... Instead of using only individual quality metric scores, a k-dimensional realism score vector is used, 20 This allows the relationships between different quality assessment metrics to be taken into account. Thus, the analyzed deepfake video has a high level of realism compared to the reference video. It is possible to determine the extent to which deepfake videos resemble the overall quality structure of real-world videos. (Invention) multidimensional statistical distance methods for the comparison process in question It can be used. For example, Mahalanobis distance or similar multivariate distances. 25 The deepfake video was analyzed using measurement methods to the reference realism model. The level of similarity can be calculated (1030 elements). As a result of the analysis performed, A final video realism score (scalar) is obtained for the input deepfake video. (Figure) In section 4, this final score is shown as VGS (Video Realism Score) (1031 elements). Using this VGS value, deepfake videos are categorized into 30 levels based on their realism. It is possible to rate videos, identify videos with a low level of authenticity, and analyze data. The sets can be arranged in a realistic manner. Comparison of Experimental Results and Literature (Previous Technique) In this section, the method proposed within the scope of the invention is described in terms of how deepfake videos affect human visual perception. Its success in determining levels of adaptive realism is measured through experimental studies. and is compared with existing IQM methods in the literature (previous technique). Experimental 13 Within the scope of the studies, FaceForensics++ (Rossler et al., which is widely used in the literature, was used. The 2019 deepfake video dataset was used. This dataset includes DeepFakes (DF) and FaceSwap. There are four different manipulation methods: (FS), NeuralTextures (NT), and Face2Face (FF). It includes deepfake videos produced with [computer name] and original source videos. Additionally, the dataset... There are three separate video formats: uncompressed (Raw), medium compression (C23), and high compression (C40). It includes a compression level. In the experiments conducted, a medium compression level was observed. Videos with C23 compression level, which is representative of the dataset, were preferred. The dataset contains 1000 items. Original source videos and 4000 fake videos showcasing different manipulation methods. This research examines different manipulations within the FaceForensics++ dataset. 100 deepfake videos produced using various techniques, and 10 people involved in creating these deepfake videos. 100 original source videos were used. These are from the FaceForensics++ dataset. any information, analysis or assessment given to the videos regarding their visual realism No tags were found. Therefore, deepfake videos in the FaceForensics++ dataset... The videos were thoroughly examined by the inventors, and 50 high-quality videos were selected from among those examined. 50 fake videos with a high level of realism and 15 videos with a low level of realism. The fake video was identified; 100 selected videos were manually categorized into two classes (realistic and unrealistic). It has been tagged. Face images are used in the process of identifying highly realistic videos. the manipulation effect is very small and it is not obvious that the faces are fake Particular attention has been paid to this. Videos with unrealistic labels, on the other hand, contain obvious visual distortions and These videos were selected from among easily recognizable fake videos. Maintaining class balance 20 An equal number of realistic and unrealistic videos were included in the same category. To analyze the performance of the method proposed within the scope of the invention, 100 selected samples were used. For deepfake videos and their sources, primarily BRISQUE (Mittal et al., 2012), DBCNN (Zhang et al., 2012) are mentioned. 2020), LIQE (Zhang et al., 2023), MANIQA (Yang et al., 2022), MUSIQ (Ke et al., 2021) and TRES 25 There are 6 different NR-IQM commonly used in the literature, including (Golestaneh et al., 2022). Quality scores were calculated using this method. NR-IQM methods work based on image quality. facial regions in frame images obtained from source and deepfake videos was used. Then, for each NR-IQM method, a deepfake video was created with the source video. The amount of quality change between them was measured using Cohen's d metric and is shown in Figure 1. As described, six different realism scores (GS-i) were obtained for six different NR-IQM methods. The realism scores obtained for each metric are presented as vectors in Figure 2. By combining the elements, a 6-dimensional fake video realism score vector was created. Realism Realism score vectors for 50 deepfake videos rated as high. The covariance matrix (6x6 dimensional) and mean representation vector shown in Figure 3 were used. A (6x1 dimensional) calculated and reference realism model has been created. Up to this point. The stages described can be defined as model training. In the testing phase, as presented in Figure 4... 14 Each of 100 fake videos selected from the FaceForensics++ dataset using the system (50 6 NR-IQM methods used in the training phase for realistic (50 unrealistic) A three-dimensional GS vector was obtained, and then this GS vector was applied to the data in the training phase. The Mahalanobis distance to the set (realistic model) was calculated, and this value was analyzed. The authenticity score (VGS) for the fake video was recorded. The lowest and highest 5 VGS scores were analyzed. Its value is 0. The closer a deepfake video's VGS value is to 0, the more realistic that video is. It can be said that it is that close to the video model. Otherwise, high VGS values ​​are the reference. This shows the deviation from the realism model. The VGS score is compared with an appropriate threshold. Deepfake videos can be used for authenticity classification. In this study, two classes were used. Within the classifier, the unrealistic class is selected as +, and the realistic class as -. 