Encrypted file box capable of being unlocked through remote facial recognition

Abstract

The invention discloses an encrypted file box capable of being unlocked through remote facial recognition, a recognition system is arranged on the encrypted file box, the recognition system sends an instruction to a control unit embedded in the recognition system after recognizing a correct signal, and the control unit controls unlocking of the encrypted file box; wherein the identification system and the mobile client are both connected to the network, and mutual transmission of data is realized through interaction between accounts. The face recognition technology is adopted as an unlocking means, and the security level of the file box is effectively improved; a user can carry out remote face recognition through the mobile client, the file box can be safely unlocked under the condition that the user is not on site, and great convenience is provided; and meanwhile, the spatial-temporal features and the biological signals are extracted, and feature fusion is performed by combining the two kinds of information, so that the recognition accuracy is enhanced, and the possibility of misrecognition is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent security, and in particular to an encrypted file box that can be unlocked by remote facial recognition. Background Art

[0002] With the rapid development of information technology, data security has become an issue that receives more and more attention.

[0003] Existing encrypted file boxes usually use passwords, fingerprint recognition and other technologies to unlock, but these technologies have certain security risks. On the one hand, the password is easy to be cracked, and fingerprint recognition may be affected by environmental factors. On the other hand, it is impossible to achieve remote recognition and unlocking. Therefore, it is necessary to develop a new encrypted file box that uses a safer and more reliable unlocking method. Summary of the invention

[0004] The purpose of this section is to summarize some aspects of embodiments of the present invention and briefly introduce some preferred embodiments. Some simplifications or omissions may be made in this section and the specification abstract and the invention title of this application to avoid blurring the purpose of this section, the specification abstract and the invention title, and such simplifications or omissions cannot be used to limit the scope of the present invention.

[0005] In view of the problems existing in the above-mentioned existing encryption method of encrypted file boxes, the present invention is proposed.

[0006] Therefore, the technical problem solved by the present invention is to solve the problem that the existing encrypted file box is not secure enough on the one hand, and cannot realize remote identification and unlocking on the other hand.

[0007] In order to solve the above technical problems, the present invention provides the following technical solutions: an encrypted file box that can be unlocked by remote facial recognition, characterized in that an identification system is installed on the encrypted file box, and after the identification system identifies the correct signal, it sends an instruction to a control unit embedded thereon, and the control unit controls the unlocking of the encrypted file box; wherein the identification system and the mobile client are both connected to the network, and the mutual transmission of data is achieved through the interaction between accounts; wherein the identification system has the following steps in the identification process: S1: the mobile client sends a signal to the identification system, and the identification system starts the identification program after receiving the signal, and sends the program to the mobile client in real time; S2: the user performs facial recognition on the mobile client according to the identification program; S3: during the facial recognition process, the recognition result is transmitted to the identification system in real time; S4: the recognition system integrates each recognition result in real time, and sends an instruction to the control unit when the recognition meets the standard.

[0008] As a preferred solution for the encrypted file box that can be unlocked by remote facial recognition as described in the present invention, when the user performs facial recognition on the mobile client according to the recognition program, the following steps are included: H1: the user inputs the current facial video in real time according to the recognition program; H2: real-time spatiotemporal feature extraction is performed on the input current facial video; H3: real-time bio-signal extraction is performed on the input current facial video; H4: feature fusion of the real-time extracted spatiotemporal features and bio-signals; H5: participate in dynamic verification decision-making based on the fused feature parameters to complete facial recognition verification.

[0009] As a preferred solution of the encrypted file box that can be unlocked by remote facial recognition as described in the present invention, wherein: after the current facial video is input in real time in step H1, it also includes data preprocessing of the input facial video; wherein the video data preprocessing includes facial ROI extraction, center cropping and bilinear interpolation, and the output size is 256×256 pixels.

[0010] As a preferred solution of the encrypted file box that can be unlocked by remote facial recognition according to the present invention, the real-time spatiotemporal feature extraction of the input current facial video specifically includes: E1: Calculate the space-time differential operator according to the following model:

[0011] Where I (x, y, t) is the video frame sequence; Spatial curvature; 0.35 is the spatiotemporal weight coefficient obtained through KL divergence optimization; E2: Hypersphere mapping compression is performed according to the following model:

[0012] Among them, W is a 128×512-dimensional trainable matrix; 0.35 is the spatiotemporal weight coefficient; is the pixel space.

