Deepfake anti-counterfeiting verification method based on face micro-motion active excitation

Abstract

The invention relates to a Deepfake anti-counterfeiting verification method based on face micro-motion active excitation, and the method comprises the steps: constructing a dynamic adaptive multi-mode active excitation signal library, and capturing face dynamic response data in combination with face feature point detection and superficial layer electromyographic signals of a verified object; performing multi-dimensional preprocessing on the response data, extracting a micro-motion track and skin deformation conduction characteristics, acquiring facial physiological signal disturbance characteristics synchronous with micro-motion, and ensuring characteristic extraction accuracy through a dynamic region-of-interest tracking mechanism; constructing a micro-motion-physiological signal correlation analysis model, calculating a time sequence correlation degree, quantifying a parameter internal dependency relationship through feature mutual information, and combining with a preset real face micro-motion feature library to complete similarity comparison; and performing hierarchical Deepfake anti-counterfeiting judgment, performing screening based on feature association strength, combining time sequence association degree and feature matching degree verification, and starting abnormal feature traceability to generate a counterfeiting type and a visualization result.

Description

Technical Field

[0001] This invention belongs to the technical field of computer vision, artificial intelligence-generated content anti-counterfeiting and biometric recognition, specifically involving a Deepfake anti-counterfeiting verification method based on active stimulation of facial micro-motions. Background Technology

[0002] With the rapid development of deep learning technology, AI-powered facial recognition technology, represented by Deepfake, has become increasingly sophisticated. Through algorithms such as Generative Adversarial Networks (GANs) and Variational Autoencoders (VAEs), it can fuse real facial images or videos with false identity information to generate visually highly realistic forgeries. This technology has been widely applied in film and television production, virtual social networking, and other fields, but it also brings serious security risks. In digital identity verification scenarios, forged facial videos can bypass traditional facial recognition systems, leading to risks such as account theft and financial fraud. In the field of information dissemination, Deepfake-generated fake news and malicious statements can easily trigger a crisis of social trust and even disrupt public order. Therefore, the demand for efficient and reliable Deepfake anti-counterfeiting verification technology is becoming increasingly urgent.

[0003] Current mainstream Deepfake detection technologies are mainly divided into two categories: passive detection and active detection. However, both have significant technical shortcomings and are difficult to meet the accuracy and generalization requirements in practical applications. Passive detection techniques dominate current research, relying on inherent flaws in forged content for identification. For example, some methods analyze visual defects such as local texture consistency, edge artifacts, and lighting / shadow mismatches—early GAN-based forgery algorithms generated facial details (like hair strands and pores) with blurred or repetitive textures, which these methods can capture through texture analysis models. Other methods focus on the passive acquisition of physiological signals, such as retrieving physiological parameters like heart rate and blood flow from subtle changes in facial skin color, attempting to detect forged content by exploiting its inability to simulate real physiological rhythms. However, with the iteration of Deepfake technology, high-fidelity forgery algorithms can effectively eliminate image artifacts, optimize texture details, and even simulate basic physiological signal fluctuations based on publicly available physiological data. This has led to a sharp decline in the generalization ability of passive detection methods relying on specific flaw features—detection models designed for a particular type of forgery algorithm experience a significant drop in accuracy when facing new forgery techniques, making them difficult to adapt to dynamically changing forgery methods.

[0004] While existing active detection technologies attempt to improve detection robustness by "actively triggering responses specific to real faces," their design has significant limitations. Current active detection schemes often employ only a single stimulus signal (such as a fixed blink command or mouth opening action) and lack a coordinated analysis mechanism linking the stimulus signal, micro-movements, and physiological signals. On one hand, a single stimulus signal is easily simulated in advance by forgery systems—forgers can train a forgery model that responds to fixed commands by collecting samples of real faces performing similar actions. On the other hand, these methods only focus on the completion of macroscopic facial movements, neglecting key features of real faces performing micro-movements, such as the delayed deformation of soft tissue due to inertia during head micro-movements and the dynamic correlation of physiological signals with the intensity of micro-movements. This means that even if the forged content completes the macroscopic movements, it is impossible to distinguish between genuine and fake content based on subtle feature differences.

[0005] The core pain points of existing deepfake anti-counterfeiting verification technologies are: passive detection relies on the inherent flaws of forgery algorithms, resulting in weak generalization and susceptibility to new forgeries; active detection has a simplistic stimulus design, lacks in-depth analysis of the linkage between micro-motion and physiological signals, and has insufficient anti-interference capabilities. Against this backdrop, there is a need for a deepfake anti-counterfeiting verification method that can trigger the unique micro-motion-physiological correlation features of real faces through dynamically adapted active stimuli and achieve accurate source tracing, in order to overcome the shortcomings of existing technologies and ensure digital identity security and the credibility of online content. Summary of the Invention

[0006] To address the aforementioned problems in the existing technology, this invention provides a Deepfake anti-counterfeiting verification method based on active excitation of facial micro-motions; The objective of this invention can be achieved through the following technical solutions: A Deepfake anti-counterfeiting verification method based on active excitation of facial micro-motions, characterized by comprising: S1: Construct a dynamic adaptive multimodal active excitation signal library. Based on the facial detection results of the verified object, select excitation signals from the multimodal active excitation signal library for output, and simultaneously start the micro-motion acquisition process to capture the dynamic response data of the face. S2: Perform a multi-dimensional preprocessing process on the facial dynamic response data. For continuous facial image sequences, use a dynamic region of interest tracking mechanism to locate the target muscle group corresponding to the excitation signal and extract the micro-motion trajectory. For facial physiological signals, use time-adaptive physiological signal time-frequency analysis technology to extract physiological signal perturbation features that are synchronized with the time period of micro-motion. S3: Construct a micro-motion-physiological signal correlation analysis model. Based on the micro-motion trajectory and the physiological fluctuation characteristics, calculate the temporal correlation degree through dynamic time warping; calculate the intrinsic dependency relationship between the parameters of the micro-motion trajectory and the physiological signal perturbation characteristics through the feature mutual information calculation process to generate feature correlation strength parameters; compare the similarity with the standard features under the corresponding excitation signal in the preset real face micro-motion feature library to generate feature matching degree parameters. S4: Based on the results of multi-dimensional correlation analysis, perform hierarchical Deepfake anti-counterfeiting judgment, screen through the feature correlation strength parameter, verify when the correlation strength reaches the preset basic threshold, combine the temporal correlation degree and the feature matching degree parameter, start the abnormal feature tracing process, analyze the breakpoint of the micro-motion trajectory and the abnormal fluctuation pattern of the physiological signal disturbance feature, generate the forgery feature type, and output the anti-counterfeiting verification result and forgery tracing information.

