Living body rechecking method and device based on deep pseudo detection model

Abstract

The invention relates to the technical field of computers, in particular to a living body rechecking method and device based on a deep pseudo detection model, and the method comprises the steps: obtaining an authentication event data set corresponding to a living body authentication event, the authentication event data set comprises a face image, an identity comparison result, a terminal equipment identifier, network address information, a geographic position and a historical authentication record; inputting the face image into a deep false detection model to obtain a deep false detection feature, constructing a risk scoring model based on the authentication event data set and the deep false detection feature, inputting data, outputting a risk score, and generating a rechecking task list according to the risk score; and outputting the rechecking task list on the manual rechecking interface, receiving a rechecking conclusion, outputting a risk control processing instruction to the business system in response to the rechecking conclusion, and using the rechecking conclusion to update the risk scoring model. According to the method, through multi-source element fusion scoring and re-check closed-loop updating, the deep pseudo attack recognition and disposal efficiency is improved.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and in particular to a liveness verification method and apparatus based on a deep fakeness detection model. Background Technology

[0002] With the widespread adoption of remote real-name registration and online account opening, payment, and credit services, liveness detection has become a critical security element. Generative technologies lower the barrier to entry for face-swapping, synthetic, and replay attacks, allowing attacks to be launched in batches and bypass defenses relying solely on face comparison or single liveness detection in a short period. Simultaneously, differences in terminal environments, network proxies, and location drift make it more difficult to accurately identify abnormal behavior using a single signal. In existing solutions, risk control rules and models are often deployed in a fragmented manner, and alarm sorting and manual review lack a unified evidence chain and closed-loop update mechanism. This results in uneven allocation of review resources, a coexistence of false positives and false negatives, and a lag in adapting to new types of attacks. Summary of the Invention

[0003] This invention provides a liveness verification method and apparatus based on a deepfake detection model, which is used to at least solve the problem of how to achieve high-priority verification scheduling and form a continuously updated risk control closed loop in scenarios where deepfake attacks and terminal network anomalies overlap.

[0004] In a first aspect, the present invention provides a liveness verification method based on a deep fake detection model, comprising the following steps: Obtain the authentication event dataset corresponding to the liveness authentication event. The authentication event dataset includes face images, identity comparison results, terminal device identifiers, network address information, geographical location, and historical authentication records. The historical authentication records include face images and authentication results of historical authentications. The face image is input into the deepfake detection model to obtain deepfake detection features. Based on the authentication event dataset and the deepfake detection features, the input data of the risk scoring model is constructed and the risk score is output. Based on the risk score, a review task list is generated. The manual review interface outputs a list of review tasks and receives review conclusions. In response to the review conclusions, it outputs risk control instructions to the business system and uses the review conclusions to update the risk scoring model.

[0005] In one possible implementation, obtaining the authentication event dataset includes: obtaining face images and identity comparison results from the identity authentication system, obtaining terminal device identifiers, network address information and geographical location from the terminal information collection system, obtaining historical authentication records from historical authentication storage, and performing field integrity verification, format conversion and deduplication merging processing on the authentication event dataset.

[0006] In one possible implementation, before inputting the face image into the deep fake detection model, face region localization, face key point localization, face image alignment and size normalization are performed on the face image, and the aligned face image is used as the input to the deep fake detection model.

[0007] In one possible implementation, the deep fake detection model includes a spatial domain feature extraction network and a frequency domain feature extraction network. The spatial domain feature extraction network extracts texture features from the face image, and the frequency domain feature extraction network performs a frequency domain transformation on the face image and extracts forgery trace features. The texture features and forgery trace features are fused to obtain the deep fake detection features.

[0008] In one possible implementation, the input data for constructing the risk scoring model includes: obtaining terminal reuse frequency features by statistically analyzing the frequency of terminal device identifiers based on terminal device identifiers and historical authentication records, and obtaining terminal consistency features by performing a matching consistency determination based on terminal device identifiers and terminal device identifiers in historical authentication records; obtaining network segment features by performing network segment division based on network address information, and obtaining network affiliation features by performing operator attribution resolution based on network address information; obtaining location change frequency features by statistically analyzing the frequency of geographical location changes based on geographical location and historical authentication records, and obtaining authentication frequency features by statistically analyzing the frequency of authentication occurrences based on historical authentication records; obtaining background region images by performing face region localization based on historical authentication face images in historical authentication records, extracting features from the background region images to obtain background feature vectors, performing unsupervised clustering analysis on the background feature vectors to obtain background cluster identifiers, and statistically analyzing background reuse frequency features based on the background cluster identifiers; and concatenating the terminal reuse frequency features, terminal consistency features, network segment features, network affiliation features, location change frequency features, authentication frequency features, background reuse frequency features, and deep fake detection features with deep fake detection features to obtain the input feature vector.

[0009] In one possible implementation, the risk scoring model is a supervised learning classification model, which is trained based on historical authentication samples labeled with review conclusions and outputs a risk score to the input feature vector.

[0010] In one possible implementation, generating the review task list includes: sorting the liveness authentication events according to their risk scores, and generating a review task entry for each liveness authentication event that includes the risk score and the liveness authentication event identifier.

[0011] In one possible implementation, the review task list output on the manual review interface includes: associating the review task items with the output of face images, historical authentication face images from historical authentication records, identity comparison results, terminal device identifiers, network address information, geographical location and risk scores, and generating and outputting deep fake detection judgment results based on deep fake detection features.