10 The experiment included 50 realistic and 50 unrealistic deepfakes. ROC (Receiver Operating Characteristic) curve based on VGS scores of the video. The area under the curve (AUC) value was calculated and measured. This is presented in Table 1. The AUC value for the VGS value was found to be 0.91. This value is relevant to this invention. The proposed method was able to successfully distinguish between deepfake videos with good and poor visual quality. 15 It shows. Table 1. Deepfake videos of NR-IQM methods and the proposed invention (VGS). Realism level (for two classes) and detection performance (AUC) comparison Method AUC  BRISQUE 0.44 DBCNN 0.47 MUSIC 0.53 MANIQA 0.48 TRES 0.42 LIQE 0.58 VGS (Patent) 0.91 Table 1 also shows that the realism detection performance of the proposed method is frequently compared with similar methods in the literature. IQM-based methods used (such as BRISQUE, DBCNN, MUSIQ, MANIQA, TRES, LIQE, etc.) A comparison with IQM is also presented. The table includes BRISQUE, DBCNN, MUSIQ, MANIQA, and TReS. and whether IQM metrics are good or bad when LIQE methods are used alone. The performance of separating deepfake videos appears to be quite limited. 25 Table 1 shows that the AUC values ​​obtained for 6 different IQM methods ranged from 0.42 to 0.58. This situation shows that using image quality (IQM) scores alone is ineffective for deepfakes. This indicates that the level of realism is insufficient for analysis and classification. Therefore... The VGS method proposed under this patent is far superior to the 6 IQM methods in the literature. It can provide a deepfake image quality analysis. Figure 5 shows the NR-IQM methods in the literature. 30 The AUC results of the proposed method (VGS) within the scope of the invention are shown in a box (bar) graph. The results obtained are compared. The proposed method is based on classical NR-IQM. It is much better at detecting deepfake videos with a low level of realism compared to other approaches. It shows that it has produced successful results, and this situation indicates the inventive step of the patent application. This strongly supports the view. When the results presented in Table 1 and Figure 5 are evaluated together, The multidimensional realism analysis method and reference realism 5 proposed within the scope of the invention Compared to IQM methods in the literature, this model is more consistent with human visual perception. It appears to provide a reality assessment mechanism. With the proposed method... The realism levels of deepfake videos can be analyzed more reliably, and data can be analyzed accordingly. The clusters can be organized with a focus on realism and more reliably detect deepfake videos. It will be possible to create these systems. 10 The main steps involved in the process are:  In the first stage, the fake video and the source video pair are separated into k image quality pairs. The data is processed using the metric; for each metric, data is obtained from source and sham video frames. The statistical distance between the resulting quality score distributions is used as an intermediate realism score of 15. k scores are determined and combined into k-dimensional vector form to create kx 1-dimensional pseudo-scores. A video realism score vector is being generated.  In the second stage, manually or automatically selected items are chosen from the dataset to be analyzed. For n number of fake videos with a high degree of realism and their source videos (1015), k 3D realism score vectors are being calculated; these vectors are kxn 3D realism 20 The scores are combined in a matrix, and a kxk dimensional covariance matrix is ​​derived from that matrix. A realistic pseudo-video reference model is created by calculating the kx 1-dimensional average representation vector. is created (1019). Here n is the degree of realism used for the reference model. 'h' represents the number of fake videos; 'k' represents the number of different image quality metrics used. It is. 25  In the third stage, the same number of copies will be given for the fake video to be tested and its source video. A test mock video realism score vector is generated using image quality metrics; This vector represents the average representation vector of the realistic pseudo-video reference model. Mahalanobis or similar multivariate statistical distance between covariance matrices It is calculated and the resulting value is produced as the final video realism score. 30 The final score approaching zero indicates that the test video is based on a reference realistic mock video model. The closer the score gets, the further it reflects a departure from visual realism. Detailed explanation of the theoretical aspects of the invention. Deepfake videos are created by manipulating real source videos in various ways. It consists of fake videos produced by applying these methods. As shown in Figure 1. Using the realism score analysis module, the IQM (Integrity Quotient) between the fake video and the corresponding source video was determined. 