[0007] Specifically, the process of selecting the output of the excitation signal based on the facial detection results of the verified object is as follows: by detecting facial feature points, the initial head posture of the verified object is obtained synchronously, and then the matching non-continuous head micro-movement commands are selected from the dynamic adaptive multimodal active excitation signal library.

[0008] Specifically, when extracting the micro-motion trajectory, skin deformation transmission feature analysis also needs to be performed simultaneously. The specific process is as follows: using the optical flow field and deformation field coupling analysis algorithm, the optical flow field is constructed by the displacement vectors of facial feature points in adjacent frames. Combined with the gradient calculation model of the deformation field, the soft tissue regions of the face cheek, perioral area, and eye area that deform due to inertia are identified. Based on the continuous face image sequence, the delayed deformation time of the soft tissue due to inertia when the head micro-motion occurs is calculated. At the same time, the wave trajectory of the deformation is extracted, including the amplitude of the deformation, the wave frequency, and the time to recover to the initial state. This is then fused with the micro-motion trajectory to generate a linkage trajectory dataset of skeletal movement and soft tissue deformation.

[0009] Specifically, the time-adaptive physiological signal time-frequency analysis technology targets subtle facial color changes caused by cardiovascular impulses. The process includes: capturing continuous facial color signals during and after head micro-movements, and segmenting them into a baseline acquisition phase before the micro-movement, a perturbation acquisition phase during the micro-movement, and a recovery acquisition phase after the micro-movement ends; calculating the degree of disorder in the color signal during the perturbation acquisition phase, which is quantified by the ratio of the signal standard deviation to the baseline acquisition phase standard deviation; simultaneously analyzing the signal return to baseline pattern during the recovery acquisition phase, showing that the real face recovery speed is positively correlated with the intensity of the micro-movement; and incorporating the degree of disorder and recovery pattern parameters as core physiological signal perturbation features into the physiological feature dataset.

[0010] Specifically, the implementation process of the dynamic region of interest tracking mechanism is as follows: In the initial stage, based on the facial feature point detection and anatomical association model, the area where the target muscle group corresponding to the excitation signal is located is marked as the initial region of interest; during the acquisition process, the facial feature point position is re-detected at a preset frame interval, and the region of interest offset is generated by calculating the displacement vector of the feature point; if the offset exceeds the preset range, the range of the initial region of interest is expanded by a preset ratio, and the number of adjustments and the offset magnitude of the region of interest are recorded.

[0011] Specifically, the steps for calculating the temporal correlation degree using dynamic time warping are as follows: aligning the time series of the micro-motion trajectory with the time series of the physiological signal perturbation features by length, and completing the sequence using linear interpolation, the interpolation process being based on the changing trends of adjacent data points; using a dynamic programming algorithm for similarity calculation, quantifying the consistency of the time dimension by finding the optimal matching path between the two types of sequences; dynamically adjusting the size of the time window according to the duration of the micro-motion, calculating the similarity after traversing all time windows, and taking the average of the similarities of all windows as the final temporal correlation degree.

[0012] Specifically, the dynamic update process of the preset real face micro-motion feature library is as follows: real face data is collected periodically to obtain a complete dataset including the micro-motion trajectory, the skin deformation transmission features, and the physiological signal perturbation features; the real face data is filtered based on feature integrity to remove data with missing micro-motion trajectories or broken physiological signals; the cleaned data is standardized and the parameters are mapped to a unified numerical range through a feature normalization algorithm.

[0013] Specifically, the operation of the feature mutual information calculation process includes: extracting two core parameters, the rate of change of velocity and the peak acceleration, from the micro-motion trajectory; extracting two corresponding parameters, the amplitude fluctuation range and the frequency change period, from the physiological signal perturbation features; constructing a probability distribution model of the four types of parameters using a Gaussian distribution fitting algorithm, and optimizing the distribution parameters through maximum likelihood estimation during the fitting process; calculating the mutual information value between the two sets of parameters using the mutual information formula, and taking the mean of the two sets of mutual information values ​​as the feature association strength parameter.

[0014] Specifically, the feature association analysis includes a cross-validation mechanism for multimodal noise suppression: the micro-motion trajectory, the skin deformation conduction features, and the physiological signal perturbation features are divided into three independent modes, and suspicious noise points are marked by modal noise detection algorithms respectively; for the noise interval marked in any mode, the synchronous data of the other two modes are called for cross-validation, and the noise interval confirmed by verification is repaired by the trend extrapolation method of adjacent effective feature segments.

[0015] Specifically, the layered Deepfake anti-counterfeiting judgment uses a real-time processing mechanism. In the feature extraction stage, a layer-adaptive amplitude-based pruning algorithm is used to compress the redundant network layers of optical flow field calculation and physiological signal time-frequency analysis. In the correlation analysis stage, a sliding window block calculation strategy is used to divide long time-series data into overlapping blocks, and dynamic time warping and mutual information calculation are performed in parallel. Finally, a complete correlation result is formed by splicing between blocks.