[0012] In one possible implementation, using the review conclusion to update the risk scoring model includes: writing the review conclusion and the input feature vector of the risk scoring model into the training sample set, and retraining the risk scoring model based on the training sample set to update the risk scoring model parameters.

[0013] Secondly, the present invention provides a liveness verification device based on a deepfake detection model, used to implement the liveness verification method based on the deepfake detection model, the system comprising: The authentication data retrieval module is used to obtain the authentication event dataset corresponding to the liveness authentication event. The authentication event dataset includes face images, identity comparison results, terminal device identifiers, network address information, geographical location and historical authentication records. The historical authentication records include face images and authentication results of historical authentications. The deep fake detection scoring module is used to input face images into the deep fake detection model to obtain deep fake detection features, construct input data for the risk scoring model based on the authentication event dataset and deep fake detection features, and output risk scores. Based on the risk scores, a review task list is generated. The review and handling module is used to output a list of review tasks and receive review conclusions on the manual review interface, output risk control handling instructions to the business system in response to the review conclusions, and use the review conclusions to update the risk scoring model.

[0014] Compared with the prior art, the advantages and beneficial effects of the present invention are as follows: By integrating deepfake detection features with terminal device identifiers, network address information, geographical location, and historical authentication records through joint modeling techniques, we achieved collaborative identification of deepfake attacks and batch reuse anomalies. Through risk score-driven review task list generation, we achieved a prioritized allocation of review resources to high-risk events. Through closed-loop update techniques involving the feedback and retraining of manual review conclusions, we achieved continuous adaptation of the risk scoring model to new attacks and reduced model drift. Finally, through integrated evidence chain presentation in the review interface, we improved the consistency of review judgments and the timeliness of processing. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the execution flow of the method of the present invention; Figure 2 This is a structural block diagram of the device of the present invention. Detailed Implementation

[0016] The embodiments of the present disclosure will now be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of the disclosure. In the following detailed description, numerous specific details are set forth to provide a thorough understanding of the embodiments of the present disclosure for ease of explanation. However, it will be apparent that one or more embodiments may be practiced without these specific details. Furthermore, descriptions of well-known structures and techniques are omitted in the following description to avoid unnecessarily obscuring the concepts of the present disclosure.

[0017] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit this disclosure. The terms “comprising,” “including,” etc., as used herein indicate the presence of the stated features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0018] All terms used herein (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein are to be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0019] Deepfake detection models are machine learning models that determine the authenticity of deepfake-generated or edited face images. Their input is typically a face image or a feature representation derived from a face image, and the output is judgment information or feature vectors characterizing the risk of forgery. The key to this type of model is extracting relatively stable clues related to the forgery process from the image, such as subtle texture discontinuities, abnormal edge transitions, and frequency domain distribution shifts introduced by synthesis, resampling, and compression processes. These clues are then mapped into quantifiable deepfake detection features. Since deepfake detection features can reflect the source of a face image's generation and traces of tampering to a certain extent, deepfake detection models can not only directly provide forgery risk judgments but also be combined with upstream feature extractors and business-side risk control elements to achieve a comprehensive assessment of authentication event risks. Based on this, this invention uses the deepfake detection model as a core component of liveness verification and coordinates it with information such as terminal, network, geographic location, and historical authentication behavior to form a closed loop for verification task generation and processing in liveness authentication scenarios.

[0020] like Figure 1 As shown, a liveness verification method based on a deep fake detection model includes the following steps: Obtain the authentication event dataset corresponding to the liveness authentication event. The authentication event dataset includes face images, identity comparison results, terminal device identifiers, network address information, geographical location, and historical authentication records. The historical authentication records include face images and authentication results of historical authentications. In one embodiment, after completing a liveness authentication, the liveness authentication system generates an authentication event identifier and aggregates it from the identity authentication system, terminal information collection system, and historical authentication storage to form an authentication event dataset. The facial images in the authentication event dataset are image frames obtained from the current liveness acquisition; the identity comparison result is the comparison judgment or score obtained by comparing the facial image with the registered identity information; the terminal device identifier is the device fingerprint identifier generated by the terminal hardware and operating environment; the network address information is the source address record when the terminal accesses the network; and the geographical location is the location information reported by the terminal. Historical authentication records are retrieved from the historical authentication storage using the user identifier, and include the facial images and authentication results of previous authentications. The authentication event dataset is written into the event data storage according to a preset field structure for direct reading by subsequent deepfake detection and risk scoring steps, maintaining data consistency.

[0021] Obtaining the authentication event dataset includes: obtaining face images and identity comparison results from the identity authentication system, obtaining terminal device identifiers, network address information and geographical location from the terminal information collection system, obtaining historical authentication records from historical authentication storage, and performing field integrity verification, format conversion and deduplication merging processing on the authentication event dataset.

[0022] In one embodiment, after completing a liveness authentication, the liveness authentication system generates an authentication event identifier and writes it into an authentication event trigger message, which carries the user identifier and authentication time. In response to the authentication event trigger message, the liveness verification system initiates data retrieval requests to the identity authentication system, the terminal information collection system, and the historical authentication storage, respectively, and aggregates these requests using the authentication event identifier as the association key to form an authentication event dataset.

