Deep fake face detection method based on fusion boundary features

Abstract

The invention relates to the technical field of face recognition, in particular to a deep counterfeit face detection method based on fusion boundary features, and the method comprises the steps: specifying a light source requirement according to the skin color contrast in an initial image of a target recognition user; determining whether the specified light source execution requirement is qualified based on the facial brightness difference; determining an action direction of a specified action based on the face brightness difference; acquiring a frame-by-frame image when the target identification user makes a specified action so as to identify a hair and face junction area of each frame of image, and extracting a fusion boundary line; dividing into a plurality of boundary sections based on the fusion boundary line; acquiring corresponding fusion boundary coordination parameters based on the plurality of boundary segments to determine a plurality of high-risk boundary segments; calculating verification consistency characterization parameters based on the boundary time sequence and the light and shadow time sequence of the plurality of high-risk boundary segments; and determining a fake face risk value of the target recognition user based on the verification consistency characterization parameter and the high-risk boundary section proportion. According to the invention, the dynamic face counterfeiting detection precision is improved.

Description

Technical Field

[0001] This invention relates to the field of facial recognition technology, and in particular to a deepfake face detection method based on fused boundary features. Background Technology

[0002] With the rapid development of deep learning and generative adversarial networks, deepfake technology, especially high-fidelity face replacement and synthesis techniques, has reached a level of realism that is indistinguishable from genuine faces. This type of technology is being abused to create fake news, commit online fraud, and infringe on personal privacy, posing a serious challenge to social trust systems and information security. Therefore, developing efficient and accurate deepfake detection technology has become a critical issue that urgently needs to be addressed.

[0003] However, existing detection methods largely focus on analyzing frequency domain statistical anomalies, facial biosignal inconsistencies, or single generative model artifacts in a single forged image. These methods have reached their limits when faced with increasingly realistic static forged images. More importantly, they generally fail to adequately consider the fundamental flaws exposed by forged content in dynamic video sequences, especially in interactions with a physical environment—namely, the inherent differences in physical properties between the forged area and the real background, and the inconsistencies and discrepancies that these differences exhibit in the spatiotemporal dimensions. This results in limited detection capabilities and high false negative and false positive rates for high-quality, dynamic deepfake faces, particularly those forged content without obvious visual flaws in a single frame.

[0004] Chinese Patent Publication No. CN120954108A discloses a method, device, and medium for identifying face forgery, including: determining a first frame image from a video to be identified; wherein the first frame image is a frame image in which the amplitude of facial movements exceeds a threshold; acquiring first optical flow motion information of the first frame image and second optical flow motion information of a second frame image; wherein the sampling time of the second frame image is before the sampling time of the first frame image; performing gradient calculation on the first frame image to obtain first gradient information corresponding to the first frame image; and obtaining a face forgery identification result of the video to be identified based on the first optical flow motion information, the second optical flow motion information, and the first gradient information.

[0005] Therefore, it is evident that the existing technology has the following problems: Current deepfake detection methods do not consider actively specifying the physical environment to trigger forgery traces, nor do they fully consider the inconsistencies in the spatiotemporal characteristics of the forged area and the light and shadow under dynamic interaction, resulting in low accuracy in detecting dynamic face forgeries. Summary of the Invention

[0006] To address this, the present invention provides a deep face forgery detection method based on fused boundary features, which overcomes the problems of low detection accuracy for dynamic face forgery caused by existing deep face forgery detection technologies that do not consider actively specifying the physical environment to trigger forgery traces and do not fully consider the spatiotemporal inconsistency of the forgery region and the light and shadow under dynamic interaction.

[0007] To achieve the above objectives, this invention provides a deepfake face detection method based on fused boundary features, comprising: The light source requirements are specified based on the skin color contrast in the user's initial image to identify the target. Acquire images of the target recognition user under specified light source requirements, and determine whether the specified light source requirements are met based on facial brightness differences; If the requirements of the specified light source are met, the direction of the action is determined based on the facial brightness difference, so that the user can complete the specified action in the direction of the action. The target recognition is obtained frame by frame based on the time sequence of the user performing the specified action, in order to identify the hair and face boundary area in each frame. Extract the fusion boundary line based on the area where the hair meets the face; The number of boundary segments is determined based on the changes in the length and curvature of the fusion boundary line of each frame image, so as to divide the image into several boundary segments. Based on several boundary segments, corresponding fusion boundary coordination parameters are obtained to determine several boundary segments of interest. The fusion boundary coordination parameters include color transition smoothness, texture continuity index, noise distribution consistency, and gradient direction matching degree. Based on the location and proportion of the aforementioned boundary segments of concern, several high-risk boundary segments are identified. Light and shadow feature parameters are obtained based on several high-risk boundary segments, wherein the light and shadow feature parameters include highlight intensity value, shadow direction value, reflection equalization value and light and shadow gradient value; Based on the fusion boundary coordination parameters and light and shadow performance parameters of several high-risk boundary segments, the boundary time sequence and light and shadow time sequence are determined to calculate the verification consistency characterization parameters, wherein the verification consistency characterization parameters include the direction change consistency degree and the hysteresis degree. The risk value of a user's fake face is determined based on the weighted fusion of the verification consistency representation parameters and the proportion of high-risk boundary segments.

[0008] Furthermore, the process of specifying light source requirements based on skin tone contrast in the user's initial image, as determined by the target identification, includes: Extract the target user's facial region from the initial image, and calculate the average brightness value of skin tone and the average brightness value of background respectively to determine skin tone contrast. The light source requirements are determined based on the skin color contrast, wherein the light source requirements include the light source type and the light angle, and the light source type includes point light sources and diffuse light.

[0009] Furthermore, the process of acquiring images of the target recognition user under specified light source requirements, and determining whether the specified light source requirements are met based on facial brightness differences, includes: Acquire images of the target recognition user under specified light source requirements, extract facial regions, and divide the face into left and right halves; Calculate the average brightness value of the left and right halves of the face to determine the difference in facial brightness. The facial brightness difference is compared with a preset facial brightness difference threshold to determine whether the requirements for executing the specified light source are met.

[0010] Furthermore, assuming the requirements for executing the specified light source are met, the process of determining the direction of the specified action based on the facial brightness difference includes: Based on the facial brightness difference, the darker side of the image that meets the requirements of the specified light source is determined, so as to determine the direction of the specified action.

[0011] Furthermore, the process of determining the number of boundary segments based on the changes in the length and curvature of the fusion boundary lines of each frame image, in order to divide the image into several boundary segments, includes: Calculate the total length of the fusion boundary line to determine the number of basic segments based on a preset length interval; Calculate the curvature of each point on the fusion boundary line, and determine the number of modified segments based on the number of curvature extreme points and the number of basic segments to divide the boundary into several segments.