[0016] Specifically, the preset real face micro-motion feature library periodically collects Deepfake fake samples, extracts the discontinuous patterns of micro-motion trajectories, non-physiological correlation features of skin deformation, and periodic abnormal patterns of physiological signals from the fake samples, and establishes a fake feature sub-library; by comparing with the real feature library, the differential dimensions of the two types of features are mined; the differential dimensions are transformed into new detection indicators and integrated into the micro-motion-physiological signal correlation analysis model.

[0017] Specifically, the detailed implementation of the abnormal feature tracing process is as follows: based on the dynamic response data of the face, a multi-dimensional feature curve synchronized with the time axis is generated, abnormal points are located and abnormal features are analyzed; according to the type and combination pattern of the abnormal features, the forgery type is classified, and a labeled visualization result map is generated.

[0018] The beneficial effects of this invention are as follows: Existing deepfake detection methods largely rely on passively extracting static features such as image texture and lighting anomalies, making them easily evaded by novel generation algorithms (such as high-fidelity forgery techniques based on neural radiation fields). This invention constructs a dynamically adaptable multimodal active excitation signal library to actively trigger the micro-movements and physiological responses of the verified object's face. Real faces, under excitation, exhibit physiologically related features of "micro-movement trajectory - skin deformation - physiological signals," while deepfake content, unable to simulate the biomechanics and physiological mechanisms of real human movement, inevitably shows feature breaks or misalignments. Combining a correlation analysis model of dynamic time warping and feature mutual information calculation, the temporal synchronicity and intrinsic dependencies of the three types of features can be accurately quantified, effectively solving the industry pain point of "false judgments caused by forged features imitating real static features."

[0019] This invention simultaneously extracts three modalities of data during the preprocessing stage: micro-motion trajectory, skin deformation transmission features, and physiological signal perturbation features. Through a multimodal noise suppression and cross-validation mechanism, it effectively filters out interference from harsh environments such as low light, motion blur, and equipment noise. For example, in low-light scenarios, the infrared and visible light fusion acquisition strategy can compensate for signal attenuation from a single channel, and skin deformation features can serve as a supplementary verification dimension for micro-motion trajectories, avoiding missed detections due to distortion of single-modal data. Furthermore, the three types of features correspond to three independent dimensions: "mechanics," "soft tissue physical properties," and "physiological metabolism." Forged content cannot simultaneously simulate the real correlation across multiple dimensions, enabling the method to identify anomalies through feature cross-validation even when faced with "partial tampering" (such as forging only micro-motion trajectories while retaining some real physiological signals). This demonstrates superior robustness compared to single-modal or dual-modal detection schemes.

[0020] Existing detection methods mostly output only a binary "true" or "false" result, failing to pinpoint the location or type of forgery, which is detrimental to subsequent traceability and algorithm optimization. This invention employs a layered Deepfake anti-counterfeiting judgment logic. First, it quickly screens for obvious forgeries based on feature correlation strength, then deeply verifies suspected samples by combining temporal correlation and feature matching. For objects determined to be forgeries, an abnormal feature tracing process is initiated—using multi-dimensional feature curves synchronized with the time axis, the location of micro-motion trajectory breakpoints and abnormal fluctuation ranges of physiological signals are intuitively marked, and forgery types are generated according to categories such as "intermittent trajectory," "temporally misaligned," and "physiologically smooth," outputting labeled and visualized results. This design not only provides users with a clear basis for "why it was determined to be forgery," but also provides direction for optimizing dynamically adaptable excitation signal libraries (e.g., adjusting the priority of corresponding excitation signals when the proportion of a certain type of forgery increases), solving the practical problem of "uninterpretable anti-counterfeiting results that are difficult to guide subsequent optimization." Attached Figure Description

[0021] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings.

[0022] Figure 1 This is a flowchart illustrating a Deepfake anti-counterfeiting verification method based on active excitation of facial micro-motions according to the present invention. Figure 2 This is a timing diagram of the method execution in this invention. Detailed Implementation

[0023] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided.

[0024] Please see Figure 1-2A Deepfake anti-counterfeiting verification method based on active excitation of facial micro-motions, characterized by comprising: S1: Construct a dynamic adaptive multimodal active excitation signal library. Based on the facial detection results of the verified object, select excitation signals from the multimodal active excitation signal library for output, and simultaneously start the micro-motion acquisition process to capture the dynamic response data of the face. S2: Perform a multi-dimensional preprocessing process on the facial dynamic response data. For continuous facial image sequences, use a dynamic region of interest tracking mechanism to locate the target muscle group corresponding to the excitation signal and extract the micro-motion trajectory. For facial physiological signals, use time-adaptive physiological signal time-frequency analysis technology to extract physiological signal perturbation features that are synchronized with the time period of micro-motion. S3: Construct a micro-motion-physiological signal correlation analysis model. Based on the micro-motion trajectory and the physiological fluctuation characteristics, calculate the temporal correlation degree through dynamic time warping; calculate the intrinsic dependency relationship between the parameters of the micro-motion trajectory and the physiological signal perturbation characteristics through the feature mutual information calculation process to generate feature correlation strength parameters; compare the similarity with the standard features under the corresponding excitation signal in the preset real face micro-motion feature library to generate feature matching degree parameters. S4: Based on the results of multi-dimensional correlation analysis, perform hierarchical Deepfake anti-counterfeiting judgment, screen through the feature correlation strength parameter, verify when the correlation strength reaches the preset basic threshold, combine the temporal correlation degree and the feature matching degree parameter, start the abnormal feature tracing process, analyze the breakpoint of the micro-motion trajectory and the abnormal fluctuation pattern of the physiological signal disturbance feature, generate the forgery feature type, and output the anti-counterfeiting verification result and forgery tracing information.