[0023] The facial image obtained from the identity authentication system is the facial image data collected for this liveness authentication. The identity comparison result is the comparison judgment information or comparison score information obtained by comparing the facial image with the registered identity information. During authentication, the terminal information collection system collects and reports the terminal device identifier, network address information, and geographical location. The terminal device identifier is a device fingerprint identifier generated from the terminal hardware information and operating environment information; the network address information is the source address record when the terminal accesses the network; and the geographical location is the location record reported by the terminal positioning module. Historical authentication storage returns historical authentication records based on the user identifier. Historical authentication records include the facial images and authentication results of previous authentications, and carry the authentication time and terminal device identifier corresponding to each historical authentication record for subsequent verification and comparison.

[0024] After the authentication event dataset is constructed, the liveness verification system performs field integrity checks on the dataset. These checks include verifying the existence and non-emptiness of fields such as face image, identity comparison result, terminal device identifier, network address information, geographical location, and historical authentication records, and verifying that the field types meet the preset field structure requirements. If a field does not meet the preset field structure requirements, the liveness verification system writes the authentication event identifier to an exception queue and records the missing field name. This prevents subsequent verification tasks from entering the risk scoring process and instead redirects them to manual data entry or re-collection, thus avoiding judgment bias caused by incomplete risk scoring input data.

[0025] After the field integrity verification passes, the liveness verification system performs format conversion on the authentication event dataset. This conversion includes uniformly converting face images to a preset image encoding format and standardizing image size annotations; uniformly converting identity comparison results to preset structured fields; uniformly converting network address information to a preset address field format; uniformly converting geographic location to a preset coordinate field format; and uniformly converting the time field returned by each source system to ensure consistency in the semantics of the same field across different sources. Subsequently, deduplication and merging are performed. This process uses the authentication event identifier as the primary key and the user identifier as the secondary key. Duplicate records generated by retry reporting from the identity authentication system and terminal information collection system are compared using hash fingerprints, retaining only unique records. Historical authentication records returned by the historical authentication storage are sorted by authentication time and merged into the authentication event dataset. Finally, the authentication event dataset is written to the event data storage and generates structured input that can be directly read by the deepfake detection model and risk scoring model, thus ensuring data consistency and reproducibility in subsequent model inference and task ranking processes.

[0026] The face image is input into the deepfake detection model to obtain deepfake detection features. Based on the authentication event dataset and the deepfake detection features, the input data of the risk scoring model is constructed and the risk score is output. Based on the risk score, a review task list is generated. In one embodiment, after the liveness verification system reads the authentication event dataset from the event data storage, it feeds the face images from the authentication event dataset into a deepfake detection model for inference. The deepfake detection model outputs deepfake detection features to characterize forgery traces. To enable the risk scoring model to stably use authentication information from different sources, the liveness verification system performs feature processing on the identity comparison results, terminal device identifiers, network address information, geographical location, and historical authentication records in the authentication event dataset. Feature processing includes converting the identity comparison results into a unified numerical field, mapping the terminal device identifier and network address information to discrete category features, mapping the geographical location to region coding features, and statistically analyzing the authentication frequency and location change count based on historical authentication records. These features are then combined with the deepfake detection features to form the input feature vector of the risk scoring model. The risk scoring model outputs a risk score based on the input feature vector, calculated using the following expression in one implementation: in, Risk score; For the Sigmoid function; This is the model weight vector; The input feature vector; This is a bias term. The liveness verification system sorts liveness authentication events based on risk scores and generates a verification task list. Each task in the verification task list is associated with an authentication event identifier and a risk score, so that it can be presented in the manual verification interface according to risk priority, thereby improving the efficiency of detecting high-risk deepfake attacks.

[0027] Before inputting the face image into the deep fake detection model, face region localization, face key point localization, face image alignment and size normalization are performed on the face image, and the aligned face image is used as the input of the deep fake detection model.

[0028] Before inputting the face image into the deep fake detection model, the liveness verification system preprocesses the face image to ensure that the input to the deep fake detection model meets uniform geometric and scale constraints. This reduces the impact of differences in camera angles, shooting distances, and facial poses from different terminals on the deep fake detection features. Face region localization is used to determine the boundaries of face regions in the original face image. Specifically, the liveness verification system calls the face detection algorithm to scan the face image, outputs the coordinates of the face bounding box, and expands the face bounding box outward based on a preset boundary expansion rule to cover the forehead, jaw, and facial contour areas. Subsequently, the face image is cropped according to the expanded face bounding box to obtain the face region image. Facial landmark localization is used to determine stable anchor points for geometric alignment in the face region image. Specifically, the liveness verification system calls the facial landmark detection algorithm to output the coordinates of the key points at the corners of the eyes, the tip of the nose, and the corners of the mouth, and performs confidence verification on the key point coordinates. Key points with confidence scores lower than a preset threshold are determined to be invalid. When invalid key points exist, the liveness verification system writes the corresponding authentication event identifier into the preprocessing anomaly record and directly uses the face region image as input to the deep fake detection model. At the same time, the preprocessing anomaly record is used as a risk factor in the subsequent risk scoring model to avoid insufficient verification coverage due to discarded samples.