[0012] Furthermore, the process of obtaining corresponding fusion boundary coordination parameters based on several boundary segments to determine several boundary segments of interest includes: The fusion boundary coordination degree of each boundary segment is determined by weighted averaging the fusion boundary coordination parameters corresponding to each boundary segment. If the fusion boundary coordination degree is less than the preset fusion boundary coordination degree threshold, the corresponding boundary segment is marked as a boundary segment of interest.

[0013] Furthermore, the process of determining several high-risk boundary segments based on the location and proportion of the aforementioned boundary segments of concern includes: Several high-risk boundary segments are determined based on several boundary segments of interest corresponding to the first preset location conditions. And determine several high-risk boundary segments based on the first preset proportion condition; The first preset location condition is that the boundary segment of concern is located in a potentially high-risk area, and the potentially high-risk area is determined based on the existence of a preset number of consecutive boundary segments of concern. The first preset proportion condition is that the proportion of the length of each segment of interest to the length of the fusion boundary line is greater than a preset threshold.

[0014] Furthermore, the process of determining the boundary temporal sequence and the light and shadow temporal sequence based on the fusion boundary coordination parameters and lighting and shadow performance parameters of several high-risk boundary segments, in order to calculate and verify the consistency characterization parameters, includes: The fusion boundary coordination degree corresponding to the fusion boundary coordination parameters of each high-risk boundary segment is arranged into a boundary time series. A weighted average is performed on the light and shadow feature parameters of each high-risk boundary segment to determine the light and shadow feature characterization value corresponding to each high-risk boundary segment, and then the light and shadow feature characterization values ​​are arranged into a light and shadow time sequence. The lag is calculated based on the boundary time series and light and shadow time series of the aforementioned high-risk boundary segments; Based on the boundary time sequence and the light and shadow time sequence, the characteristics of coordinated change of the fusion boundary and the characteristics of change of light and shadow performance are determined respectively, so as to calculate the consistency of the direction change; The characteristics of coordinated change at the fusion boundary include mean drift amplitude, fluctuation frequency, mutation density, and trend correlation; the characteristics of change in light and shadow performance include instantaneous jump intensity, inertial intensity, self-similar decay rate, and periodic significance.

[0015] Furthermore, the process of calculating the lag degree based on the boundary time series and light and shadow time series of the aforementioned high-risk boundary segments, and calculating the consistency of the trend change based on the coordinated change characteristics of the fused boundary and the change characteristics of the light and shadow performance, includes: Based on the boundary time sequence and light and shadow time sequence of each high-risk boundary segment, the time shift between the two sequences is calculated, and the time shift is normalized to determine the alignment time shift amount of the corresponding high-risk boundary segment. The hysteresis is determined based on the mean of the alignment time shift of all high-risk boundary segments; Based on the fusion boundary coordination change characteristics and light and shadow change characteristics of each high-risk boundary segment, corresponding feature vectors are constructed respectively. Similarity is calculated based on the feature vectors, and the consistency of trend change is determined based on the average similarity of all high-risk boundary segments.

[0016] Furthermore, the process of determining the risk value of a fake face for a target user based on the weighted fusion of the verification consistency representation parameters and the proportion of high-risk boundary segments includes: The weighted fusion calculation of the change consistency comprehensive score is based on the verification consistency characterization parameters; The risk value of a user's fake face is determined by weighted fusion based on the overall score of consistency of changes and the proportion of high-risk boundary segments.

[0017] Compared with existing technologies, the beneficial effects of this invention lie in providing a systematic solution to the problem of deepfake detection by constructing a complete closed-loop process from environmental standardization and proactive interactive guidance to multi-dimensional temporal feature analysis. By specifying the light source and guiding the user to perform specific actions, the originally passive and random video analysis is transformed into an active and standardized stress test, thereby stably stimulating the most vulnerable features of the forged content at the fusion boundary. By comprehensively analyzing the static coordination of the hair-face junction region and the temporal consistency of dynamic lighting and shadow, this method can capture physical law violations and spatiotemporal inconsistencies that are difficult for generative models to simulate. Finally, by fusing evidence of temporal inconsistency and evidence of spatial anomalous distribution to calculate the risk value, the detection results are more robust and reliable, significantly reducing the dependence on the quality of a single image or a single feature, and improving the accuracy of identifying highly realistic and dynamically forged faces.

[0018] Furthermore, this invention specifies the light source type and angle based on the user's initial skin tone saturation, enabling preliminary personalized adaptation of detection conditions and laying the foundation for obtaining high-quality analysis data. Determining the light source type based on skin tone saturation ensures that the recommended light source is easy for users to obtain and set, while also producing a clear light and shadow direction, which is beneficial for highlighting boundary features later. The uniform and clear light angle guidance greatly reduces the user's understanding and execution threshold, ensuring the method's operability in general scenarios and avoiding decreased user cooperation due to complex parameter adjustments, thus ensuring the smooth start and execution of the entire detection process.

[0019] Furthermore, this invention provides an objective and measurable standard for qualified lighting by quantitatively calculating the brightness difference between the left and right halves of the face and comparing it with a preset threshold range. This step, as a quality control step, ensures that the input video data has appropriate light and shadow contrast—neither too flat to stimulate features nor too strong to produce interfering shadows. Clearly defined pass / fail criteria improve the quality of input data for all subsequent feature extraction and analysis steps from the source, and are a key prerequisite for ensuring the stable and reliable detection performance of the entire method.

[0020] Furthermore, this invention utilizes the inherent brightness distribution information in qualified illumination images, including the brighter and darker sides, to simply and directly determine the direction of head turning, thus automating and intelligentizing the detection logic. This design ensures that the action command is strongly correlated with the current lighting conditions, guaranteeing that the user's head rotation generates the most effective relative motion with the fixed light source. This rule of turning to the darker side maximizes the amplitude and continuity of light and shadow changes in the fusion boundary region during the action, thereby creating optimal experimental conditions for subsequent accurate calculation of key temporal features related to light effect response delay and change consistency, and enhancing the method's sensitivity to forgery traces.

[0021] Furthermore, this invention achieves intelligent adjustment of the fusion boundary line analysis by combining fixed-length intervals with curvature extrema for adaptive segmentation. Basic segmentation based on length intervals ensures the comprehensiveness and uniformity of the analysis, avoiding the omission of any region. Adding segments based on curvature extrema allows for more precise local focusing on key areas with drastic contour changes, such as hairline corners, where forgery traces are easily detected, without significantly increasing computational load. This coarse-to-fine segmentation strategy enables feature extraction to cover both the overall picture and delve into details, improving the method's ability to capture local anomalies and enhancing analytical efficiency.