[0025] Specifically, the process of selecting the output of the excitation signal based on the facial detection results of the verified object is as follows: by detecting facial feature points, the initial head posture of the verified object is obtained synchronously, and then the matching non-continuous head micro-movement commands are selected from the dynamic adaptive multimodal active excitation signal library.

[0026] Specifically, when extracting the micro-motion trajectory, skin deformation transmission feature analysis also needs to be performed simultaneously. The specific process is as follows: using the optical flow field and deformation field coupling analysis algorithm, the optical flow field is constructed by the displacement vectors of facial feature points in adjacent frames. Combined with the gradient calculation model of the deformation field, the soft tissue regions of the face cheek, perioral area, and eye area that deform due to inertia are identified. Based on the continuous face image sequence, the delayed deformation time of the soft tissue due to inertia when the head micro-motion occurs is calculated. At the same time, the wave trajectory of the deformation is extracted, including the amplitude of the deformation, the wave frequency, and the time to recover to the initial state. This is then fused with the micro-motion trajectory to generate a linkage trajectory dataset of skeletal movement and soft tissue deformation.

[0027] Specifically, the time-adaptive physiological signal time-frequency analysis technology targets subtle facial color changes caused by cardiovascular impulses. The process includes: capturing continuous facial color signals during and after head micro-movements, and segmenting them into a baseline acquisition phase before the micro-movement, a perturbation acquisition phase during the micro-movement, and a recovery acquisition phase after the micro-movement ends; calculating the degree of disorder in the color signal during the perturbation acquisition phase, which is quantified by the ratio of the signal standard deviation to the baseline acquisition phase standard deviation; simultaneously analyzing the signal return to baseline pattern during the recovery acquisition phase, showing that the real face recovery speed is positively correlated with the intensity of the micro-movement; and incorporating the degree of disorder and recovery pattern parameters as core physiological signal perturbation features into the physiological feature dataset.

[0028] In this embodiment, the green channel of the image acquisition component is used. Because hemoglobin has a higher absorption coefficient for green light than other visible light bands, this channel can more accurately capture subtle changes in facial color caused by blood flow. Continuous facial color signals are captured during and after head micro-movements and segmented into "baseline acquisition stage before micro-movement → perturbation acquisition stage during micro-movement → recovery acquisition stage after micro-movement ends." The instantaneous disorder of the color signal during the perturbation acquisition stage is calculated—quantified by the ratio of the standard deviation of the signal in this stage to the standard deviation of the baseline acquisition stage. The ratio of the actual face should be within the normal fluctuation range of the cardiovascular system after being stimulated by movement. At the same time, the pattern of signal return to the baseline during the recovery acquisition stage is analyzed. The actual face usually shows a gradual recovery trend, and the recovery speed is positively correlated with the intensity of the micro-movement (the greater the intensity of the movement, the slower the recovery speed). The disorder level and recovery pattern parameters are used as core physiological signal perturbation features and incorporated into the physiological feature dataset. The two types of parameters must meet the synchronicity in the time dimension (the time difference between the disorder peak and the micro-movement peak is within a preset physiologically reasonable range).

[0029] Specifically, the implementation process of the dynamic region of interest tracking mechanism is as follows: In the initial stage, based on the facial feature point detection and anatomical association model, the area where the target muscle group corresponding to the excitation signal is located is marked as the initial region of interest; during the acquisition process, the facial feature point position is re-detected at a preset frame interval, and the region of interest offset is generated by calculating the displacement vector of the feature point; if the offset exceeds the preset range, the range of the initial region of interest is expanded by a preset ratio, and the number of adjustments and the offset magnitude of the region of interest are recorded.

[0030] In this embodiment, in the initial stage, based on the facial feature point detection and anatomical association model, the region where the target muscle group corresponding to the excitation signal is located is marked as the initial region of interest (ROI). For example, when performing left and right micro-movements of the head, the target muscle group is the sternocleidomastoid muscle of the neck, and its corresponding facial association region is the two sides of the mandible. This association is determined by the muscle attachment point and motion transmission path in anatomy. During the acquisition process, the position of facial feature points is re-detected at preset frame intervals, and the displacement vector of the feature points is calculated to determine whether the ROI has shifted due to head movement. If the shift exceeds a preset reasonable range (this range is determined based on the common head movement amplitude and the coverage radius of the ROI), the original region is automatically expanded by a preset ratio to ensure that the shifted target muscle group is still completely within the ROI. At the same time, the number of adjustments and the shift amplitude of the ROI are recorded. These parameters serve as the basis for judging the stability of the subsequent micro-movement trajectory. Too many adjustments or a shift amplitude exceeding the normal range may indicate instability in the micro-movement trajectory, which needs to be focused on in subsequent association analysis.

[0031] Specifically, the steps for calculating the temporal correlation degree using dynamic time warping are as follows: The time series of the micro-motion trajectory is aligned in length with the time series of the physiological signal perturbation features; the sequence is completed using linear interpolation, with the interpolation process based on the changing trends of adjacent data points; similarity calculation employs a dynamic programming algorithm, quantifying the consistency of the time dimension by finding the optimal matching path between the two types of sequences; the size of the time window is dynamically adjusted according to the duration of the micro-motion; after traversing all time windows to calculate the similarity, the average of all window similarities is taken as the final temporal correlation degree. Based on the correlation degree distribution of real face samples, the correlation degree is divided into three levels: high correlation, medium correlation, and low correlation, where high correlation represents a high degree of synchronization between micro-motion and physiological signals in the time dimension.