[0029] In the face image alignment stage, the liveness verification system calculates the geometric transformation relationship based on the keypoint coordinates and the coordinates of a preset alignment template. The preset alignment template coordinates are the standard coordinates of the face keypoints consistent with the training input of the deep fake detection model. The liveness verification system establishes affine transformation parameters based on the keypoints of the eyes and the tip of the nose, and performs affine transformation on the face region image to obtain the aligned face image. The pixel resampling of the affine transformation uses bilinear interpolation to reduce the jaggedness and texture breaks introduced by the geometric transformation. Size normalization processing is used to convert the aligned face image to a preset input size. Specifically, the liveness verification system scales the aligned face image to a preset resolution and performs pixel normalization processing according to a preset pixel value range. Pixel normalization processing includes mapping pixel values ​​to the numerical range agreed upon by the deep fake detection model and performing channel order unification and data type conversion. Finally, the aligned face image is written into the model inference buffer as input to the deep fake detection model, establishing a one-to-one correspondence with the authentication event identifier, so that the subsequent output deep fake detection features can be traced back to the original authentication event dataset. Through the above preprocessing process, the input to the deepfake detection model remains consistent in terms of face position, scale, and rotation angle, reducing interference from non-fake factors caused by terminal differences and acquisition posture. This allows the deepfake detection features to more centrally represent forgery traces, improving the stability and repeatability of the review task list sorting.

[0030] The deep fake detection model includes a spatial domain feature extraction network and a frequency domain feature extraction network. The spatial domain feature extraction network extracts texture features from the face image, while the frequency domain feature extraction network performs frequency domain transformation on the face image and extracts fake trace features. The texture features and fake trace features are fused to obtain the deep fake detection features.

[0031] In one embodiment, the deepfake detection model is deployed in a server-side model inference service. The model inference service receives the aligned face image and outputs deepfake detection features. These features are then associated with authentication event identifiers and written into event data storage for the risk scoring model to read. The deepfake detection model employs a dual-branch structure: a spatial domain feature extraction network and a frequency domain feature extraction network. Both branches share the same aligned face image as input to ensure that the deepfake detection features can simultaneously reflect image texture anomalies and frequency domain forgery traces.

[0032] Spatial domain feature extraction networks are used to extract texture features from aligned face images. These networks consist of sequentially connected convolutional layers, normalization layers, and non-linear activation layers, forming multi-level feature representations through residual connections. Convolutional layers cover local regions to capture subtle differences in skin texture, edge transitions, and illumination discontinuities, while residual connections maintain the transitivity of low-level texture information across deeper layers. At the end, a global pooling layer is applied to compress the spatial dimension, outputting a fixed-length texture feature vector. This vector represents the fine-grained texture distribution of the aligned face image in the spatial domain.

[0033] A frequency domain feature extraction network is used to extract forgery traces from aligned face images. The network performs a frequency domain transformation on the aligned face image at the input to obtain a frequency domain representation. The frequency domain transformation employs either Discrete Fourier Transform or Discrete Cosine Transform. The frequency domain representation is then processed by taking the amplitude spectrum and performing amplitude normalization to suppress the influence of brightness differences between different devices on the frequency domain amplitude. The network then performs convolutional feature extraction on the frequency domain representation. The convolutional kernel covers frequency components at different scales to capture spectral anomalies introduced by interpolation, resampling, and compression in the generative model, outputting a fixed-length forgery trace feature vector. This forgery trace feature vector characterizes the inconsistency between the frequency domain distribution of the aligned face image and the real-world image.

[0034] To enable the combined use of the two types of features within the same feature space, the deep fake detection model performs feature fusion processing on texture feature vectors and forgery trace feature vectors. Feature fusion employs a combination of feature concatenation and fusion mapping. Feature concatenation connects the texture feature vector and the forgery trace feature vector dimensionally to form a fused input vector. Fusion mapping maps the fused input vector to deep fake detection features through a fully connected layer and performs normalization processing on the deep fake detection features to stabilize the input distribution of the subsequent risk scoring model. The deep fake detection features, as a unified representation of the aligned face image, are used as input data in subsequent steps to construct the risk scoring model, thus utilizing the complementary information of spatial texture anomalies and frequency domain forgery traces simultaneously during risk score calculation. Through this dual-branch and fusion structure, the deep fake detection model can output stable deep fake detection features when facing different types of attacks such as face swapping, synthesis, and replay, improving the consistency of the review task list and reducing the risk of missed detections due to relying solely on identity comparison results.

[0035] The input data for constructing the risk scoring model includes: terminal reuse frequency characteristics obtained by statistically analyzing the frequency of terminal device identifiers based on terminal device identifiers and historical authentication records; terminal consistency characteristics obtained by performing a matching consistency judgment based on terminal device identifiers and those in historical authentication records; network segment characteristics obtained by performing network segment division based on network address information; network affiliation characteristics obtained by performing operator attribution resolution based on network address information; location change frequency characteristics obtained by statistically analyzing the frequency of geographical location changes based on geographical location and historical authentication records; authentication frequency characteristics obtained by statistically analyzing the frequency of authentication occurrences based on historical authentication records; background region images obtained by performing face region localization based on historical authentication face images in historical authentication records; background feature vectors obtained by extracting features from the background region images; background cluster identifiers obtained by performing unsupervised clustering analysis on the background feature vectors; and background reuse frequency characteristics obtained by statistically analyzing the background cluster identifiers. Finally, the input feature vector is obtained by concatenating the terminal reuse frequency characteristics, terminal consistency characteristics, network segment characteristics, network affiliation characteristics, location change frequency characteristics, authentication frequency characteristics, background reuse frequency characteristics, and deep fake detection characteristics.