[0022] Furthermore, this invention achieves rapid and automated initial screening of anomalous boundary segments by comprehensively weighting four different dimensions of fusion boundary coordination parameters into a single fusion boundary coordination degree index and setting a clear threshold for filtering. This method simplifies complex multi-feature judgment into a scalar comparison, significantly improving processing efficiency. The weighted averaging method allows for flexible adjustment based on the importance of different parameters, while the threshold setting is based on real data distribution, ensuring the scientific nature of the screening. This step effectively filters out a large number of boundary segments that appear normal, allowing subsequent computational resources to be focused on in-depth analysis of a few boundary segments of interest, improving analysis efficiency and targeting.

[0023] Furthermore, this invention employs two complementary rules—spatial continuity and anomaly proportion—to screen high-risk boundary segments, making the judgment logic more comprehensive and robust. The spatial continuity rule can keenly capture obvious forgery artifacts that appear concentrated in local areas; these traces are usually highly indicative. The anomaly proportion rule, on the other hand, can prevent missed detections caused by scattered but widely distributed forgery traces due to insignificant single-point anomalies. The two rules work together to ensure that both locally obvious suspected forgeries and globally weak suspected forgeries can be effectively identified and prioritized for further investigation, enhancing robustness.

[0024] Furthermore, this invention extracts higher-order variation features from the original temporal sequence of coordination and light and shadow representation values, achieving a deep characterization of temporal behavior patterns. These features go beyond simple numerical comparisons and can describe the stability, regularity, and synchronicity of the sequence. This higher-order description of temporal patterns provides richer and more essential characteristic evidence for ultimately determining whether it conforms to natural physical laws.

[0025] Furthermore, this invention calculates hysteresis through cross-correlation analysis and directional change consistency through eigenvector cosine similarity, providing a precise and calculable mathematical measure for assessing temporal consistency. Hysteresis quantifies the delay of light and shadow changes relative to boundary attribute changes, directly reflecting the sluggishness commonly found in forged content. Direction change consistency, by comparing the directions of two sets of higher-order eigenvectors, assesses whether static attribute change patterns and dynamic light and shadow change patterns are coordinated in their overall trend, avoiding the difficulty of directly comparing parameters with different physical meanings. These two parameters provide interpretable mathematical evidence for inconsistencies from both temporal and pattern dimensions.

[0026] Furthermore, this invention achieves comprehensive decision-making regarding spatiotemporal evidence by weightedly fusing the overall consistency score of changes with the proportion of high-risk boundary segments to arrive at the final forgery risk value. The advantage of this fusion mechanism lies in its consideration of both the dynamic flaws (inconsistencies) exhibited by forgery traces in the temporal dimension and their widespread spatial distribution (anomaly proportion). A single temporal inconsistency might be caused by accidental factors, but if it is accompanied by a high proportion of spatial anomalies, the likelihood of forgery increases significantly. This comprehensive assessment avoids misjudgments that may result from relying on a single type of evidence, making the final risk value more comprehensive and stable, and providing a more sufficient basis for decision-making. Attached Figure Description

[0027] Figure 1 This is a flowchart of a deepfake face detection method based on fused boundary features according to an embodiment of the present invention; Figure 2 This is a flowchart illustrating the process of specifying light source requirements based on skin color contrast in the user's initial image according to an embodiment of the present invention; Figure 3 This is a flowchart illustrating how to determine whether the requirements for a specified light source are met based on facial brightness differences, according to an embodiment of the present invention. Figure 4 This is a flowchart illustrating the process of dividing boundary segments based on the changes in the length and curvature of the fusion boundary lines of each frame image, as described in an embodiment of the present invention. Detailed Implementation

[0028] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0029] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0030] It should be noted that, in the description of this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0031] For ease of understanding, the scenarios and terms used in this invention are explained as follows: The deepfake face detection method based on fused boundary features of the present invention is applied to real-time face detection to determine that the behavior of issuing instructions is the user's true intention. For example, it can be applied to various verification scenarios such as APP, application, and terminal device that require authorization. The initial image is the user's current image initially acquired by the camera device when performing face detection.

[0032] Please see Figure 1 The diagram shows a flowchart of a deepfake face detection method based on fused boundary features according to an embodiment of the present invention. The deepfake face detection method based on fused boundary features according to an embodiment of the present invention includes: Step S1: Specify the light source requirements based on the skin color contrast in the user's initial image for target identification; Step S2: Obtain the image of the target recognition user under the specified light source requirements, and determine whether the specified light source requirements are met based on the facial brightness difference; Step S3: If the requirements for executing the specified light source are met, determine the direction of the action based on the facial brightness difference so that the user can complete the specified action in the direction of the action. Step S4: Obtain frame-by-frame images of the target recognition user performing a specified action according to the time sequence, so as to identify the hair and face boundary area of ​​each frame image. Step S5: Extract the fusion boundary line based on the area where the hair meets the face; Step S6: Determine the number of boundary segments to be divided based on the length and curvature changes of the fusion boundary lines of each frame image, so as to divide the image into several boundary segments. Step S7: Obtain corresponding fusion boundary coordination parameters based on several boundary segments to determine several boundary segments of interest. The fusion boundary coordination parameters include color transition smoothness, texture continuity index, noise distribution consistency, and gradient direction matching degree. Step S8: Determine several high-risk boundary segments based on the location and proportion of the aforementioned boundary segments of concern; Step S9: Obtain light and shadow feature parameters based on several high-risk boundary segments, wherein the light and shadow feature parameters include highlight intensity value, shadow direction value, reflection equalization value, and light and shadow gradient value; Step S10: Determine the boundary time sequence and the light and shadow time sequence based on the fusion boundary coordination parameters and light and shadow performance parameters of several high-risk boundary segments, so as to calculate the verification consistency characterization parameters, wherein the verification consistency characterization parameters include the direction change consistency degree and the hysteresis degree. Step S11: Determine the risk value of the fake face of the target user based on the weighted fusion of the verification consistency characterization parameters and the proportion of high-risk boundary segments; Step S12: Based on the fact that the fake face risk value is in the preset fake face risk value area, a correction light source requirement and action direction are issued to update the fake face risk value.

[0033] Please see Figure 2 As shown, it is a flowchart of an embodiment of the present invention that specifies the light source requirements based on the skin color contrast in the user's initial image.