[0032] Specifically, the dynamic update process of the preset real face micro-motion feature library is as follows: real face data is collected periodically to obtain a complete dataset including the micro-motion trajectory, the skin deformation transmission features, and the physiological signal perturbation features; the real face data is filtered based on feature integrity to remove data with missing micro-motion trajectories or broken physiological signals; the cleaned data is standardized and the parameters are mapped to a unified numerical range through a feature normalization algorithm.

[0033] Specifically, the operation of the feature mutual information calculation process includes: extracting two core parameters, the rate of change of velocity and the peak acceleration, from the micro-motion trajectory; extracting two corresponding parameters, the amplitude fluctuation range and the frequency change period, from the physiological signal perturbation features; constructing a probability distribution model of the four types of parameters using a Gaussian distribution fitting algorithm, and optimizing the distribution parameters through maximum likelihood estimation during the fitting process; calculating the mutual information value between the two sets of parameters using the mutual information formula, and taking the mean of the two sets of mutual information values ​​as the feature association strength parameter.

[0034] In this embodiment, the first step is to extract two types of core parameters from the micro-motion trajectory: the rate of change of velocity (calculated at high frequency time intervals, the ratio of the difference in velocity between adjacent moments to the time interval, reflecting the intensity of velocity change) and the peak acceleration (extracted at a preset time window, reflecting the force intensity of the action); and to extract two corresponding parameters from the physiological signal disturbance characteristics: the range of signal amplitude fluctuation (calculated at high frequency time intervals, the difference between the maximum and minimum values ​​of the signal within the interval, reflecting the intensity of signal fluctuation) and the frequency change period (calculated at the time interval between two adjacent peaks of the signal, reflecting the periodic change pattern of the signal).

[0035] The second step involves constructing a probability distribution model for four types of parameters using a Gaussian distribution fitting algorithm. The fitting process optimizes the distribution parameters through maximum likelihood estimation to ensure that the model accurately reflects the distribution characteristics of the real parameters. The third step involves calculating the mutual information value between the two sets of parameters using the mutual information formula. The mutual information value of "rate of change of velocity - amplitude fluctuation range" reflects the correlation strength between the change of micro-motion speed and the amplitude fluctuation of physiological signals, while the mutual information value of "peak acceleration - frequency change period" reflects the correlation strength between the intensity of micro-motion force and the periodic change of physiological signals. The fourth step involves taking the mean of the two sets of mutual information values ​​as the feature correlation strength parameter. Based on the mutual information distribution of real face samples, the parameter is divided into three levels: strong dependence, moderate dependence, and weak dependence. Strong dependence indicates that there is a significant intrinsic correlation between micro-motion parameters and physiological signal parameters.

[0036] Specifically, the feature association analysis includes a cross-validation mechanism for multimodal noise suppression: the micro-motion trajectory, the skin deformation conduction features, and the physiological signal perturbation features are divided into three independent modes, and suspicious noise points are marked by modal noise detection algorithms respectively; for the noise interval marked in any mode, the synchronous data of the other two modes are called for cross-validation, and the noise interval confirmed by verification is repaired by the trend extrapolation method of adjacent effective feature segments.

[0037] Specifically, the layered Deepfake anti-counterfeiting judgment uses a real-time processing mechanism. In the feature extraction stage, a layer-adaptive amplitude-based pruning algorithm is used to compress the redundant network layers of optical flow field calculation and physiological signal time-frequency analysis. In the correlation analysis stage, a sliding window block calculation strategy is used to divide long time-series data into overlapping blocks, and dynamic time warping and mutual information calculation are performed in parallel. Finally, a complete correlation result is formed by splicing between blocks.

[0038] In this embodiment, the detailed logic of the layered Deepfake anti-counterfeiting judgment is as follows: In the first round of screening, if the feature association strength parameter does not reach the preset strong dependence standard, "preliminary judgment of forgery" is directly output, and the key defects that lead to the judgment are marked. If the skin deformation delay time in the micro-motion trajectory exceeds the reasonable physiological range of the real human face, or the deformation amplitude is much lower than the range of soft tissue movement of the real human face, "abnormal skin deformation feature" is marked. If the standard deviation ratio of the physiological signal disturbance stage does not reach the normal fluctuation range after the cardiovascular system is stimulated, or the recovery stage duration is much longer than the normal recovery range corresponding to the same intensity of micro-motion, "abnormal physiological signal disturbance feature" is marked.

[0039] In the second round of verification, if the feature correlation strength parameter reaches the preset strong or medium dependence standard, a comprehensive judgment is made by combining temporal correlation and feature matching degree: when both reach the preset high correlation and high matching standards, "true judgment" is output; when either parameter does not reach the corresponding standard, the abnormal feature tracing process is initiated; if the temporal correlation does not reach the preset medium-high correlation standard, the focus is on analyzing the time misalignment between micro-movements and physiological signals - for example, the difference between the peak occurrence time of physiological signal disturbance and the peak occurrence time of micro-movement actions exceeds the physiological synchronization range of real faces; if the feature matching degree does not reach the preset high matching standard, the focus is on comparing the difference between skin deformation parameters and the standard values ​​of the feature library - for example, the deviation between the deformation amplitude and the standard value exceeds the reasonable range of individual differences in real faces.

[0040] Specifically, the preset real face micro-motion feature library periodically collects Deepfake fake samples, extracts the discontinuous patterns of micro-motion trajectories, non-physiological correlation features of skin deformation, and periodic abnormal patterns of physiological signals from the fake samples, and establishes a fake feature sub-library; by comparing with the real feature library, the differential dimensions of the two types of features are mined; the differential dimensions are transformed into new detection indicators and integrated into the micro-motion-physiological signal correlation analysis model.

[0041] Specifically, the detailed implementation of the abnormal feature tracing process is as follows: based on the dynamic response data of the face, a multi-dimensional feature curve synchronized with the time axis is generated, abnormal points are located and abnormal features are analyzed; according to the type and combination pattern of the abnormal features, the forgery type is classified, and a labeled visualization result map is generated.