[0036] In one embodiment, after obtaining deep fake detection features, the liveness verification system further extracts behavioral statistics from the authentication event dataset, including terminal, network, geographic location, and historical authentication records. It also supplements this with unsupervised clustering analysis of the background of historically authenticated face images to obtain background reuse frequency features. These features, along with the deep fake detection features, construct the input feature vector for the risk scoring model. To ensure statistical consistency, the liveness verification system limits historical authentication records to a preset time window and uses the liveness authentication event identifier to associate and obtain the terminal device identifier, network address information, and geographic location corresponding to the current liveness authentication event.

[0037] The terminal reuse frequency feature is obtained by statistically analyzing the number of times the terminal device identifier appears in historical authentication records. The liveness verification system filters historical authentication records and counts them using the terminal device identifier as the key, and the counting result is written into the input feature vector. The terminal consistency feature is obtained by performing a consistency judgment between the terminal device identifier of the current liveness authentication event and the terminal device identifier in the historical authentication records. The consistency judgment outputs a consistency flag or an inconsistency flag, and the consistency flag is written into the input feature vector. The network segment feature is obtained by performing network segmentation on network address information. The liveness verification system performs normalized parsing on network address information and extracts the address prefix as the network segment identifier. The network affiliation feature is obtained by performing operator affiliation parsing on network address information. The operator affiliation parsing maps network address information to operator affiliation identifiers based on a pre-set network address affiliation database. The location change frequency feature is obtained by statistically analyzing the number of geographical location changes in historical authentication records. The liveness verification system converts geographical locations into area codes and accumulates statistics on area code changes of adjacent authentication records. The authentication frequency feature is obtained by statistically analyzing the number of authentication occurrences in historical authentication records. The liveness verification system counts historical authentication records within a preset time window and writes the count into the input feature vector.

[0038] To identify the attack pattern of black market operators reusing the same photo background and replacing different faces, the liveness verification system performs face region localization based on historically authenticated face images from historical authentication records to obtain the background region image. Face region localization can use face detection algorithms to output face bounding boxes. The liveness verification system then performs masking processing on the historically authenticated face images based on these bounding boxes. The masking process sets the area covered by the face bounding box to a preset pixel value and preserves the remaining area, thus obtaining the background region image. The background feature vector is obtained by feature extraction from the background region image. Feature extraction can be implemented using either a convolutional neural network feature extractor or an image hash feature extractor, outputting a fixed-length background feature vector. The liveness verification system performs unsupervised clustering analysis on the background feature vector to obtain background cluster labels. Unsupervised clustering analysis can use a distance-based clustering algorithm to divide the background feature vector into different clusters and output the corresponding background cluster labels. The background reuse frequency feature is obtained by counting the number of times the background appears based on the background cluster identifier. The liveness verification system counts the historical authentication records with the same background cluster identifier within a preset time window and writes them into the input feature vector.

[0039] In the input feature vector construction stage, the liveness verification system concatenates terminal reuse frequency features, terminal consistency features, network segment features, network affiliation features, location change frequency features, authentication frequency features, background reuse frequency features, and deepfake detection features to obtain the input feature vector. This input feature vector is then fed into the risk scoring model to output a risk score. By jointly incorporating background reuse patterns and deepfake detection features into the risk scoring, the system can improve the risk score differentiation in scenarios where face-swapping attacks reuse the same background template but replace it with different faces. This allows the verification task list to prioritize coverage of potential attack events and enhances the consistency and traceability of evidence in manual verification.

[0040] The risk scoring model is a supervised learning classification model, which is trained based on historical authentication samples labeled with review conclusions and outputs a risk score to the input feature vector.

[0041] In one embodiment, the risk scoring model is deployed in a model training service and a model inference service. The model training service trains a supervised learning classification model based on historical authentication samples labeled with review conclusions. The model inference service outputs risk scores from the input feature vectors for sorting in the review task list. Historical authentication samples are obtained by aggregating event data storage and review label storage. Each historical authentication sample includes at least an input feature vector and a review conclusion label. The review conclusion labels are generated by the manual review interface and written into the review label storage. To ensure the labels are clear, the review conclusion labels are labeled as either attack or non-attack. Attack corresponds to deep fake attacks or other violations, while non-attack corresponds to normal authentication events or false alarm events. The model training service performs supervised learning training on the samples based on the review conclusion labels.

[0042] Before training, the model training service performs data preparation processing on historical authentication samples. This data preparation includes: imputing missing tags in missing fields of the input feature vector, encoding conversion for discrete fields, range pruning and normalization for numerical fields, and deduplication for duplicate samples with the same authentication event identifier to ensure consistent training input. To reduce the impact of sample distribution skew on model judgment, the model training service performs sample balancing processing on attack-type and non-attack-type samples. Sample balancing processing includes oversampling attack-type samples or downsampling non-attack-type samples. After training, the threshold of the output risk score is calibrated based on validation samples to match the threshold with the business-side review resources. The supervised learning classification model can be implemented using one of the following methods: logistic regression, gradient boosting decision tree, or multilayer perceptron. The model type is written into the model configuration file as a model configuration and fixed upon deployment so that the training process can be reproduced during auditing and backtracking.

[0043] The model inference service receives the input feature vector from the risk scoring model and outputs a risk score. The risk score characterizes the degree of risk of an authentication event being an attack. The risk score can be calculated using the aforementioned risk score output expression or directly output by the selected supervised learning classification model. The model inference service writes the risk score and authentication event identifier back to the event data storage, and the task generation service sorts the authentication events based on the risk score to generate a review task list. To ensure the model continuously adapts to changes in attack methods, the model training service performs retraining at a preset training period or by triggering incremental review annotation samples. After training, the new model version is registered in the model version management storage, and the model parameters in the model inference service are replaced through a canary release. During the canary release, the difference in risk scores and review hit rates between the old and new models under the same input feature vector are recorded. When the difference exceeds a preset range, the model is rolled back to the previous version, thereby continuously improving the effectiveness of deepfake attack identification and review ranking without affecting business stability.