[0034] Specifically, in step S1, the process of specifying the light source requirements based on the skin color contrast in the target-identified user's initial image includes: Step S1001: Extract the target recognition user's facial region from the initial image, and calculate the average brightness value of skin tone and the average brightness value of background respectively to determine skin tone contrast. Step S1002: Determine the light source requirements based on the skin color contrast, wherein the light source requirements include the light source type and the light angle, and the light source type includes point light source and diffuse light.

[0035] In this embodiment, a face detection algorithm such as MTCNN in the prior art is used to crop the facial region from the user's initial image. Any method in the prior art can also be used, which will not be elaborated here. In the LAB color space of the facial region pixels, the average brightness value of the skin color, usually referring to the main area such as the cheeks, and the average brightness value of the background / non-skin color area, such as hair and dark clothing, are calculated respectively. Then, the skin color contrast can be defined as the ratio of the absolute value of the difference between the two to the mean of the two, that is: Skin color contrast = |mean value of skin color brightness - mean value of background brightness| / [(mean value of skin color brightness + mean value of background brightness) / 2].

[0036] In step S1002, a light source is selected from the preset light source library based on skin tone contrast. The selection considers both effect and feasibility, prioritizing point light sources due to their strong directionality and ease of user setup, such as desk lamps and spotlights. These sources also produce clear highlights and shadows, which are most beneficial for highlighting boundaries. For users with low skin tone contrast, i.e., skin tone contrast less than the preset skin tone contrast threshold, diffuse light is added to soften shadows and preserve details until the skin tone contrast is greater than or equal to the preset skin tone contrast threshold. Preferably, the preset skin tone contrast threshold is set to 0.2. The lighting angle is uniformly set to a 45-degree angle from the upper left, which is easy for users to understand and reproduce. This establishes the left-right axis of symmetry of the user's face. The system then instructs the user to "ensure the main light source is located at a 45-degree angle from the upper left" or "a 45-degree angle from the upper right" on the camera display screen, or provides a corresponding indication of the light source's position (e.g., using a circular dotted line to mark the point light source position for user understanding and execution). The determination of whether the point light source is on the left or right side is based on the brightness uniformity index of the left and right halves of the face in the preliminary test image. The standard deviation of pixel brightness in each half of the face is calculated; a smaller standard deviation indicates greater uniformity, and the side with the smaller standard deviation is selected as the recommended light source side. If the uniformity on both sides is similar, the left side is selected by default.

[0037] It is understood that in this invention, by first setting the light source requirements, an environment that is easy to perform fake face recognition is constructed. In order to accurately identify the probability and risk of fake faces by analyzing the consistency of light and shadow features and action features in frame-by-frame images when performing image analysis of a specified action.

[0038] Please see Figure 3 As shown, it is a flowchart of an embodiment of the present invention for determining whether the requirements for executing a specified light source are qualified based on facial brightness difference.

[0039] Specifically, in step S2, the process of acquiring an image of the target recognition user performing the specified light source requirement, and determining whether the performance of the specified light source requirement is qualified based on the facial brightness difference, includes: Step S21: Obtain the image under the specified light source requirements for target recognition performed by the user, extract the facial region and divide the left and right halves of the face; Step S22: Calculate the average brightness value of the left and right halves of the face to determine the facial brightness difference; Step S23: Compare the facial brightness difference with a preset facial brightness difference threshold to determine whether the requirements for executing the specified light source are met.

[0040] In this embodiment, after the user adjusts the light source according to the instructions and maintains a frontal view, a test image is captured. Facial landmark detection technology is used to locate the facial contours, and the facial area is divided into the left and right halves along the midline of the nose. The average pixel brightness values ​​of the two regions on the grayscale image are calculated separately. The formula for calculating the facial brightness difference is: Facial brightness difference = |Left half face brightness - Right half face brightness| / [(Left half face brightness + Right half face brightness) / 2]. Preferably, the preset acceptable threshold range is a lower limit of 0.15 and an upper limit of 0.40. That is, if the calculated brightness difference is between 0.15 and 0.40, the current light source requirement is considered acceptable, and the light contrast is moderate. If the brightness difference is less than 0.15, it indicates insufficient light contrast; please increase side lighting or adjust the angle. If the brightness difference is greater than 0.40, it indicates excessive light contrast; please reduce the light or increase supplementary lighting. The user can adjust according to the prompts. Preferably, three judgment cycles are allowed to ensure that the acquisition conditions meet the standards.

[0041] Specifically, in step S3, the process of determining the direction of the specified action based on the facial brightness difference, assuming that the requirements for executing the specified light source are met, includes: Step S31: Based on the facial brightness difference, determine the darker side in the image that meets the requirements of the specified light source, so as to determine the direction of the specified action.

[0042] In this embodiment, the average brightness values ​​of the left and right halves of the face have been calculated for the qualified image determined in step S2. By directly comparing these two values, the brighter and darker sides of the left and right halves of the face can be determined. The determination of the direction of movement follows a core principle: guiding the user to turn their head from the brighter side to the darker side. For example, if the analysis shows that the user's left face is brighter, i.e., the light source is in the upper left, the generated instruction is "Please slowly turn your head to your right." This design ensures that during the head-turning process, the bright area that was originally directly illuminated by the light source, including the fusion boundary to be detected, will gradually move into the shadow, or the shadow area will be illuminated, thereby generating a maximized and continuous change in light and shadow in time, creating conditions for subsequent detection of lag and inconsistency.

[0043] In this embodiment, in step S4, the system records the entire process of the user performing a specified head-turning action at a fixed frame rate, such as 30fps. After decoding the video frame by frame, a face detection and keypoint alignment algorithm, such as MTCNN, is first applied to each frame to ensure that the facial region is stably tracked and cropped to a normalized size. Subsequently, for each cropped facial image, a lightweight but high-precision semantic segmentation neural network is used, for example, the BiSeNet model based on the MobileNetV2 backbone. The training samples of this model include publicly available face parsing datasets (such as CelebAMask-HQ) and self-built hair-face annotation data covering various hairstyles, skin colors, and lighting conditions. The model outputs a probability map of each pixel belonging to hair and facial skin. By performing thresholding processing on the probability map, such as taking the maximum probability category, and morphological post-processing, such as closing operations, to fill small holes, accurate binary masks of the hair region and facial region in each frame are obtained. The adjacent boundary regions of these two masks constitute the hair-face interface region with a width of several pixels, which is extracted for subsequent analysis.