16 Realism score of the fake video obtained based on Image Quality Metrics score relationship. Intermediate realism is determined in step 1001 based on a single image quality metric. The fake video whose score will be calculated and the source video from which this video was produced will be electronically processed. It is received as input to the unit, then by the user in process step 1003. By selecting an image quality evaluation metric, fake and 5 were detected in process step 1002. Quality of regions of interest in selected (all or part) frame images of source videos. Scores are being calculated. Image quality used in step 1003. Evaluation metrics; BRISQUE, LIQE, MANIQA, MUSIQ, DBCNN, TReS or similar. one of the reference or non-reference image quality assessment methods (IQM, NR-IQM) It is possible. After calculating the IQM scores of the source video and the fake video, number 1004 is the correct answer. Statistical analysis of the quality score distributions of the fake and source videos during the processing step. The statistical distance (realism score: GS) between the two distributions is measured at this stage. There are no method limitations in distance measurement. Thus, the selected image quality Intermediate realism score of the fake video given as input for evaluation metric 1003 (1005) This score is obtained. This score determines the final fake video realism score (VGS) (1031) 15 It is used as an interim score in this stage and is called GS-i. Here, the number i represents... It represents the number / index of the image quality evaluation method selected by the user. Figure 2 shows multiple image quality options for the mock video given as input. (1007, 1008, 1009) were generated using the evaluation metric (k different IQM methods). The steps for calculating the realism score vector GS (1014) are shown. 20 numbered 1006. In the processing step, the same pair of fake video and source video is used to create a multidimensional realism score vector. k different images expressed by process steps numbered (1007, 1008, 1009) to be created The quality metrics are transferred to analysis. A different image quality evaluation is performed in each module. Realism scores (GS-1, GS-2, ... , GS-k) for the fake video are generated using the IQM metric. (These outputs are shown as 1010, 1011, and 1012.) The k realism scores obtained are 1013 and 25. In step number 1014, it is combined into a vector form, and in step number 1014... A 1D realism score vector is being generated for the fake video (kx, number 1014). In this processing step, this vector is referred to as the GS vector. It was created for a fake video. The realism score vector can be expressed as follows. Here, GS(i) is the realism score vector of the i-th fake video; GS-1(i), GS-2(i), ... , GS-k(i) is the i-th fake video. image quality evaluation metrics (IQM) obtained for the fake video These represent realism scores. Thus, a fake video can only be judged on a single quality metric. not with, but with, multidimensional realism information obtained from different quality assessment metrics. This is represented using [method]. Figure 3 shows highly realistic fake videos. 35 using the covariance matrix (kxk dimensional) of the reference realistic mock video model and The generation of the average representation vector (kx 1-dimensional) is shown. This part of the invention 17 This can be defined as the training phase of the proposed method. Realistic fake video The quality model is used as input during the final video realism score (VGS) determination phase. The realism score vector of a fake deepfake video compared to the distribution of realistic fake videos. It is used in measuring similarity. Figure 3 shows the realism used as input. Videos with a high viewing level are determined manually or automatically by the user, and 5 The videos have such high visual quality that it's almost impossible to tell they're fake. It is expected that in step 1015 of the process shown in Figure 3, there are n high-realism levels. The fake video and n source videos belonging to this video are entered into the electronic processing unit. is obtained. Then, in process step 1016, each fake video applied to the input Realism 10 was achieved using the processing steps described in Figure 1 and Figure 2 for the source video pair. Score vectors are calculated for each highly realistic fake video. A k-dimensional realism score vector is obtained. The obtained realism score vectors are further... then combined in vector form to form a kxn dimensional Realism Score Matrix (GSM) is created (1017). Here k is the number of different IQM methods used, and n is 1015 The user-selected, high-realism level specified in step number 15 It indicates the number of fake videos. The realism score matrix GSM (kxn dimensional) is generally used. It can be expressed as follows: Here, the realism of fake videos with a high level of realism is discussed. They represent the score vectors (columns of the GSM matrix). Then operation number 1018 is 20 In this step, the covariance of the realistic fake video model is calculated using the realism score matrix. The matrix Σ (kxk dimensional) and the mean representation vector μ (kx1 dimensional) are calculated. The elements of the average representation vector are calculated as follows: Here, μ(i) represents the i-th element of the mean representation vector of the realistic mock video model. GSM(i,j) represents the i-th row and j-th column of the GSM matrix. i takes values ​​in the range 1,...,k. It represents the kx 1-dimensional mean representation vector and the rows of the μ GSM matrix. It is calculated by taking the averages. Similarly, the realism score matrix belongs to The covariance matrix can be expressed as follows: Here, Σ represents the covariance matrix showing the relationships between the realism score vectors. The calculated covariance matrix and mean representation vector together form number 1019. It creates a realistic mock video model in the process step. Figure 4 shows a mock video. and when the source video is received as input to the electronic processing unit, the final video belongs to the fake video. The process of obtaining the realism score (VGS) is shown. This stage demonstrates the proposed method for testing 35. 