[0044] The process of generating a review task list includes: sorting liveness authentication events according to their risk scores, and generating a review task entry for each liveness authentication event that includes the risk score and the liveness authentication event identifier.

[0045] In one embodiment, after the liveness verification system outputs a risk score from the risk scoring model, the task generation service generates a list of verification tasks and pushes it to the manual verification interface. To ensure that verification resources prioritize high-risk events, the task generation service sorts liveness authentication events according to their risk scores, from highest to lowest. When multiple liveness authentication events have the same risk score or the difference is less than a preset threshold, the task generation service uses the authentication time as a secondary sorting key, prioritizing the output of newer liveness authentication events to promptly address ongoing attacks.

[0046] The review task entry serves as the smallest scheduling unit for the manual review interface. Each review task entry must include at least a liveness authentication event identifier and a risk score. The liveness authentication event identifier is a unique identifier generated by the identity authentication system and carried in the authentication event dataset. The task generation service reads the authentication event dataset corresponding to the liveness authentication event identifier from the event data storage based on the liveness authentication event identifier and writes the risk score and the liveness authentication event identifier into the review task entry. To ensure traceability and repeatability of the review, the task generation service records the generation time and task status for each review task entry. The task status includes at least three states: pending review, under review, and reviewed. The task status is updated by the manual review interface when the task is received and the review conclusion is submitted.

[0047] The review task list is an ordered set of review task items. This list is written to the review task queue, which supports retrieving tasks by risk score range and filtering tasks by status. The manual review interface requests a specified number of review task items from the queue upon loading and presents them in the order of the review task list. To reduce duplicate reviews, the task generation service performs task deduplication before writing review task items to the queue. Deduplication uses the liveness authentication event identifier as the unique key; if the same liveness authentication event identifier already exists in the queue and the task status is pending review or under review, it will not be written again. Through the above sorting, task item construction, and queue management mechanisms, the liveness review system can use risk scores as the core scheduling basis, directly converting the risk assessment results after fusing deepfake detection and terminal network behavior characteristics into executable review tasks. This achieves a closed-loop connection from model judgment to manual handling and improves the timeliness of handling high-risk events.

[0048] The manual review interface outputs a list of review tasks and receives review conclusions. In response to the review conclusions, it outputs risk control instructions to the business system and uses the review conclusions to update the risk scoring model.

[0049] In one embodiment, the manual review interface retrieves a list of review tasks from the review task queue and reads the corresponding authentication event dataset from the event data storage using the liveness authentication event identifier as an index. It then displays the reviewer a face image, identity comparison result, terminal device identifier, network address information, geographical location, and historically authenticated face images and results from historical authentication records. The reviewer selects a review conclusion based on the displayed information and submits it. The review conclusion is written to the review annotation storage and sent back to the task generation service to update the task status of the review task entry. The risk control handling service responds to the review conclusion by generating risk control handling instructions and sending them to the business system. These instructions include at least account restriction instructions and secondary authentication instructions. The business system executes account freezing, login blocking, or triggering secondary authentication based on the risk control handling instructions. The model training service periodically reads the review conclusion from the review annotation storage and merges the review conclusion with the corresponding input feature vector to form incremental training samples. These samples are used to retrain and update the risk scoring model, ensuring the model remains adaptable to new deepfake attacks and terminal network reuse patterns, thereby improving the effectiveness of the review task list sorting and the timeliness of risk control handling.

[0050] The manual review interface outputs a review task list, which includes: associated face images, historical authentication face images from historical authentication records, identity comparison results, terminal device identifiers, network address information, geographical location and risk scores for each review task item, and generates and outputs deep fake detection judgment results based on deep fake detection features.

[0051] In one embodiment, the manual review interface serves as the interaction terminal of the live review system, which is used to aggregate and display the key information corresponding to the review task entries with the same authentication event identifier as the main line, and receive the review conclusions submitted by the reviewers. After reading the review task entries from the review task queue, the manual review interface initiates a query to the event data storage based on the live authentication event identifier, reads the authentication event dataset corresponding to the live authentication event identifier, and associates and outputs the face image, identity comparison result, terminal device identifier, network address information, geographical location, historical authentication record, and risk score in the authentication event dataset. To reduce misjudgments caused by reviewers switching between multi-source data, the manual review interface displays the face image of the current live authentication and the historical authentication face image in the historical authentication record on the same page in parallel, and outputs the corresponding authentication result and authentication time beside the historical authentication face image, enabling reviewers to directly compare the facial detail differences and historical behavior consistency.