[0044] In this embodiment, in step S5, after obtaining the binary mask of the boundary region for each frame, a representative boundary line with a single pixel width and smoothness needs to be extracted. First, edge detection algorithms such as the Canny operator are applied to the hair mask and the face mask respectively to obtain their initial contour lines. Then, for each point on the contour line, the nearest corresponding point on another contour line is found in the normal direction, and the midpoints of these two points are connected to form an initial but potentially uneven midline. To obtain a more stable and continuous fusion boundary line, a B-spline curve fitting algorithm is used to fit this discrete set of points composed of midpoints. The fitted curve is not only smooth but also retains the geometric features of the original region. Finally, this parametric curve is re-discretized at a fixed sampling interval to generate an ordered sequence of two-dimensional coordinate points. This sequence is the final fusion boundary line for this frame, which will serve as the basis for all spatial segmentation and feature calculations.

[0045] Please see Figure 4 As shown, it is a flowchart of the process of dividing boundary segments based on the length and curvature changes of the fusion boundary line of each frame image in an embodiment of the present invention.

[0046] Specifically, in step S6, the process of determining the number of boundary segments based on the length and curvature changes of the fusion boundary lines of each frame image, and dividing the image into several boundary segments, includes: Step S61: Calculate the total length of the fusion boundary line to determine the number of basic segments based on a preset length interval; Step S62: Calculate the curvature of each point on the fusion boundary line, and determine the number of modified segments based on the number of curvature extreme points and the number of basic segments, so as to divide the boundary into several segments.

[0047] In this embodiment, the total length of the ordered point sequence of the fusion boundary line in the current frame is calculated. A preset length interval related to the actual size of the face is set, for example, corresponding to the actual length of the face of 1.5 cm, which is converted to pixel value according to the image resolution, such as 20 pixels. The basic number of segments = rounded (total length of fusion boundary line / preset length interval). The curvature value of each point on the boundary line is calculated by fitting a local curve and taking the derivative through a sliding window with a window size of 7 points. All local curvature maxima are found, that is, the curvature is greater than its immediate neighbors and exceeds the overall average curvature by 1.5 times, and their number is counted. The final corrected number of segments is determined by the formula Corrected number of segments = Basic number of segments + rounded (number of curvature maxima / 2), that is, on the basis of equal division, segments are added near feature points with significant curvature changes. In step S62, after determining the corrected number of segments, non-uniform segmentation is performed through the following steps: the start and end points of the boundary line are used as initial segmentation points; all curvature maxima are traversed, and a segmentation point is inserted at the position of each curvature maxima. This ensures that key points with sharp contour curvature themselves become the boundaries of a segment. The length of each segment into which the boundary line is divided is calculated using existing points. If the length of any segment is greater than (total length / corrected number of segments) * 1.5, i.e., significantly greater than the average length, a new segment point is inserted in the middle of that long segment until all segment lengths meet the balance requirement or the total number of segments reaches the corrected number of segments. Finally, all inserted points, along with the start and end points, sequentially divide the fused boundary line into several boundary segments of unequal length. This method prioritizes segmentation at feature points and ensures that the overall segments are not too long, thereby achieving fine focusing on key areas.

[0048] Specifically, in step S7, the process of obtaining corresponding fusion boundary coordination parameters based on several boundary segments to determine several boundary segments of interest includes: Step S71: Perform a weighted average based on the fusion boundary coordination parameters corresponding to each boundary segment to determine the fusion boundary coordination degree corresponding to each boundary segment. Step S72: If the fusion boundary coordination degree is less than the preset fusion boundary coordination degree threshold, then the corresponding boundary segment is marked as a boundary segment of interest.

[0049] In this embodiment, for each boundary segment obtained by division, a narrow band region with a bandwidth of 5 pixels is selected on both sides. The four fusion boundary coordination parameters for this segment are calculated, including: Color transition smoothness: Extract the a and b chromaticity components of all pixels in the LAB color space of both sides, calculate the Euclidean distance between the mean chromaticity vectors of both sides, called the chromaticity distance, and smoothness = exp[-(chromaticity distance)² / (adjustment parameter)²], where the adjustment parameter can be set to 5.

[0050] Texture continuity index: Extract the local binary pattern (LBP) feature histograms of the two regions, calculate the Bach distance between the two histograms, and the continuity index = 1 - Bach distance.

[0051] Noise distribution consistency: Perform Daubechies wavelet one-level decomposition on the images of both sides, extract the high-frequency sub-band coefficients, calculate the variance of the amplitude of the high-frequency coefficients on both sides, and calculate the correlation coefficient between the two variances as a consistency index.

[0052] Gradient direction matching degree: Calculate the Sobel gradients of the two regions in the x and y directions, count the gradient direction histogram, divide it into 8 directional intervals, and calculate the cosine similarity between the two histograms as the matching degree.

[0053] Empirical weights were assigned to these four parameters as follows: 0.3, 0.3, 0.2, and 0.2. A weighted average was then used to obtain the fusion boundary coordination degree for that segment. The calculation formula is: Fusion Boundary Coordination Degree = 0.3 * Color Transition Smoothness + 0.3 * Texture Continuity Index + 0.2 * Noise Distribution Consistency + 0.2 * Gradient Direction Matching Degree. Coordination degree is a value between 0 and 1; the closer to 1, the more coordinated and natural the two sides of the segment. A preset fusion boundary coordination degree threshold was set. This threshold was obtained by statistically analyzing the coordination degree distribution of a large amount of known real face data and taking its 2nd quantile. All boundary segments were iterated over. If the coordination degree of a segment was less than the threshold, the segment was considered abnormal and marked as a boundary segment of interest. All marked segments formed a set for further filtering.

[0054] Specifically, in step S8, the process of determining several high-risk boundary segments based on the location and proportion of the several boundary segments of concern includes: Step S81: Determine several high-risk boundary segments based on several boundary segments of interest corresponding to the first preset location conditions; Step S82, and determine several high-risk boundary segments based on the first preset proportion condition; The first preset location condition is that the boundary segment of concern is located in a potentially high-risk area, and the potentially high-risk area is determined based on the existence of a preset number of consecutive boundary segments of concern. The first preset proportion condition is that the proportion of the length of each segment of interest to the length of the fusion boundary line is greater than a preset threshold.