[0042] In this embodiment, the detailed implementation of the abnormal feature tracing process is as follows: First, the stored raw data and preprocessing results of "micro-motion trajectory - skin deformation - physiological signal" are called, and a multi-dimensional feature curve synchronized with the time axis is generated through data visualization technology. The horizontal axis represents the time process, and the vertical axis corresponds to the micro-motion speed, skin deformation amplitude, and physiological signal intensity, respectively. The three curves are completely aligned on the time axis, which facilitates intuitive observation of the feature correlation. Second, abnormal points are located and abnormal features are analyzed: If the micro-motion trajectory is unnaturally discontinuous, the time node corresponding to the breakpoint is marked, and the difference of micro-motion parameters before and after the breakpoint is calculated. If the difference exceeds the normal fluctuation range of real face micro-motion parameters, it is determined to be "trajectory breakpoint abnormal". If the physiological signal fluctuates abnormally, the time interval corresponding to the abnormal fluctuation is marked, and the frequency distribution characteristics of the signal in the interval are analyzed. For example, if the signal frequency remains fixed and there is no dynamic fluctuation that should exist in real physiological signals, it is determined to be "signal frequency abnormal".

[0043] The third step is to classify the forgery type based on the type and combination pattern of abnormal features: if there are many trajectory breakpoints and abnormal skin deformation features, it is judged as "micro-motion trajectory discontinuity forgery"; if the temporal correlation does not meet the standard and there is obvious time misalignment, it is judged as "temporal misalignment forgery"; if the physiological signal fluctuation amplitude is much lower than the normal range and the frequency change period deviation is significant, it is judged as "physiological signal smoothness forgery". The fourth step is to generate a labeled visualization result chart - the abnormal interval is selected on the feature curve with special visual markers, and the specific features of the abnormal parameters are labeled with text (such as "the micro-motion speed drops sharply at a certain time node, and the difference exceeds the normal fluctuation range"). This chart is output along with the anti-counterfeiting verification results, providing an intuitive basis for subsequent forgery tracing.

[0044] The detailed process of the result feedback optimization step is as follows: First, establish a dynamic mapping relationship between "spoofing type - excitation signal optimization": For "micro-motion trajectory discontinuity spoofing", appropriately increase the amplitude of head micro-motion posture changes and extend the duration of micro-motion, thereby increasing the detection sensitivity of trajectory continuity by expanding the difference in action features; For "temporal misalignment spoofing", shorten the time interval between visual cues and voice cues in the excitation signal to ensure that the action execution of the verified object is synchronized with the signal cues, reducing human temporal deviation; For "physiological signal smoothness spoofing", increase the intensity of physiological reflex excitation signals (such as increasing the intensity of gradient light stimulation) to enhance the perturbation degree of physiological signals, making the smoothness of the spoofed signal easier to expose.

[0045] The second step involves accumulating a preset scale of verification results and adjusting the screening priority of the dynamically adaptable multimodal active stimulus signal library based on the mapping relationship. For example, if the proportion of a certain type of forgery increases significantly in a specific scenario, the priority of the corresponding optimized stimulus signal is increased to ensure that signals that are easier to identify forgeries are used in that scenario. The third step involves verifying and evaluating the optimized stimulus signals: a verification set containing real and forged samples is selected, and cross-validation is used to evaluate the verification accuracy of the optimized signals. If the accuracy improvement reaches a preset significant standard, the optimization strategy is fixed. If the accuracy improvement does not reach the preset standard, the adaptation defects between the forgery type and the stimulus signal are re-analyzed, and the parameters in the mapping relationship are adjusted (such as further optimizing the intensity or complexity of the stimulus signal) until the verification effect meets the requirements.

[0046] In one embodiment, the user identity verification process of an online financial platform needs to identify whether the real-time facial video collected from the user is Deepfake content. Core component letter identifier definition Signal library and detection: Dynamically adaptable multimodal active excitation signal library: Library A (including autonomous control signals A1 and physiological reflex signals A2); Face detection module: M1 (used to extract initial head pose and facial feature points); Micro-motion acquisition: C1 (including image acquisition submodule C) 11 Physiological signal acquisition submodule C 12 ).

[0047] Preprocessing algorithms and tracking mechanisms: Dynamic region of interest tracking mechanism: T1; Optical flow field and deformation field coupled analysis algorithm: A3 (used to extract skin deformation conduction characteristics); Time-adaptive physiological signal time-frequency analysis technology: A4 (used to extract physiological signal perturbation features).

[0048] Association analysis and feature library: Micromotor-physiological signal correlation analysis model: M2; Dynamic Time Warping Algorithm: A5 (for calculating temporal correlation); Feature mutual information calculation module: A6 (generates feature association strength parameters); Preset real human face micro-motion feature library: Library B (containing real micro-motion features B1, skin deformation features B2, and physiological signal features B3); Forgery Feature Sub-library: Library B1 (containing forged trajectory discontinuity pattern B) 11 Non-physiological deformation characteristics B 12 Abnormal physiological signals B 13 ).

[0049] Judgment and Source Tracing Module: Layered Deepfake anti-counterfeiting judgment module: J1; Anomaly Feature Tracing Module: T2; Real-time optimization module: O1 (including network pruning submodule O) 11 Block-based computation submodule O 12 ).

[0050] Step 1: Excitation signal selection and data acquisition Start module M1 to perform face detection on the object being verified and output the initial head pose parameters P1; Based on parameter P1, select suitable excitation signals from library A (if P1 is a frontal view, select horizontal micro-motion signal A1; if it is a side view, select vertical micro-motion signal A2). Output the filtered excitation signal and synchronously start the acquisition module C1: Submodule C 11 Capture a continuous sequence of human face images (I); Submodule C 12 Simultaneous acquisition of facial physiological signals S; Module C1 outputs dynamic face response data D (D=I+S).