[0013] As a preferred solution of the encrypted file box that can be unlocked by remote facial recognition according to the present invention, the real-time biometric signal extraction of the input current facial video specifically includes: P1: Skin color fluctuation modeling is performed based on the following model:

[0014] P2: Spectral analysis after modeling.

[0015] As a preferred solution of the encrypted file box that can be unlocked by remote facial recognition according to the present invention, the feature fusion of the spatiotemporal features and bio-signals extracted in real time specifically includes: Y1: Perform feature fusion model based on the following model to complete feature interaction calculation:

[0016] Among them, ⊙ is the Hadamard product operation; tanh is the hyperbolic tangent function, and the output is normalized to [-1, 1]; Y2: Dynamic template update based on the following model:

[0017] Among them, 50 is the feature norm normalization factor.

[0018] As a preferred solution of the encrypted file box that can be unlocked by remote facial recognition according to the present invention, the following model is used to participate in dynamic verification decision-making based on the fused feature parameters and complete facial recognition verification:

[0019] Among them, w k is the feature weight based on the attention mechanism; 0.45 is the confidence coefficient of the living signal.

[0020] As a preferred solution for the encrypted file box that can be unlocked by remote facial recognition as described in the present invention, when Score≥0.85, direct verification is passed; when 0.7≤Score<0.85, secondary verification is performed; when Score<0.7, rejection and alarm are given.

[0021] The present invention provides an encrypted file box that can be unlocked by remote facial recognition, which has the following beneficial effects: Improved security: Facial recognition technology is used as a means of unlocking. Compared with traditional passwords or fingerprint recognition, facial recognition technology is more difficult to crack, which effectively improves the security level of the file box.

[0022] Realize remote unlocking: Users can perform remote facial recognition through the mobile client, which enables them to safely unlock the file box even when they are not on site, providing great convenience.

[0023] Dynamic verification decision: Dynamic verification decisions are made through the integration of feature parameters. Different levels of verification measures can be taken in different situations, such as secondary verification, thereby ensuring security while improving user experience.

[0024] Combining spatiotemporal features with biological signals: By extracting spatiotemporal features and biological signals and combining these two types of information for feature fusion, the accuracy of recognition is enhanced and the possibility of misidentification is reduced.

[0025] Real-time data processing: Extracting spatiotemporal features and biometric signals from facial videos in real time and making verification decisions quickly ensures the responsiveness and efficiency of the encrypted file box.

[0026] Strong adaptability: Due to the use of facial recognition technology, the encrypted file box is not affected by environmental factors, such as moisture and dirt that may affect fingerprint recognition, which improves the adaptability of the use environment.

[0027] User-friendly: Users do not need to remember complex passwords or carry keys. They can unlock the device simply through facial recognition, which simplifies the usage process and improves user-friendliness. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative labor. Among them: Figure 1 A method flow chart of the identification process of the identification system provided by the present invention.

[0029] Figure 2 A flow chart of a method provided by the present invention for a user to perform facial recognition on a mobile client according to a recognition program. DETAILED DESCRIPTION

[0030] In order to make the above-mentioned purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in the art without creative work should fall within the scope of protection of the present invention.

[0031] The encryption method used in the existing encrypted file box is easy to be cracked, and fingerprint recognition may be affected by environmental factors. On the other hand, it is impossible to achieve remote recognition and unlocking. Therefore, it is necessary to develop a new encrypted file box with a safer and more reliable unlocking method.

[0032] Therefore, the present invention provides an encrypted document box that can be unlocked by remote facial recognition. A recognition system is installed on the encrypted document box. After the recognition system recognizes the correct signal, it sends a command to a control unit embedded thereon, and the control unit controls the unlocking of the encrypted document box. Among them, the identification system and the mobile client are both connected to the network, and the data is transmitted to each other through the interaction between accounts; Among them, see Figure 1 , the recognition system has the following steps in the recognition process: S1: The mobile client sends a signal to the recognition system. After receiving the signal, the recognition system starts the recognition program and sends the program to the mobile client in real time. S2: The user performs facial recognition on the mobile client according to the recognition procedure; S3: During the facial recognition process, the recognition results are transmitted to the recognition system in real time; S4: The recognition system integrates the recognition results in real time and sends instructions to the control unit when the recognition meets the standard.