[0055] In this embodiment, step S8 describes two parallel rules for selecting higher-risk boundary segments from the boundary segments of interest. The first rule is based on location, i.e., the first preset location condition: check the index position of all boundary segments of interest on the boundary line. If a preset number of consecutive boundary segments of interest are found to be adjacent, such as 3 or more, these consecutive segments and the continuous boundary areas they cover are considered to constitute a potential high-risk area. Each boundary segment contained within this potential high-risk area, regardless of whether it was previously marked as a boundary segment of interest separately, is identified as a high-risk boundary segment. The second rule is based on proportion, i.e., the first preset proportion condition: calculate the ratio of the sum of the total lengths of all boundary segments of interest to the total length of the entire fusion boundary line. Set a proportion threshold, such as 30%. If the calculated proportion is greater than the proportion threshold, it indicates that the abnormal distribution is widespread, and all boundary segments of interest will be regarded as high-risk boundary segments. If the proportion is less than or equal to the proportion threshold and there is no region with M consecutive segments, then only the Top-N boundary segments of interest with the lowest coordination degree are selected as high-risk boundary segments, such as setting M=3 and N=2. By combining these two rules, we can ensure that we can capture obvious signs of forgery in local clusters, while also preventing the detection of scattered but significant anomalies.

[0056] In this embodiment, in step S9, for each boundary segment identified as high-risk, four types of temporal lighting and shadow feature parameters need to be calculated within its spatial neighborhood, for example, a strip-shaped region extending 5-10 pixels to both sides of the segment as the center line. These parameters are independent instantaneous values ​​for each frame and are used to subsequently construct the temporal sequence, including: Spectrum intensity value: Convert the image to the HSV color space and count the proportion of pixels with a saturation of less than 0.2 and a brightness of greater than 0.9 in this area out of the total number of pixels. This is used as the spectrum intensity value for this segment of the frame.

[0057] Shadow direction value: Calculate the gradient field of the region on the grayscale image, and determine the principal direction angle of the gradient direction through principal component analysis. The range is 0-180 degrees, which is used as the shadow direction value of this segment in the frame.

[0058] Reflection equalization value: Calculate the average pixel value of the R, G, and B channels in this area, and calculate the standard deviation between the three. Reflection equalization value = 1 / (1 + standard deviation of the three channels). The closer the value is to 1, the more balanced the three channels are.

[0059] Light and shadow gradient value: Extract a brightness profile curve that passes through the region along the direction perpendicular to the boundary segment, calculate the mean of the absolute value of the first derivative of the curve, and use it as the light and shadow gradient value of the segment in the frame. The larger the value, the more abrupt the brightness change may be.

[0060] These four parameters characterize the instantaneous physical manifestation of light and shadow in a single frame image from different dimensions, and their temporal sequence will be used for subsequent analysis.

[0061] Specifically, in step S10, the process of determining the boundary temporal sequence and the light and shadow temporal sequence based on the fusion boundary coordination parameters and lighting performance parameters of several high-risk boundary segments, in order to calculate and verify the consistency characterization parameters, includes: Step S101: Arrange the fusion boundary coordination degree corresponding to the fusion boundary coordination parameters of each high-risk boundary segment into a boundary time sequence; Step S102: Perform a weighted average based on the light and shadow feature parameters of each high-risk boundary segment to determine the light and shadow feature representation value corresponding to each high-risk boundary segment, and arrange the light and shadow feature representation values ​​into a light and shadow time sequence. Step S103: Calculate the lag degree based on the boundary time sequence and light and shadow time sequence of the several high-risk boundary segments; Step S104: Based on the boundary time sequence and the light and shadow time sequence, determine the coordination change characteristics of the fusion boundary and the change characteristics of the light and shadow performance, respectively, so as to calculate the consistency of the direction change.

[0062] The characteristics of coordinated change at the fusion boundary include mean drift amplitude, fluctuation frequency, mutation density, and trend correlation; the characteristics of change in light and shadow performance include instantaneous jump intensity, inertial intensity, self-similar decay rate, and periodic significance.

[0063] In this embodiment, for each determined high-risk boundary segment, throughout the entire temporal sequence of the action (i.e., the total number of frames), it has a corresponding fusion boundary coordination degree and a vector composed of four light and shadow feature parameters calculated in step S9 in each frame. To obtain a comprehensive light and shadow representation scalar for sequence construction, the four light and shadow parameters are weighted and averaged in each frame to obtain the light and shadow feature representation value for that frame. After normalizing the four light and shadow feature parameters to unify the dimensions, the calculation formula for the light and shadow feature representation value is: Light and shadow feature representation value = 0.3 * Specular intensity value + 0.3 * Shadow direction value + 0.2 * Reflection equalization value + 0.2 * Light and shadow gradient value. Arranging the fusion boundary coordination degrees in chronological order yields the boundary temporal sequence of the high-risk segment. Similarly, arranging the light and shadow feature representation values ​​in chronological order yields the light and shadow temporal sequence. Feature extraction is performed on both original sequences to calculate the change features. The fusion boundary coordination change features are calculated from the boundary temporal sequence, including: Mean shift magnitude: Calculate the absolute value of the difference between the mean of the last 1 / 3 frames and the mean of the first 1 / 3 frames in the sequence.

[0064] Fluctuation frequency: The sequence is detrended and then subjected to a fast Fourier transform to obtain the frequency corresponding to the largest amplitude value in the amplitude spectrum.

[0065] Mutation density: Calculate the difference sequence of the sequence, count the number of data points whose absolute difference exceeds twice the standard deviation of the difference sequence, and divide by the total number of frames.

[0066] Trend correlation: Calculate the Spearman rank correlation coefficient between the high-risk sequence and all other high-risk sequences, and then calculate the average of these correlation coefficients.

[0067] Calculate the characteristics of light and shadow changes from the light and shadow time sequence, including: Instantaneous jump strength: Calculate the difference sequence of the sequence and take the 95th percentile of the absolute value of the difference sequence.

[0068] Inertia strength: Calculate the autocorrelation coefficient of the first-order difference sequence of the light and shadow time sequence, with a lag of 1 frame. The absolute value of this autocorrelation coefficient is the inertia strength.

[0069] Self-similar decay rate: Calculate the autocorrelation function of the light and shadow temporal sequence, starting from time shift 0 frames and extending to a preset maximum time shift, such as 1 / 4 of the total number of frames. Analyze the rate at which the autocorrelation function value decays as the time shift increases. Specifically, a linear fitting of the decay curve can be used, with the time shift on the x-axis and the autocorrelation coefficient on the y-axis, and the absolute value of the slope of the fitted line is taken as the self-similar decay rate.

[0070] Periodic significance: The light and shadow time series is detrended and then subjected to a Fast Fourier Transform (FFT). In the resulting spectrum, the frequency component with the largest amplitude is found, the DC component is excluded, and its power is calculated as the square of its amplitude. Periodic significance is defined as the power of the largest frequency component divided by the total power of the spectrum.

[0071] These eight higher-order change features are direct inputs for calculating the consistency of trend changes.