[0051] Step 2: Multi-dimensional preprocessing Micro-motion trajectory and skin deformation extraction: Input image sequence I to tracking mechanism T1 to lock the target muscle group region R corresponding to the excitation signal; Algorithm A3 is invoked to construct the optical flow field using the displacement vectors of feature points in adjacent frames, and then the deformation field gradient is calculated. Extract the micro-motion trajectory T of region R (including displacement trend T1 and velocity change T2); Calculate the delayed deformation time t of soft tissue, extract the deformation amplitude A, fluctuation frequency F, and recovery time tᵣ, and generate the linkage trajectory dataset D1 (D1=T+t+A+F+tᵣ).

[0052] Physiological signal perturbation feature extraction: Input the blood flow color change signal S2 from the physiological signal S to technique A4, and process it in time segments: Baseline S 21 (Before micro-motion), disturbance segment S 22 (During micro-movement), recovery phase S 23 (After micro-exercise); Calculate the disturbance segment S 22 The degree of signal disturbance C (C = standard deviation of disturbance segment / standard deviation of baseline segment); Analysis and recovery of segment S23 The regression model P (P = the correlation between recovery speed and micro-movement intensity); Generate a physiological signal perturbation feature set D2 (D2=C+P).

[0053] Step 3: Association Analysis and Parameter Calculation Start the association analysis model M2, and input the dataset D1 and feature set D2: Call algorithm A5 to process the trajectory time series T in D1. t With the characteristic time series D in D2 2t Execute length alignment, find the optimal matching path through dynamic programming, and output the temporal correlation degree R. t ; Call module A6 to extract the rate of change of velocity V and peak acceleration A of D1. p Extract the amplitude fluctuation range S of D2 a Given the frequency and period T, calculate the mutual information value I between the two sets of parameters. m The mean value is used to generate the feature association strength parameter Rᵢ.

[0054] Feature similarity comparison: Call library B to compare D1 with features B1 and B2, and compare D2 with feature B3, outputting the feature matching degree parameter R. m ; To update library B, call library B1 and compare D1 with B. 11 D2 and B 13 We explored differentiating dimensions and added new detection indicators to model M2.

[0055] Step 4: Stratification and Source Tracing Start the decision module J1 and input parameters Rᵢ and R. t R m : First round of screening: If Rᵢ < preset threshold Tᵢ, directly output the forgery judgment result J. 11 ; Second round of verification: If Rᵢ≥Tᵢ, combined with R... t ≥T t (Time series threshold), R m ≥T m (Matching threshold), output the true judgment result J 12 If any parameter fails to meet the standard, the tracing module T2 is triggered.

[0056] Anomaly identification and real-time optimization: Module T2 calls data D, D1, and D2 to generate a time-axis synchronization characteristic curve, locates the micro-motion trajectory breakpoint Pᵦ and the abnormal physiological signal interval Sᵦ, and outputs the forgery type T. 21(e.g., discontinuous trajectory, time-series misalignment) and visual annotation diagrams; Startup optimization module O1: Submodule O 11 Redundant network layers in algorithms A3 and A4 are pruned, and submodule O is... 12 Long-term time-series data is divided into blocks for parallel computation to ensure the real-time performance of the decision-making process.

[0057] The final output includes the anti-counterfeiting verification result J (genuine / counterfeit) and the counterfeit feature type T. 21 Visualized traceability diagrams, and key parameters (Rᵢ, Rᵢ, Rᵢ). t R m The quantification value of ) completes a full Deepfake anti-counterfeiting verification.

[0058] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A Deepfake anti-counterfeiting verification method based on active excitation of facial micro-motions, characterized in that, include: S1: Construct a dynamic adaptive multimodal active excitation signal library, and based on the facial detection results of the verified object, select excitation signals from the multimodal active excitation signal library for output, start the micro-motion acquisition process, and capture the dynamic response data of the face; S2: Perform a multi-dimensional preprocessing process on the dynamic response data of the face. For a continuous face image sequence, use a dynamic region of interest tracking mechanism to lock the region where the target muscle group corresponding to the excitation signal is located and extract the micro-motion trajectory. For facial physiological signals, time-frequency analysis technology for time-series adapted physiological signals is used to extract physiological signal perturbation features that are synchronized with the time period of micro-movement. S3: Construct a micro-motor-physiological signal correlation analysis model, and calculate the temporal correlation degree based on the micro-motor trajectory and the physiological fluctuation characteristics through dynamic time warping; The intrinsic dependency between the parameters of the micro-motion trajectory and the physiological signal perturbation features is calculated through a feature mutual information calculation process, and feature correlation strength parameters are generated. The similarity is compared with the standard features under the corresponding excitation signal in the preset real face micro-motion feature library to generate feature matching degree parameters; S4: Based on the results of multi-dimensional correlation analysis, perform hierarchical Deepfake anti-counterfeiting judgment, screen through the feature correlation strength parameter, verify when the correlation strength reaches the preset basic threshold, combine the temporal correlation degree and the feature matching degree parameter, start the abnormal feature tracing process, analyze the breakpoint of the micro-motion trajectory and the abnormal fluctuation pattern of the physiological signal disturbance feature, generate the forgery feature type, and output the anti-counterfeiting verification result and forgery tracing information.

2. The method according to claim 1, characterized in that, In S1, the specific process of filtering the excitation signal output based on the facial detection results of the verified object is as follows: by detecting facial feature points, the initial head posture of the verified object is obtained synchronously, and then the matching discontinuous head micro-movement commands are filtered from the dynamic adaptive multimodal active excitation signal library.