[0033] For further information, see Figure 2 When a user performs facial recognition on a mobile client according to the recognition procedure, the following steps are included: H1: The user inputs the current facial video in real time according to the recognition program; Specifically, it also includes data preprocessing of the input facial video; The video data preprocessing includes facial ROI extraction, center cropping and bilinear interpolation, and the output size is 256×256 pixels.

[0034] Right now: Input: 30fps video stream (resolution ≥ 720p); Color space: YUV420 format (compatible with mobile hardware encoding); ROI extraction: locate the face through lightweight YOLO-Face (latency < 5ms), and the output size is 256×256 pixels after center cropping and bilinear interpolation; Among them, the facial ROI is extracted, the center cropped and bilinear interpolated using the existing conventional lightweight YOLO-Face model, which will not be described in detail here. The following is the conventional code for facial ROI: # Use lightweight YOLO-Face model bbox = yoloface.predict(frame) face_roi = frame[bbox.y:bbox.y+bbox.h, bbox.x:bbox.x+bbox.w] H2: Extract the spatiotemporal features of the input current facial video in real time, including: E1: Calculate the space-time differential operator according to the following model:

[0035] Where I (x, y, t) is the video frame sequence; is the spatial curvature; 0.35 is the spatiotemporal weight coefficient obtained through KL divergence optimization; The meanings of the parameters of the above-mentioned space-time differential operators are explained in turn as follows: ①G(x,y,t) Meaning: Spatiotemporal joint feature map.

[0036] Physical meaning: The output result is the dynamic feature intensity value of each pixel (x, y) at time t.

[0037] High-value areas indicate dramatic changes in facial texture (such as micro-expressions, blood vessel pulsation, etc.).

[0038] Value range: R (real number domain, can be positive or negative, needs to be normalized).

[0039] ②I(x,y,t) Meaning: Pixel intensity function of a video frame.

[0040] Specific definition: x, y: image plane space coordinates (unit: pixels).

[0041] t: time variable (unit: seconds).

[0042] I∈[0, 255] represents the grayscale intensity value (if it is RGB, it needs to be processed by channels).

[0043] Function: Serves as basic input data for spatiotemporal analysis.

[0044] ③ (Second-order mixed derivative) Mathematical definition:

[0045] Physical meaning: Measures the cross curvature variation of pixel intensity in the x and y directions.

[0046] Used to capture subtle spatial differences in facial texture (e.g. wrinkles, pore distribution).

[0047] Implementation: Calculated by Sobel cross convolution kernel:

[0048] ④ (Time derivative) Mathematical definition:

[0049] Physical meaning: Represents the instantaneous rate of change of pixel intensity between adjacent frames.

[0050] Used to detect facial movements (such as blinks, muscle tremors) and lighting changes.

[0051] Discrete computing:

[0052] ⑤ (Laplacian operator) Mathematical definition:

[0053] Physical meaning: Detect edges and curvature in images and enhance high-frequency details (e.g. facial contours, hair).

[0054] Used to enhance the discrimination of spatial features.

[0055] Implementation: Use the Laplacian filter.

[0056] ⑥ 0.35 (time-space weight coefficient) Meaning: Relative weight of the time derivative term.

[0057] Physical meaning: Controlling time dynamic information The contribution ratio to the overall feature G(x, y, t).

[0058] The larger the value, the more significant the effect of temporal changes on the feature.

[0059] Optimization basis: By maximizing the KL divergence, it was determined that among 100,000 groups of samples, when the weight was 0.35, the distinction between live and attack samples was the highest.

[0060] Experimental data: Weight = 0.2 → Liveness Detection Rate (TAR) = 98.3%; Weight = 0.35 → TAR = 99.7% (optimal); Weight = 0.5 → TAR = 97.1% (overfitting temporal noise).

[0061] As follows, this solution additionally provides implementation code for calculating spatiotemporal differential operators: # Sobel operator calculates the second-order derivative I_xy = cv2.Sobel(frame, cv2.CV_32F, 1, 1, ksize=3) # Time difference I_t = (current_frame - prev_frame) / Δt # Laplacian operator laplacian = cv2.Laplacian(frame, cv2.CV_32F) G = I_xy + 0.35 * I_t * laplacian E2: Hypersphere mapping compression is performed according to the following model:

[0062] Among them, W is a 128×512-dimensional trainable matrix; 0.35 is the spatiotemporal weight coefficient; is the pixel space.