[0072] Specifically, in step S103, the process of calculating the lag based on the boundary time series and light and shadow time series of the aforementioned high-risk boundary segments includes: Step S1031: Based on the boundary time sequence and light and shadow time sequence of each high-risk boundary segment, calculate the time shift between the two sequences, and normalize the time shift to determine the alignment time shift amount of the corresponding high-risk boundary segment. Step S1032: Determine the hysteresis based on the average alignment time shift of all high-risk boundary segments.

[0073] Specifically, in step S104, the process of calculating the consistency of the direction change based on the fusion boundary coordination change characteristics and the light and shadow performance change characteristics includes: Step S1041: Construct corresponding feature vectors based on the fusion boundary coordination change characteristics and light and shadow change characteristics of each high-risk boundary segment; Step S1042: Calculate the similarity based on the feature vector, and determine the consistency of the trend change based on the average similarity of all high-risk boundary segments.

[0074] In this embodiment, for the hysteresis, for each high-risk boundary segment, its boundary time series and light and shadow time series are taken. The normalized cross-correlation function of the two sequences is calculated: for time shift k, the normalized cross-correlation coefficient is... k =Σ[(Boundary Coordination Degree) t -Mean boundary compatibility)*(Light and shadow characterization value) t+k -Mean of light and shadow representation value) / (Standard deviation of boundary coordination * Standard deviation of light and shadow representation value), where the summation range is determined by the overlapping part. t can be any time point in the sequence. Traversing a reasonable time shift range, such as from the negative maximum time shift to the positive maximum time shift, find the time shift that maximizes the normalized cross-correlation coefficient (k), called the optimal time shift. This optimal time shift is the offset frame number when the two sequences are most aligned, and its absolute value reflects the lag of the light and shadow changes relative to the boundary coordination changes. Divide it by the total number of frames for normalization to obtain the alignment time shift amount for that segment: Alignment time shift amount = |Optimal time shift| / Total number of frames. It can be understood that the unit of optimal time shift is a frame, and the unit of the total number of frames is a frame, so the alignment time shift amount is dimensionless. Calculate the arithmetic mean of the alignment time shift amounts for all high-risk boundary segments to obtain the overall lag.

[0075] For the consistency of directional changes, the coordinated change features of the fusion boundary of each high-risk segment are concatenated into a vector, called the fusion boundary coordinated change feature vector, and the changes in light and shadow performance are concatenated into another vector, called the light and shadow performance change feature vector. Before calculation, the fusion boundary coordinated change feature vector and the light and shadow performance change feature vector of all high-risk segments are Z-score standardized, i.e., the mean is subtracted and the standard deviation is divided to eliminate dimensions. The cosine similarity between the two standardized vectors of the high-risk segment is calculated: Similarity = (fusion boundary coordinated change feature vector · light and shadow performance change feature vector) / (modulus of fusion boundary coordinated change feature vector * modulus of light and shadow performance change feature vector). The closer the similarity is to 1, the more coordinated the static attribute change pattern and the dynamic light and shadow change pattern of the segment are. Finally, the arithmetic mean of the similarities of all high-risk boundary segments is calculated to obtain the overall consistency of directional changes.

[0076] Specifically, in step S11, the process of determining the risk value of a fake face for the target user based on the weighted fusion of the verification consistency representation parameters and the proportion of high-risk boundary segments includes: Step S111: Calculate the comprehensive score of change consistency based on the weighted fusion of the verification consistency characterization parameters; Step S112: Based on the comprehensive score of change consistency and the proportion of high-risk boundary segments, a weighted fusion is performed to determine the risk value of the fake face of the target user.

[0077] In this embodiment, two change consistency representation parameters are weighted and fused to obtain a comprehensive change consistency score. The formula is: Comprehensive Change Consistency Score = Consistency Weight * Trend Change Consistency + Lag Weight * (1 - Lag). The consistency weight and lag weight are corresponding weighting factors, and their sum is 1. For example, consistency weight = 0.7, lag weight = 0.3. The higher the consistency and the lower the lag, the closer the comprehensive score is to 1, indicating a more realistic result. The proportion of high-risk boundary segments is calculated, which is the ratio of the total length of all high-risk segments to the total length of the fused boundary line.

[0078] The final risk value for forged faces is calculated using the following formula: Forged Face Risk Value = Consistency Contribution Weight * (1 - Overall Consistency Score for Changes) + (1 - Consistency Contribution Weight) * Proportion of High-Risk Boundary Segments. Here, the consistency contribution weight is a balancing parameter, for example, 0.6; (1 - Overall Consistency Score for Changes) converts the consistency score into a risk contribution, with a higher value indicating a higher risk. This formula organically combines temporal inconsistency evidence with spatially anomalous distribution evidence, jointly determining the final risk assessment.

[0079] In this embodiment, in step S12, a predefined ambiguous risk value range for fake faces is set, for example, the risk value is between 0.4 and 0.6. When the risk value for fake faces is less than 0.4, the risk of fake faces in the currently read image is determined to be low; when the risk value for fake faces is greater than 0.6, the risk of fake faces in the currently read image is determined to be high; when the risk value calculated for the first time falls within the range of 0.4 to 0.6, it is impossible to make a clear judgment on whether it is real or fake, and the correction process is triggered.

[0080] The calibration process includes: issuing new instructions based on the initial risk value and the initial light source angle: calibrating the light source requirements, such as prompting "please slightly raise the light source angle" or "please add a weak light on the other side" and updating the direction of action, such as turning right this time if turning left the first time.

[0081] In a specific embodiment, after the correction process is triggered, the side lighting is strengthened based on the facial brightness difference being less than 0.25; the light source angle is slightly raised based on the correction segment number being less than (base segment number + 2), preferably raised by 15° to 25°; if none of the above conditions are met, the action direction is updated first.

[0082] Following the new instructions, the user re-executes the action, thus repeating the core process after step S2: from acquiring a qualified image to calculating the risk value, resulting in a new risk value. The system can compare the two risk values, or take the maximum or average of the two as the basis for the final judgment. This iterative mechanism, by actively changing the detection conditions (lighting and motion direction) when the initial detection is uncertain, stimulates potentially hidden forgery traces from another dimension, thereby improving the decision confidence in edge cases.