18 It can be defined as a stage. In Figure 4, the input given in process step number 1020 is... For both the fake video and the source video, the elements described in Figure 1 and Figure 2 are applied, and In step 1028, the realism score vector of the fake video is obtained. This Realism score analysis modules in steps (1021, 1022, 1023) Realism scores for different quality assessment metrics are calculated using 5 (1024, 1025, 1026), then the scores obtained are vectored in step 1027. They are combined in this way. Then, the kx 1-dimensional structure obtained in step 1028 is combined. The operations shown in Figure 3 were applied in step 1029 with the realism score vector. The covariance matrix and mean representation vector of the realistic fake video model obtained. Using this, the Mahalanobis distance is calculated in step 1030. This 10 The distance value tested on the fake video indicates a difference in visual quality compared to realistic fake video quality. It represents the statistical distance to the model. The value obtained from this distance measurement. The input deepfake / fake video is referred to as the Video Realism Score (VGS). The final Video Realism Score (VGS) based on Mahalanobis distance is as follows: It is calculated as: 15 Here, GS is the realism score vector of the tested mock video, and μ is the realistic mock video vector. Σ represents the mean representation vector of the model, and Σ represents the covariance matrix. Last In this case, the output of the method is the fake input applied to the input in step 1031. The video's final realism score is generated and recorded as VGS. 20 The steps involved in the process are as follows: – This fake video will have an interim realism score calculated based on a single image quality metric. The source video from which the video was produced is taken as input to the electronic processing unit (1001). – Pre-processing of selected frames or faces or areas of interest from the introductory videos. and based on the image quality metric selected from these regions, for both fake and source video. Calculation of separate quality score distributions (1002). – Selecting the image quality metric method to be used in calculating the interim realism score. (1003). – The statistically significant difference between the quality score distributions of the fake and source videos is 30. measurement (1004). – Intermediate realism score (GS-i) of the measured statistical distance for the selected image quality metric. produced as (1005). – A multi-dimensional realism score vector will be created using the fake video and its corresponding source video. The pair uses 35 as input to k different realism analysis modules for k different image quality metrics. transfer (1006). 19 – Performing the first realism score analysis according to the first image quality metric. (1007). – Performing a second realism score analysis based on the second image quality metric. (1008). – Performing the Kth realism score analysis according to the Kth image quality metric 5 (1009). – Obtaining the first intermediate realism score (GS-1) (1010). – Obtaining the second intermediate realism score (GS-2) (1011). – Obtaining the kth intermediate realism score (GS-k) (1012). – Combining the k intermediate realism scores obtained in vector form (1013). 10 – Generation of a kx 1D realism score vector (GS vector) for the fake video. (1014). – To create a reference model, n highly realistic fake videos were used, and these were... n source videos belonging to the electronic processing unit are received as input (1015). – For each pair of fake videos and source videos received at the input, a k-dimensional realism score vector of 15 calculation (1016). – n realism score vectors in a kxn dimensional realism score matrix (GSM) merging (1017). – kxk-dimensional covariance matrix and kx 1-dimensional mean representation from the realism score matrix. Calculation of the vector (1018). 20 – Realistic mock video reference including covariance matrix and mean representation vector. creation of the model (1019). – For testing purposes, a mock video will be used to calculate the final video realism score, along with its accompanying data. Receiving the source video as input to the electronic processing unit (1020). – Performing the first test realism score analysis according to the first image quality metric 25 (1021). – Performing a second test realism score analysis based on the second image quality metric. (1022). – Performing a K-th test realism score analysis based on the K-th image quality metric. (1023). 30 – Obtaining the first test interim realism score (GS-1) (1024). – Obtaining the second test interim realism score (GS-2) (1025). – Obtaining the k-th test interim realism score (GS-k) (1026). – Combining the k test realism scores obtained in vector form (1027). – Creation of the kx 1D realism score vector (GS vector) for the test mock video 35 (1028). – Covariance matrix and mean representation vector of the realistic pseudo-video reference model to be taken as an introduction to the testing phase (1029). – Realistic fake video reference with fake video realism score vector for testing phase. Mahalanobis or similar multivariate statistical distance between models calculation (1030). 5 – The calculated distance is generated as the final video realism score (VGS) and electronically recording to the recording medium (1031).