[0063] It should be noted that the following are the physical meanings of the parameters in the model: ①F proj Meaning: The projected eigenvector.

[0064] Physical meaning: The output is a dynamic feature representation after dimensionality reduction, which is used for subsequent liveness detection and identity authentication.

[0065] Through normalization and nonlinear activation, the discriminability and robustness of features are enhanced.

[0066] Dimension: R 128 (Assume W is a 128×512128×512 matrix).

[0067] Example values: [0.45, 1.2, 0.0, …, 0.8], ReLU will set negative values ​​to zero.

[0068] ②W Meaning: Trainable projection matrix.

[0069] Mathematical definition: W∈R 128×512 ,in: Row number 128128: target feature dimension (compressed dimension).

[0070] Number of columns: 512512: The original dimension of the input feature G.

[0071] effect: The original 512-dimensional spatiotemporal feature G is linearly mapped to a 128-dimensional space.

[0072] The most discriminative feature directions are learned to be preserved through training.

[0073] Training method: End-to-end optimization using Triplet Loss.

[0074] Initialize to Xavier normal distribution to prevent gradient disappearance.

[0075] ③ ∥G∥2 Mathematical definition: L2 norm (Euclidean norm) of vector G:

[0076] effect: Normalize the feature vector to eliminate the influence of light intensity.

[0077] Ensure the stability of characteristic scale under different lighting conditions.

[0078] Example calculation: If G=[3, 4], then ∥G∥2=5.

[0079] ④ϵ Meaning: Numerical stabilization constant.

[0080] Value: ϵ=1×10 −6 .

[0081] effect: Prevent the denominator from being zero (avoid division by zero errors when ∥G∥2 approaches zero).

[0082] Ensure the stability of gradient calculation.

[0083] Selection basis: empirical value, usually a value close to the lower limit of floating-point precision.

[0084] ⑤ReLU(⋅) Mathematical definition: Rectified Linear Unit:

[0085] use: Introduce nonlinearity to enhance the expressiveness of the model.

[0086] Filter out negative features and retain positive activations (consistent with the non-negative nature of biological features).

[0087] Example: Input x=−0.5, output 00; input x=1.2, output 1.2.

[0088] It should be noted that in the process of calculating the space-time differential operator, the traditional scheme only uses ∇ 2I or ∂I / ∂t, this model captures the curvature changes of facial micro-expressions through the second-order mixed derivative G(x, y, t), and experiments show that it can improve the dynamic feature discrimination by 23%. When performing hypersphere mapping compression, this scheme compresses 512-dimensional features to 128 dimensions and retains 99.2% of the original information (traditional PCA only retains 89%).

[0089] H3: Real-time bio-signal extraction of the current facial video input, including: P1: Skin color fluctuation modeling is performed based on the following model:

[0090] P2: Spectral analysis after modeling.

[0091] It should be noted that the physical meanings of the parameters in the model are as follows: ①ΔC(t) Meaning: Synthesized blood flow fluctuation signal.

[0092] Physical meaning: Reflects the periodic volume changes of subcutaneous blood vessels caused by heartbeat.

[0093] Used for liveness detection and heart rate estimation.

[0094] Unit: dimensionless (normalized).

[0095] ②

[0096] Meaning: The average brightness value of the RGB channels of the facial region (ROI).

[0097] Calculation method: After detecting the face area, calculate the RGB mean of all pixels in the area:

[0098] N: total number of pixels in ROI; R i : Red channel value of the i-th pixel (0-255); Physical meaning: Red light (R) is sensitive to the absorbance of hemoglobin and reflects changes in blood oxygen.

[0099] Green light (G) is susceptible to motion interference but contains some blood flow information.

[0100] Blue light (B) is sensitive to changes in venous blood and assists in noise reduction.

[0101] ③

[0102] Meaning: The time derivative of the mean brightness of the RGB channels.

[0103] Mathematical definition:

[0104] Discrete calculation (taking 30fps video as an example):

[0105] Physical meaning: Capture microsecond brightness fluctuations caused by a heartbeat (amplitude typically <1%).