3. The method according to claim 1, characterized in that, In S2, when extracting the micro-motion trajectory, skin deformation transmission feature analysis also needs to be performed simultaneously. The specific process is as follows: using the optical flow field and deformation field coupling analysis algorithm, the optical flow field is constructed by the displacement vectors of facial feature points in adjacent frames. Combined with the gradient calculation model of the deformation field, the soft tissue regions of the face cheek, perioral area, and eye area that deform due to inertia are identified. Based on the continuous face image sequence, the delayed deformation time of the soft tissue due to inertia when the head micro-motion occurs is calculated. At the same time, the wave trajectory of the deformation is extracted, including the amplitude of the deformation, the wave frequency, and the time to recover to the initial state. This is then fused with the micro-motion trajectory to generate a linkage trajectory dataset of skeletal movement and soft tissue deformation.

4. The method according to claim 1, characterized in that, In S2, the time-adaptive physiological signal time-frequency analysis technology specifically targets the subtle facial color change signals caused by cardiovascular pulses. The process includes: capturing continuous facial color signals during and after head micro-movements, and segmenting them into a baseline acquisition phase before the micro-movement, a perturbation acquisition phase during the micro-movement, and a recovery acquisition phase after the micro-movement ends; calculating the degree of disorder in the color signal during the perturbation acquisition phase, which is quantified by the ratio of the signal standard deviation to the baseline acquisition phase standard deviation; simultaneously analyzing the signal return to baseline pattern during the recovery acquisition phase, showing that the real face recovery speed is positively correlated with the intensity of the micro-movement; and incorporating the degree of disorder and recovery pattern parameters as core physiological signal perturbation features into the physiological feature dataset.

5. The method according to claim 1, characterized in that, In S2, the specific implementation process of the dynamic region of interest tracking mechanism is as follows: In the initial stage, based on the facial feature point detection and anatomical association model, the area where the target muscle group corresponding to the excitation signal is located is marked as the initial region of interest; during the acquisition process, the facial feature point position is re-detected at a preset frame interval, and the region of interest offset is generated by calculating the displacement vector of the feature point; if the offset exceeds the preset range, the range of the initial region of interest is expanded by a preset ratio, and the number of adjustments and the offset magnitude of the region of interest are recorded.

6. The method according to claim 1, characterized in that, In S3, the steps for calculating the temporal correlation degree using dynamic time warping are as follows: the time series of the micro-motion trajectory is aligned with the time series of the physiological signal perturbation features by length, and the sequence is completed by linear interpolation. The interpolation process needs to be based on the changing trend of adjacent data points. The similarity calculation adopts a dynamic programming algorithm to quantify the consistency of the time dimension by finding the optimal matching path between the two types of sequences. The size of the time window is dynamically adjusted according to the duration of the micro-motion. After traversing all time windows to calculate the similarity, the average of the similarities of all windows is taken as the final temporal correlation degree.

7. The method according to claim 1, characterized in that, In S3, the dynamic update process of the preset real face micro-motion feature library is as follows: collect real face data periodically and obtain a complete dataset including the micro-motion trajectory, the skin deformation transmission features, and the physiological signal perturbation features. The real face data is filtered based on feature integrity, removing data with missing micro-motion trajectories or broken physiological signals; the cleaned data is then standardized, and the parameters are mapped to a unified numerical range using a feature normalization algorithm.

8. The method according to claim 1, characterized in that, In S3, the operation of the feature mutual information calculation process includes: extracting two core parameters, the rate of change of velocity and the peak acceleration, from the micro-motion trajectory; extracting two corresponding parameters, the amplitude fluctuation range and the frequency change period, from the physiological signal perturbation features; constructing a probability distribution model of the four types of parameters using a Gaussian distribution fitting algorithm, and optimizing the distribution parameters through maximum likelihood estimation during the fitting process; calculating the mutual information value between the two sets of parameters using the mutual information formula, and taking the mean of the two sets of mutual information values ​​as the feature association strength parameter.

9. The method according to claim 1, characterized in that, In S3, the feature association analysis includes a cross-validation mechanism for multimodal noise suppression: the micro-motion trajectory, the skin deformation conduction feature, and the physiological signal perturbation feature are divided into three independent modes, and suspicious noise points are marked by modal noise detection algorithms respectively; for the noise interval marked in any mode, the synchronous data of the other two modes are called for cross-validation, and the noise interval confirmed by verification is repaired by the trend extrapolation method of adjacent effective feature segments.

10. The method according to claim 1, characterized in that, In S4, the layered Deepfake anti-counterfeiting judgment uses a real-time processing mechanism. In the feature extraction stage, a layer-adaptive amplitude-based pruning algorithm is used to compress the redundant network layers of optical flow field calculation and physiological signal time-frequency analysis. In the correlation analysis stage, a sliding window block calculation strategy is used to divide long time series data into overlapping blocks, and dynamic time warping and mutual information calculation are performed in parallel. Finally, a complete correlation result is formed by splicing between blocks.

11. The method according to claim 1, characterized in that, In S4, the preset real face micro-motion feature library periodically collects Deepfake fake samples, extracts the discontinuous patterns of micro-motion trajectories, non-physiological correlation features of skin deformation, and periodic abnormal patterns of physiological signals of the fake samples, and establishes a fake feature sub-library; by comparing with the real feature library, the differential dimensions of the two types of features are mined; the differential dimensions are transformed into new detection indicators and integrated into the micro-motion-physiological signal correlation analysis model.

12. The method according to claim 1, characterized in that, In S4, the detailed implementation of the abnormal feature tracing process is as follows: based on the dynamic response data of the face, a multi-dimensional feature curve synchronized with the time axis is generated, abnormal points are located and abnormal features are analyzed; according to the type and combination pattern of the abnormal features, the forgery type is classified, and a labeled visualization result map is generated.

Images