[0052] In terms of the output of the deepfake detection determination result, after the deepfake detection model outputs the deepfake detection features, the live review system calls the deepfake detection determination unit to generate the deepfake detection determination result for the deepfake detection features. The deepfake detection determination result includes the deepfake determination label and the deepfake confidence level. The deepfake determination label is used to represent whether there is a deepfake risk in the face image, and the deepfake confidence level is used to represent the credibility of the determination. The deepfake detection determination unit uses the classification head supporting the deepfake detection model to map the deepfake detection features, outputs the deepfake confidence level, and compares the deepfake confidence level with the threshold rule to generate the deepfake determination label. The threshold rule is provided by the model configuration file and is bound to the model version in the model version management storage. The manual review interface displays the deepfake determination label and the deepfake confidence level while showing the risk score, and clearly marks the review task entries with the deepfake determination label indicating a deepfake risk to prompt reviewers to prioritize the verification of details such as the facial edge transition, skin texture continuity, and lighting consistency.

[0053] Through the above-mentioned associated output and determination result presentation method, the manual review interface aligns the determination information of the deepfake detection model with the terminal network behavior elements and historical authentication comparison information in the same review process, enabling reviewers to submit review conclusions based on a unified evidence chain, reducing missed detections caused by relying solely on identity comparison results, and improving the handling efficiency and traceability of high-risk live authentication events.

[0054] Using the review conclusion to update the risk scoring model includes: writing the review conclusion and the input feature vector of the risk scoring model into the training sample set, and performing retraining on the risk scoring model based on the training sample set to update the risk scoring model parameters.

[0055] In one embodiment, the liveness verification system feeds back the verification conclusions to the risk scoring model to form a closed-loop update. After a reviewer submits their verification conclusion, the manual verification interface writes the conclusion and a liveness authentication event identifier into the verification annotation storage. The verification annotation storage and the event data storage are linked through the liveness authentication event identifier. The model training service periodically scans the verification annotation storage for confirmed verification conclusions and, based on the liveness authentication event identifier, reads the risk scoring model input feature vector corresponding to that liveness authentication event from the event data storage. It then combines the verification conclusion and the input feature vector and writes it into the training sample set. The training sample set is stored using an append-only method. Each training sample includes at least the input feature vector, verification conclusion label, and generation time, and records the model version identifier to support traceability of the training process.

[0056] To ensure the availability of the training sample set, the model training service performs a sample consistency check before writing samples into the training sample set. This check includes verifying that the input feature vector field structure is consistent with the current model configuration file and verifying that the review conclusion labels belong to a preset label set. If the check fails, a liveness detection event identifier is written to the sample anomaly record, and the model does not proceed to the retraining process. The model training service retrains the risk scoring model when retraining trigger conditions are met. These trigger conditions include reaching a preset training period or the number of newly written samples in the training sample set reaching a preset threshold. The retraining process uses a combination of historical and incremental samples from the training sample set for training. After training, a new model version is generated and written to the model version management storage, while simultaneously recording the training sample set range and training time corresponding to the new model version.

[0057] When a new model version is released, the model inference service loads the parameters of the new model version and enters the gray-scale verification phase. During the gray-scale verification phase, the risk score of the new model version and the risk score of the old model version are calculated simultaneously for the same input feature vector, and the difference in risk scores and the hit rate of the verification are recorded. When the gray-scale verification phase meets the preset stability conditions, the new model version is switched to as the default model version; otherwise, it rolls back to the old model version. By writing the verification conclusions and the input feature vector of the risk scoring model into the training sample set and performing retraining to update the risk scoring model parameters, the risk scoring model can continuously absorb the latest judgment results from manual verification, ensuring that the risk score remains consistent with the actual attack pattern. This improves the targeting of the verification task list and reduces the risk of model drift during long-term operation.

[0058] like Figure 2 As shown, a liveness verification device based on a deepfake detection model is used to implement the liveness verification method based on the deepfake detection model. The system includes: The authentication data retrieval module is used to acquire the authentication event dataset corresponding to the liveness authentication event. The authentication event dataset includes face images, identity comparison results, terminal device identifiers, network address information, geographical location, and historical authentication records. Historical authentication records include face images and authentication results of historical authentications. The authentication data retrieval module is a hardware execution unit for data acquisition and aggregation, integrating a network interface controller, a system bus interface, and a storage controller. It is used to pull data from the identity authentication system, the terminal information acquisition system, and the server or database node corresponding to the historical authentication storage and complete the aggregation. The authentication data retrieval module receives face image data streams and identity comparison result data frames through the network interface controller, writes the terminal device identifier, network address information, and geographical location into the memory buffer through the system bus, and reads historical authentication records from external storage or distributed storage nodes into the local cache by the storage controller, thereby forming the authentication event dataset corresponding to the liveness authentication event and maintaining a consistent association with the liveness authentication event identifier.

[0059] The deepfake detection module is used to input face images into the deepfake detection model to obtain deepfake detection features. Based on the authentication event dataset and the deepfake detection features, it constructs input data for the risk scoring model and outputs risk scores. Based on the risk scores, it generates a review task list. The deepfake detection module is a hardware computing unit for model inference and score calculation, including a central processing unit, a graphics processing unit or neural network accelerator, and high-speed memory. It is used to perform deepfake detection model inference on face images and generate deepfake detection features. The deepfake detection module performs feature concatenation and tensor encapsulation of the structured fields of the authentication event dataset and the deepfake detection features in memory, and the computing unit performs risk scoring model inference to output risk scores. The deepfake detection module further sorts the risk scores and assembles task items on the processor to generate a review task list and writes it to the task queue for storage to ensure the schedulability and traceability of the review tasks.