[0106] The red light derivative mainly contributes to the pulse wave, while the green light derivative contains more motion noise.

[0107] ④Weight coefficient 0.76, −0.35, 0.55 effect: Through linear combination optimization, blood flow signals are enhanced and noise is suppressed.

[0108] Red (R) has the highest weight: it is most sensitive to changes in blood oxygen.

[0109] Green (G) negative weights: counteract motion artifacts (such as lighting changes caused by head shaking).

[0110] Blue (B) auxiliary weight: supplements venous blood information and balances skin color differences.

[0111] Optimization basis: Principal Component Analysis (PCA): Decomposes the feature directions related to heart rate on the MIT rPPG dataset.

[0112] Maximizing the signal-to-noise ratio: Solved by constrained optimization:

[0113] The experiment shows that the optimal weight is α R =0.76,α G =−0.35,α B =0.55.

[0114] Normalization verification:

[0115] It should be noted that the spectrum analysis after modeling includes the following details: ①Signal preprocessing Input signal: ΔC(t) (blood flow fluctuation signal, length N frames) Processing steps: Detrending:

[0116] Bandpass filtering: Adopt 5th order Butterworth filter, cut-off frequency 0.75Hz-4Hz (corresponding to 45-240bpm) Transfer function:

[0117] standardization:

[0118] ②Windowing Window function selection: Using the Blackman-Harris window, the expression is:

[0119] The coefficients are: a0=0.35875, a1=0.48829, a2=0.14128, a3=0.01168; ③FFT transformation and spectrum refinement FFT calculation:

[0120] Spectrum refinement technology: Zero padding: Expand the FFT points to 4096 and improve the frequency accuracy to 0.0073Hz; Resampling: 8x linear interpolation in the range of 0.75-4Hz; Energy correction: compensates for the amplitude attenuation caused by the window function; Frequency-heart rate conversion:

[0121] ③Peak detection algorithm Three-step verification method: Candidate peak location: Find the local maximum point, the amplitude must be more than 3 times the average energy; Harmonic Verification: Check if Equal harmonic components; Continuity test: the heart rate change rate of 5 consecutive frames is ≤15% (suppressing transient noise); Mathematical conditions:

[0122] ⑤Multi-frame fusion technology Sliding window processing: Window length W = 10 seconds, overlap rate 50%; The heart rate value output for each frame is the weighted average of three adjacent windows:

[0123] ⑥Motion artifact suppression Accelerometer Fusion: Modeling motion noise using the phone's built-in accelerometer data:

[0124] α i : Coupling coefficient estimated online by the least squares method; Adaptive filtering: The NLMS (normalized least mean square) algorithm was used with a step factor of μ = 0.02; ⑦Light mutation treatment Robust STFT: When the light gradient dI / dt>50 lux / s is detected: Shorten the window to N=256 (improve time domain resolution) Enhanced instantaneous frequency capture using Wigner-Ville distribution ⑧Performance verification indicators 1. Signal-to-noise ratio calculation

[0125] 2. Heart rate error assessment

[0126] H4: Feature fusion of real-time extracted spatiotemporal features and biological signals specifically includes: Y1: Perform feature fusion model based on the following model to complete feature interaction calculation:

[0127] Among them, ⊙ is the Hadamard product operation; tanh is the hyperbolic tangent function, and the output is normalized to [-1,1]; Y2: Dynamic template update based on the following model:

[0128] Among them, 50 is the feature norm normalization factor.

[0129] It should be noted that the following are the physical meanings of the current model parameter symbols: ①T n With T n+1 Meaning: Dynamic feature template at time steps n and n+1.

[0130] Physical meaning: T n : The facial feature template at the current moment, including historical information.

[0131] T n+1 : The updated template combines the weighted results of the historical template and the new changes.

[0132] Data type: multi-dimensional vector (e.g. 128-dimensional features).

[0133] ②γ (attenuation coefficient) Mathematical definition: γ∈[0, 0.8], dynamically adjusted by the magnitude of feature changes.

[0134] Physical meaning: High γ (close to 0.8): When the feature changes slowly, more historical template information is retained to enhance system stability.

[0135] Low γ (close to 0): When features change dramatically, new features are quickly absorbed to improve adaptability to dynamic scenes.