Claims

21 REQUESTS 1. Deepfake videos have a level of realism that is compatible with human visual perception. Detected by electronic processing units, highly realistic counterfeit. Creating a statistical reference model from the videos and 5 obtained from the fake videos based on the distances of the quality score vectors from the reference model in question deep fake videos with low levels of realism by rating them a method for detecting and separating these videos from a deep forged dataset Its characteristic is; - To create a multidimensional realism score vector (MR), a fake video was used and 10 The source video pair is used as input to k different reality score analysis modules. transfer (1006), - For each of k different image quality metrics, a mock video was used to create the source. Comparison of video quality score distributions and k intermediate realisms generating the score, 15 - Combining the k intermediate realism scores obtained in vector form (1013), - Generation of a k-dimensional realism score vector (GS vector) for the fake video. (1014), - Using n highly realistic simulated videos to create a reference model. 20 The input of n source videos belonging to these into the electronic processing unit (1015), - k-dimensional realism score for each pair of fake videos and source videos received at the input. Calculation of the vector (1016), - n realism score vectors in a kxn dimensional realism score matrix (GSM) merging (1017), 25 - kxk dimensional covariance matrix and kx 1 dimensional mean from the realism score matrix. Calculation of the representation vector (1018), - Realistic mock video reference including covariance matrix and mean representation vector. creation of the model (1019), - The final video realism score will be calculated using a mock test video and its source code 30 receiving the video as input to the electronic processing unit (1020), - For each of k different image quality metrics, a test mock video will be used to compare them. Comparison of the quality score distributions of the source videos and the k items in question. Generating k test realism scores corresponding to the image quality metric, - Combining the k test realism scores obtained in vector form (1027), 35 22 - kx 1-dimensional realism score vector (GS vector) for the test mock video creation (1028), - Covariance matrix and mean representation of the realistic mock video reference model. taking the vector as input to the test phase (1029), - Test fake video realism score vector with realistic fake video reference model 5 Mahalanobis or similar multivariate statistical distance between them calculation (1030) and - Generating the calculated distance as the final video realism score (VGS) and recording in electronic recording medium (1031) It includes the steps of the process. 10 2. This method complies with Claim 1 and is characterized by its use of k different image quality metrics. The process of generating k intermediate realism scores; - First realism score analysis according to the first image quality metric. implementation (1007), 15 - Performing a second realism score analysis based on the second image quality metric. (1008), - Performing the Kth realism score analysis according to the Kth image quality metric. (1009),  Obtaining the first intermediate realism score (GS-1) (1010), 20  Obtaining the second intermediate realism score (GS-2) (1011), - Obtaining the kth intermediate realism score (GS-k) (1012), It includes the steps of the process.