[0060] The review and handling module is used to output a review task list and receive review conclusions on the manual review interface. In response to the review conclusions, it outputs risk control handling instructions to the business system and uses the review conclusions to update the risk scoring model. This module is a hardware execution unit for human-computer interaction and business linkage, integrating a display controller, an input interface controller, and a network communication interface. It outputs a review task list and receives review conclusions on the manual review interface. The review and handling module drives the display terminal to present facial images, historical authentication records, and risk scores associated with the review task items through the display controller. It receives the review conclusions from reviewers through the input interface controller and writes them to the local cache and backend storage. The review and handling module sends risk control handling instructions to the business system through the network communication interface and writes the review conclusions and corresponding input feature vectors into the training sample set storage, triggering the model training server to retrain and update the risk scoring model, thus forming a closed-loop update mechanism for deepfake attacks.

[0061] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects.

[0062] The above are merely embodiments of the present invention and are not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the present invention should be included within the scope of the claims of the present invention.

Claims

1. A liveness review method based on a deep pseudo-detection model, characterized in that, Includes the following steps: Obtain the authentication event dataset corresponding to the liveness authentication event. The authentication event dataset includes face images, identity comparison results, terminal device identifiers, network address information, geographical location, and historical authentication records. The historical authentication records include face images and authentication results of historical authentications. The face image is input into the deepfake detection model to obtain deepfake detection features. Based on the authentication event dataset and the deepfake detection features, the input data of the risk scoring model is constructed and the risk score is output. Based on the risk score, a review task list is generated. The manual review interface outputs a list of review tasks and receives review conclusions. In response to the review conclusions, it outputs risk control instructions to the business system and uses the review conclusions to update the risk scoring model.

2. The method of claim 1, wherein, Obtaining the authentication event dataset includes: obtaining face images and identity comparison results from the identity authentication system, obtaining terminal device identifiers, network address information and geographical location from the terminal information collection system, obtaining historical authentication records from historical authentication storage, and performing field integrity verification, format conversion and deduplication merging processing on the authentication event dataset.

3. The method according to claim 1, characterized in that, Before inputting the face image into the deep fake detection model, face region localization, face key point localization, face image alignment and size normalization are performed on the face image, and the aligned face image is used as the input of the deep fake detection model.

4. The method according to claim 3, characterized in that, The deep fake detection model includes a spatial domain feature extraction network and a frequency domain feature extraction network. The spatial domain feature extraction network extracts texture features from the face image, while the frequency domain feature extraction network performs frequency domain transformation on the face image and extracts fake trace features. The texture features and fake trace features are fused to obtain the deep fake detection features.

5. The method according to claim 1, characterized in that, The input data for constructing the risk scoring model includes: terminal reuse frequency characteristics obtained by statistically analyzing the frequency of terminal device identifiers based on terminal device identifiers and historical authentication records; terminal consistency characteristics obtained by performing a matching consistency judgment based on terminal device identifiers and those in historical authentication records; network segment characteristics obtained by performing network segment division based on network address information; network affiliation characteristics obtained by performing operator attribution resolution based on network address information; location change frequency characteristics obtained by statistically analyzing the frequency of geographical location changes based on geographical location and historical authentication records; authentication frequency characteristics obtained by statistically analyzing the frequency of authentication occurrences based on historical authentication records; background region images obtained by performing face region localization based on historical authentication face images in historical authentication records; background feature vectors obtained by extracting features from the background region images; background cluster identifiers obtained by performing unsupervised clustering analysis on the background feature vectors; and background reuse frequency characteristics obtained by statistically analyzing the background cluster identifiers. Finally, the input feature vector is obtained by concatenating the terminal reuse frequency characteristics, terminal consistency characteristics, network segment characteristics, network affiliation characteristics, location change frequency characteristics, authentication frequency characteristics, background reuse frequency characteristics, and deep fake detection characteristics.

6. The method according to claim 5, characterized in that, The risk scoring model is a supervised learning classification model, which is trained based on historical authentication samples labeled with review conclusions and outputs a risk score to the input feature vector.

7. The method according to claim 1, characterized in that, The process of generating a review task list includes: sorting liveness authentication events according to their risk scores, and generating a review task entry for each liveness authentication event that includes the risk score and the liveness authentication event identifier.

8. The method according to claim 7, characterized in that, The manual review interface outputs a review task list, which includes: associated face images, historical authentication face images from historical authentication records, identity comparison results, terminal device identifiers, network address information, geographical location and risk scores for each review task item, and generates and outputs deep fake detection judgment results based on deep fake detection features.

9. The method according to claim 1, characterized in that, Using the review conclusions to update the risk scoring model includes: writing the review conclusions and the input feature vectors of the risk scoring model into the training sample set, and retraining the risk scoring model based on the training sample set to update the risk scoring model parameters.

10. A liveness verification device based on a deepfake detection model, used to implement the liveness verification method based on a deepfake detection model as described in any one of claims 1-9, characterized in that, The system includes: The authentication data retrieval module is used to obtain the authentication event dataset corresponding to the liveness authentication event. The authentication event dataset includes face images, identity comparison results, terminal device identifiers, network address information, geographical location and historical authentication records. The historical authentication records include face images and authentication results of historical authentications. The deep fake detection scoring module is used to input face images into the deep fake detection model to obtain deep fake detection features, construct input data for the risk scoring model based on the authentication event dataset and deep fake detection features, and output risk scores. Based on the risk scores, a review task list is generated. The review and handling module is used to output a list of review tasks and receive review conclusions on the manual review interface, output risk control handling instructions to the business system in response to the review conclusions, and use the review conclusions to update the risk scoring model.

Images