[0136] Design basis: Exponential decay function: smoothly adjusts weights to avoid sudden changes (such as sudden changes in lighting or occlusion).

[0137] 0.8 is an empirical value, which ensures that 80% of historical information is retained when the feature is stable.

[0138] ③ (Characteristic time derivative) Mathematical definition:

[0139] Physical meaning: Captures the rate of change of the feature vector F over time and reflects facial dynamics (such as blood flow pulsation and muscle tremors).

[0140] Example: If there are 30 frames per second (Δt=1 / 30s), calculate the feature difference between two adjacent frames.

[0141] ④∥F fusion ∥2 (L2 norm) Mathematical definition:

[0142] Physical meaning: It represents the overall strength of the feature vector and quantifies the significance of the feature at the current moment.

[0143] High norm: Features change dramatically (such as rapid head turns, strong lighting changes).

[0144] Low norm: The feature is stable (like a stationary state).

[0145] ⑤ Constant 50 (normalization factor) Function: Scale the feature norm to a reasonable range and balance the sensitivity of the exponential function.

[0146] Design basis: The experimental results show that ∥F fusion The average value of ∥2 is 30-40, and the maximum value does not exceed 100.

[0147] When ∥F fusion When ∥2=50, γ=0.8⋅e −1 ≈0.294, and the new change weight (1−γ)≈0.706.

[0148] ⑥0.8 (basic attenuation weight) Physical meaning: The characteristic is completely stable (∥F fusion ∥2=0) retains 80% of the historical templates.

[0149] Optimization basis: Determined by cross validation: Too low (such as 0.6): The template is updated too quickly and is easily affected by noise.

[0150] Too high (such as 0.9): The response is slow and cannot adapt to real dynamic changes.

[0151] Dynamic adjustment example: Scenario 1: Static state ∥F fusion ∥2=20 γ=0.8⋅e −20 / 50 ≈0.8⋅e −0.4 ≈0.536 Update weight: 53.6% historical template + 46.4% new changes.

[0152] Scenario 2: Quick head turn ∥F fusion ∥2=60 γ=0.8⋅e −60 / 50 ≈0.8⋅e −1.2 ≈0.241 Update weight: 24.1% historical template + 75.9% new changes, quickly adapt to motion blur.

[0153] Scene 3: Sudden change under strong light ∥F fusion ∥2=80 γ=0.8⋅e −80 / 50 ≈0.8⋅e −1.6 ≈0.161 Update weight: 16.1% historical template + 83.9% new changes, suppressing outdated lighting information.

[0154] H5: Participate in dynamic verification decision-making based on the fused feature parameters and complete facial recognition verification based on the following model:

[0155] It should be noted that the physical meanings of the above model parameter symbols are as follows: ① (Spatial similarity term) Physical meaning: Weighted cosine similarity: The facial features are divided into 64 sub-regions, and the similarity of each region is calculated and then weighted summed.

[0156] Used to measure the spatial matching degree between the current facial features and the registered template.

[0157] Parameter breakdown: k=1, 2, ..., 64: Represents 64 characteristic sub-regions of the face (such as left eye, right cheek, nose bridge, etc.), divided by feature points or convolutional neural network (CNN) activation maps.

[0158] w k (Weight): The importance weight of the k-th sub-region in the validation.

[0159] Determined by attention mechanisms or statistical learning (e.g., eye regions are usually weighted higher).

[0160] Constraints:

[0161] cosθ k (Cosine similarity): The cosine value of the eigenvector angle of the kth sub-region:

[0162] The value is [−1, 1][−1, 1]. The closer it is to 1, the more matching the features of the region are.

[0163] ② (Live signal strength item) Physical meaning: Calculate the average energy intensity of physiological signals (such as blood flow fluctuations) in the past 2 seconds to verify the existence of living bodies.

[0164] Parameter breakdown: Integration interval [t−2, t]: A 2-second sliding window is used to balance real-time performance (short-term response) and stability (suppressing transient noise).

[0165] ∥ΔC∥2 (L2 norm): ΔC: blood flow fluctuation signal extracted by rPPG (remote photoplethysmography).

[0166] L2 norm: , which represents the total energy of the signal in the time domain.

[0167] Physiological signals of living bodies (such as heartbeats) exhibit periodic high energy, while attacks (such as photos / videos) have weak or disordered energy.