3. The method compliant with Claim 1 is characterized by its ability to include a test mock video and its corresponding source video. k tests based on k different image quality metrics yielded a realism score of 25. the process of producing; - For the testing process of the first realism score analysis according to the first image quality metric. implementation (1021), - For the testing process of the second realism score analysis according to the second image quality metric. implementation (1022), 30 - For the testing process of the Kth realism score analysis according to the Kth image quality metric. implementation (1023), - Obtaining the first intermediate realism score (GS-1) (for the test) (1024), - Obtaining the second intermediate realism score (GS-2) (for the test) (1025),  Obtaining the Kth intermediate realism score (GS-k) for the test) (1026) 35 It includes the steps of the process. 23 4. This method complies with Claim 1 and its characteristic is that it provides interrealism for each image quality metric. the single analysis sub-stream that enables the generation of the score; - A mock video will have an interim realism score calculated based on a single image quality metric.

5. The source video from which this video was produced is taken as input to the electronic processing unit. (1001), - The image quality metric method to be used in calculating the interim realism score. selection (1003), - Faces in selected frames from the intro videos (both for the fake and the source video) or preprocessing of regions of interest, selecting image quality 10 for these regions. Calculation of IQM scores according to the metric (IQM) (1002), - The statistical distance between the quality score distributions of the fake and source videos. measurement (1004) and - The statistical distance calculated in step 1004 to the selected image quality metric The intermediate realism score (GS-i) of (1005) is produced as 15 It includes the steps of the process.

5. This method complies with Claim 4 and its characteristic is that the image quality metric is determined in process step 1002. their scores were based on facial expressions or points of interest in all or some of the frames selected from the videos. Calculation on regions and face detection, alignment, cropping on those frames 20 and the process step must include the application of at least one of the preprocessing operations.

6. The method is compliant with claim 4 or 5, and its characteristic is that it differs from each other in transaction step 1003. different image quality metrics of BRISQUE, DBCNN, LIQE, MANIQA, MUSIQ, TReS and similar non-referenced image quality metrics, referenced image quality metrics, or human 25 Selecting one of the deep learning-based quality analysis methods related to visual perception It includes the process step.

7. This method is suitable for any of the previous requests and its characteristic feature is that it includes fake and source videos. The statistical distance between the image quality metric distributions is measured using Cohen's d or 30. Measurement using a similar statistical distance method and i. the realism score of the fake video. the process of creating the vector through its relationship It includes the step.

8. A method that conforms to any of the previous requirements, characterized by having n degrees of realism. vectors of realism score of high fake video 24 in the kxn-dimensional realism score matrix with its relationship It includes the merging process step.

9. The method is compliant with Claim 8 and is characterized by its average of realistic fake video reference models. the relationship between the elements of the representation vector and the covariance matrix 5 It involves the calculation step of the relationship.

10. A method that is suitable for any of the previous requests, and its characteristic is; belonging to the test mock video. with the equation of the final video realism score The calculation involves the step of determining the Mahalanobis distance using the defined relationship. 10 11. This method complies with Claim 10 and its characteristic is that the final video is produced in process step 1031. The test video, where the realism score approaches zero, is a realistic mock video reference. The score's proximity to the model indicates that the score is higher, while an increase in the score indicates a greater distance from that model; The video is deemed realistic or unrealistic based on a comparison of the score to an acceptable threshold. labeling as non-existent and separating the unrealistic video from the dataset It includes the process step.

12. This method complies with Claim 1 and its characteristic is that it has n realities in step 1015. Manual or automatic identification of highly fraudulent videos, number of videos n 20 The number of image quality metrics, k values, are selected according to the application, and number 1031. The final video realism score generated in the processing step is recorded on an electronic medium. The process involves recording the data and presenting it as a printout via an electronic device.

Images