[0083] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A deepfake face detection method based on fusion boundary features, characterized in that, The method comprises the following steps: specifying light source requirements according to skin color contrast in an initial image of a target user to be identified; acquiring an image of the target user to be identified performing the specified light source requirements, and determining whether the specified light source requirements are qualified based on a face brightness difference; in the case of determining that the specified light source requirements are qualified, determining a motion direction based on the face brightness difference, so that the user completes a specified motion in the motion direction; acquiring frame-by-frame images of the target user to be identified performing the specified motion according to a time sequence, so as to identify a hair and face junction area of each frame image; extracting a fusion boundary line based on the hair and face junction area; determining the number of divided boundary segments based on the length and curvature change of the fusion boundary line of each frame image, so as to divide the fusion boundary line into a plurality of boundary segments; acquiring corresponding fusion boundary coordination parameters based on the plurality of boundary segments, so as to determine a plurality of concerned boundary segments, wherein the fusion boundary coordination parameters comprise color transition smoothness, texture continuity index, noise distribution consistency and gradient direction matching degree; determining a plurality of high-risk boundary segments based on the positions and proportions of the plurality of concerned boundary segments; acquiring light and shadow feature parameters based on the plurality of high-risk boundary segments, wherein the light and shadow feature parameters comprise highlight intensity value, shadow direction value, reflection balance value and light and shadow gradient value; determining boundary time sequence and light and shadow time sequence based on the fusion boundary coordination parameters and light and shadow performance parameters of the plurality of high-risk boundary segments, so as to calculate verification consistency representation parameters, wherein the verification consistency representation parameters comprise trend change consistency degree and lag degree; determining a fake face risk value of the target user to be identified based on the weighted fusion of the verification consistency representation parameters and the proportion of the high-risk boundary segments.

2. The method of claim 1, wherein the method further comprises: The process of specifying light source requirements according to skin color contrast in an initial image of a target user to be identified comprises the following steps: extracting a face area of the target user to be identified in the initial image, and calculating skin color average brightness value and background average brightness value respectively, so as to determine skin color contrast; determining light source requirements based on the skin color contrast, wherein the light source requirements comprise light source type and light angle, and the light source type comprises point light source and diffuse reflection light.

3. The method of claim 2, wherein the method further comprises: The process of acquiring an image of the target user to be identified performing the specified light source requirements, and determining whether the specified light source requirements are qualified based on a face brightness difference comprises the following steps: acquiring an image of the target user to be identified performing the specified light source requirements, extracting a face area and dividing left and right half faces; calculating average brightness values of the left and right half faces to determine a face brightness difference; comparing the face brightness difference with a preset face brightness difference threshold, so as to determine whether the specified light source requirements are qualified.

4. The method of claim 3, wherein the method further comprises: The process of determining a motion direction of the specified motion based on the face brightness difference in the case of determining that the specified light source requirements are qualified comprises the following steps: determining a darker side in a specified light source requirement qualified image based on the face brightness difference, so as to determine a motion direction of the specified motion.

5. The method of claim 4, wherein the method further comprises: The process of determining the number of divided boundary segments based on the length and curvature change of the fusion boundary line of each frame image, so as to divide the fusion boundary line into a plurality of boundary segments comprises the following steps: calculating the total length of the fusion boundary line, so as to determine the number of basic segments based on a preset length interval; The curvature of each point of the fusion boundary line is calculated, the number of extreme points of the curvature and the number of base segments are determined to determine the number of modified segments for division into a plurality of boundary segments.

6. The method of claim 5, wherein the method further comprises: The process of obtaining corresponding fusion boundary coordination parameters based on a plurality of boundary segments to determine a plurality of concerned boundary segments includes: Weighted average is performed on the fusion boundary coordination parameters corresponding to each boundary segment to determine the fusion boundary coordination degree corresponding to each boundary segment. If the fusion boundary coordination degree is less than a preset fusion boundary coordination degree threshold, the corresponding boundary segment is marked as a concerned boundary segment.

7. The method of claim 6, wherein the method further comprises: The process of determining a plurality of high-risk boundary segments based on the positions and proportions of the plurality of concerned boundary segments includes: A plurality of high-risk boundary segments are determined based on a plurality of concerned boundary segments corresponding to a first preset position condition; and A plurality of high-risk boundary segments are determined based on a first preset proportion condition. The first preset position condition is that the concerned boundary segment is in a potential high-risk area, and the potential high-risk area is determined according to the existence of a continuous preset number of concerned boundary segments. The first preset proportion condition is that the proportion of the length of each concerned segment to the length of the fusion boundary line is greater than a preset threshold.

8. The method of claim 7, wherein the method further comprises: The process of determining boundary timing sequences and light and shadow timing sequences based on the fusion boundary coordination parameters and light and shadow performance parameters of a plurality of high-risk boundary segments to calculate a verification consistency representation parameter includes: The fusion boundary coordination degrees corresponding to the fusion boundary coordination parameters of each high-risk boundary segment are arranged into a boundary timing sequence. Weighted average is performed on the light and shadow feature parameters of each high-risk boundary segment to determine a light and shadow feature representation value corresponding to each high-risk boundary segment, and the light and shadow timing sequence is arranged based on the light and shadow feature representation value. The lag degree is calculated based on the boundary timing sequence and the light and shadow timing sequence of the plurality of high-risk boundary segments. The fusion boundary coordination change feature and the light and shadow performance change feature are determined based on the boundary timing sequence and the light and shadow timing sequence, respectively, to calculate the change consistency degree of the trend. The fusion boundary coordination change feature includes mean shift amplitude, fluctuation frequency, mutation density, and trend correlation; and the light and shadow performance change feature includes instantaneous jump intensity, inertia intensity, self-similarity decay rate, and period significance.

9. The method of claim 8, wherein the method further comprises: The process of calculating the lag degree based on the boundary timing sequence and the light and shadow timing sequence of the plurality of high-risk boundary segments and calculating the change consistency degree of the trend based on the fusion boundary coordination change feature and the light and shadow performance change feature includes: The time shift between the boundary timing sequence and the light and shadow timing sequence of each high-risk boundary segment is calculated, and the time shift is normalized to determine the alignment time shift amount of the corresponding high-risk boundary segment. The lag degree is determined based on the mean value of the alignment time shift amounts of all high-risk boundary segments. The feature vectors corresponding to the fusion boundary coordination change feature and the light and shadow change feature of each high-risk boundary segment are constructed. The similarity is calculated based on the feature vectors, and the change consistency degree of the trend is determined based on the mean value of the similarities of all high-risk boundary segments.

10. The method of claim 1, wherein the method further comprises: The process of determining the risk value of the fake face of the target identified user based on the weighted fusion of the verification consistency representation parameter and the proportion of the high-risk boundary segment includes: The change consistency comprehensive score is calculated based on the weighted fusion of the verification consistency representation parameter. The change consistency comprehensive score and the proportion of high-risk boundary segments are weighted and fused to determine a fake face risk value of the target identification user.

Images