[0168] 1 / 2 (normalization factor): The results of the 2-second window integration are averaged to eliminate the effect of the window length on the value range.

[0169] The typical value of the living body signal term is made to fall in [0,1][0,1], which is consistent with the dimension of the spatial similarity term.

[0170] Furthermore, when Score≥0.85, the verification is passed directly; when 0.7≤Score<0.85, a secondary verification is performed; when Score<0.7, it is rejected and an alarm is issued.

[0171] In addition, in order to verify the beneficial effects of the present invention, the following simulation experiments are carried out simultaneously: Purpose of the test: Verify the security and accuracy of remote facial recognition unlocking of encrypted document boxes under different conditions, including different lighting, angles, and expressions, as well as the ability to resist spoofing attacks.

[0172] Test equipment: Remote facial recognition to unlock encrypted file boxes Multiple mobile client devices (such as smartphones) Standard light test box High-speed camera Data acquisition and processing system Test steps: System calibration: Ensure that all equipment is calibrated correctly, including the identification system of the encrypted document box, mobile clients, and high-speed cameras.

[0173] Data collection: In a controlled laboratory environment, facial video data of the same user is collected under different lighting conditions (strong light, weak light, indoor, outdoor), angles (front, side, looking down, looking up) and expressions (natural, smiling, serious, angry). At the same time, the environmental parameters under each condition are recorded.

[0174] Deception attack simulation: Use masks, photos, video playback and other means to simulate deception attacks and collect deception data.

[0175] Feature extraction and fusion: Extract spatiotemporal features and bio-signals from the collected real user facial videos and spoofing data, and perform feature fusion.

[0176] Verification test: Input the fused feature parameters into the dynamic verification decision model for verification, and record the verification score and results.

[0177] Result Record: Record detailed information of each verification, including user ID, test conditions, spatiotemporal feature scores, biosignal scores, fusion feature scores, and verification results.

[0178] Data table: Table 1 User ID Lighting conditions angle expression Spatiotemporal feature scores Biosignal score Fusion feature score Verify the results 001 Indoor strong light front nature 0.92 0.89 0.91 pass 002 Indoor low light side Smile 0.78 0.80 0.79 Secondary verification 003 outdoor front serious 0.85 0.83 0.84 pass 004 outdoor side laugh 0.69 0.71 0.70 reject ... ... ... ... ... ... ... ... Mask deception indoor front - 0.50 0.55 0.52 reject Photo Deception indoor - - 0.60 0.62 0.61 reject ... ... ... ... ... ... ... ... Analysis of test results: Through the specific data in the above table, the following analysis can be carried out: Accuracy evaluation: Analyze the pass rate of real users under different conditions, and the rejection rate of spoofing attacks.

[0179] Stability evaluation: Analyze the stability of the system under different lighting, angles and expression conditions, that is, the fluctuation range of the feature score.

[0180] Security assessment: Evaluate the system's ability to resist spoofing attacks, measured by the rejection rate.

[0181] In summary, as shown in Table 2 below, this is a comparison table of the technical advantages of this solution and the traditional solution: Table 2 Technical indicators Traditional solutions (such as FaceNet) This solution (ST-DFE) Liveness detection method Action commands (blink, etc.) Non-sensitized microvascular pulsation analysis Feature Dimension 512-dimensional static vector 128-dimensional dynamic tensor Anti-counterfeiting ability Resist 2D attacks Resisting 3D printing / video attacks Computational complexity 1.2 GOPS 0.35 GOPS Model size 12.5 MB 3.8 MB This solution uses spatiotemporal-biophysical joint modeling to implement a facial recognition system with a security level of FIDO2 Level 3 on the mobile terminal, and meets the real-time computing requirements of the Android terminal (single frame processing time ≤ 15ms).

[0182] The present invention provides an encrypted file box that can be unlocked by remote facial recognition, which has the following beneficial effects: Improved security: Facial recognition technology is used as a means of unlocking. Compared with traditional passwords or fingerprint recognition, facial recognition technology is more difficult to crack, which effectively improves the security level of the file box.

[0183] Realize remote unlocking: Users can perform remote facial recognition through the mobile client, which enables them to safely unlock the file box even when they are not on site, providing great convenience.

[0184] Dynamic verification decision: Dynamic verification decisions are made through the integration of feature parameters. Different levels of verification measures can be taken in different situations, such as secondary verification, thereby ensuring security while improving user experience.

[0185] Combining spatiotemporal features with biological signals: By extracting spatiotemporal features and biological signals and combining these two types of information for feature fusion, the accuracy of recognition is enhanced and the possibility of misidentification is reduced.

[0186] Real-time data processing: Extracting spatiotemporal features and biometric signals from facial videos in real time and making verification decisions quickly ensures the responsiveness and efficiency of the encrypted file box.

[0187] Strong adaptability: Due to the use of facial recognition technology, the encrypted file box is not affected by environmental factors, such as moisture and dirt that may affect fingerprint recognition, which improves the adaptability of the use environment.

[0188] User-friendly: Users do not need to remember complex passwords or carry keys. They can unlock the door simply by facial recognition, which simplifies the usage process and improves user-friendliness.

[0189] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. An encrypted file box that can be unlocked remotely by facial recognition, characterized in that: The encrypted document box is provided with an identification system, and after the identification system identifies a correct signal, it sends a command to a control unit embedded thereon, and the control unit controls the unlocking of the encrypted document box; The identification system and the mobile client are both connected to the network, and data is transmitted to each other through interaction between accounts; The identification system comprises the following steps in the identification process: S1: The mobile client sends a signal to the recognition system, and the recognition system starts the recognition program after receiving the signal and sends the program to the mobile client in real time; S2: The user performs facial recognition according to the recognition procedure on the mobile client; S3: During the facial recognition process, the recognition result is transmitted to the recognition system; S4: The recognition system receives the recognition result in real time, and sends a command to the control unit when the recognition result indicates a pass.

2. The encrypted file box that can be unlocked by remote facial recognition according to claim 1, characterized in that: When a user performs facial recognition on a mobile client according to the recognition procedure, the following steps are included: H1: The user inputs the current facial video in real time according to the recognition program; H2: Extract real-time spatiotemporal features of the input current facial video; H3: Real-time biometric signal extraction for the input current facial video; H4: Fuse the spatiotemporal features and biological signals extracted in real time; H5: Participate in dynamic verification decision-making based on the fused feature parameters to complete facial recognition verification.

3. The encrypted document box that can be unlocked by remote facial recognition according to claim 2, characterized in that: After the current facial video is input in real time in step H1, it also includes data preprocessing of the input facial video; The video data preprocessing includes facial ROI extraction, center cropping and bilinear interpolation, and the output size is 256×256 pixels.

4. The encrypted file box that can be unlocked by remote facial recognition according to claim 3, characterized in that: The real-time spatiotemporal feature extraction of the input current facial video specifically includes: E1: Calculate the space-time differential operator according to the following model: Where I (x, y, t) is the video frame sequence; is the spatial curvature; 0.35 is the spatiotemporal weight coefficient obtained through KL divergence optimization; E2: Hypersphere mapping compression is performed according to the following model: Among them, W is a 128×512 dimensional trainable matrix.

5. The encrypted file box that can be unlocked by remote facial recognition according to claim 4, characterized in that: Real-time bio-signal extraction of the input current facial video specifically includes: P1: Skin color fluctuation modeling is performed based on the following model: P2: Spectral analysis after modeling.

6. The encrypted document box that can be unlocked by remote facial recognition according to claim 5, characterized in that: The feature fusion of the spatiotemporal features extracted in real time and the biological signals specifically includes: Y1: Perform feature fusion model based on the following model to complete feature interaction calculation: Among them, ⊙ is the Hadamard product operation; tanh is the hyperbolic tangent function, and the output is normalized to [-1, 1]; Y2: Dynamic template update based on the following model: Among them, 50 is the feature norm normalization factor.

7. The encrypted document box capable of being unlocked by remote facial recognition according to claim 6, characterized in that: Based on the fused feature parameters, dynamic verification decisions are made to complete facial recognition verification based on the following model: Among them, w k is the feature weight based on the attention mechanism; 0.45 is the confidence coefficient of the living signal.

8. The encrypted document box capable of being unlocked by remote facial recognition according to claim 7, characterized in that: When Score≥0.85, the verification is passed directly; when 0.7≤Score<0.85, a secondary verification is performed; when Score<0.7, it is rejected and an alarm is issued.

Images