A method and system for verifying authenticity based on a cloned virtual image

Abstract

The application provides a kind of based on clone virtual image authenticity verification method and system, comprising: from virtual image data, initial texture and color feature are extracted by convolutional neural network, and dimension reduction is carried out using principal component analysis, to generate compressed feature vector;Then, the system discriminates the type of virtual image according to the vector, and enables type adaptive verification path: for cartoon image, construct contour topology and carry out structure consistency matching;For realistic portrait, then analyze micro-expression sequence through space-time attention network to capture dynamic artifacts.Subsequently, using conditional random field model, the preliminary judgment result and the type adaptive verification feature are fused to generate the final authenticity judgment.The application also includes: the image identification and feature fingerprint that pass through verification are recorded to block chain, and chain data is periodically analyzed, and the verification model is adaptively optimized through genetic algorithm, to form a closed-loop self-evolution system.

Description

Technical Field

[0001] This invention relates to the field of information technology, and in particular to a method and system for verifying the authenticity of cloned virtual images. Background Technology

[0002] With the rise of concepts such as metaverse, virtual social interaction, and digital entertainment, virtual avatars, as users' representations in the digital world, are increasingly widely used, encompassing multiple fields including virtual social interaction, digital entertainment, and identity authentication. Their authenticity and credibility directly impact user experience, data security, and platform trust, making them a crucial and indispensable element in the digital economy. The authenticity, uniqueness, and credibility of virtual avatars directly affect user experience, data security, and even the stability of the entire digital economic ecosystem. However, with the development of avatar synthesis technology, highly accurate cloned virtual avatars are emerging in large numbers. These avatars not only closely resemble the original avatar in appearance but can even mimic specific actions and expressions, posing significant challenges to identity authentication, copyright protection, and platform supervision.

[0003] The limitations of existing virtual avatar verification methods are becoming increasingly apparent. These limitations primarily lie in the insufficient comprehensiveness and accuracy of extracting complex avatar features, especially when faced with high-precision cloned avatars, making it difficult to effectively distinguish between the original and the counterfeit. Furthermore, existing methods lack targeted verification strategies for different types of virtual avatars, resulting in insufficient verification efficiency and accuracy, and failing to meet the needs of diverse scenarios.

[0004] In the process of verifying the authenticity of virtual avatars, the core challenges lie in how to accurately capture and compare the unique features of the avatar, and how to formulate differentiated verification rules based on the type of avatar. First, the pixel-level features of virtual avatars are complex and varied, containing high-dimensional information such as texture, color distribution, and micro-expressions. Existing technologies often lose key details when extracting these features, causing verification systems to be unable to accurately distinguish highly similar clones. For example, in virtual social platforms, a highly realistic clone might be mistaken for the original avatar through simple feature comparison, leading to the risk of identity theft. Second, because different types of virtual avatars (such as cartoon characters, realistic portraits, and animal characters) differ significantly in feature expression and application scenarios, a unified verification standard is difficult to adapt to diverse needs. For example, cartoon characters may focus more on outline features, while realistic portraits rely on facial details. The lack of a classification system can lead to biased verification results, affecting the reliability of subsequent avatar management and traceability.

[0005] Therefore, how to achieve efficient and accurate identification of the authenticity of virtual images through high-precision feature extraction technology and differentiated verification strategies, and provide reliable data support for the construction of image family tree chains, has become a key problem that this study urgently needs to solve. Summary of the Invention

[0006] The purpose of this application is to provide a method and system for verifying the authenticity of cloned virtual images, aiming to solve the problems of existing technologies such as single verification strategies, insufficient accuracy, untraceable results, and inability of the system to self-optimize.

[0007] To achieve the above objectives, this application provides the following technical solution:

[0008] On the one hand, this application provides a method for verifying the authenticity of cloned virtual images, including the following steps:

[0009] S1: Obtain the image data of the virtual avatar to be verified, and extract its initial feature representation through a pre-defined feature extraction network, such as a convolutional neural network (CNN). This initial feature representation may include a first texture feature set and a first color distribution feature set extracted from the image data, and a compressed feature vector formed after processing by dimensionality reduction methods such as principal component analysis (PCA).

[0010] S2: Based on the initial feature expression described in S1, a lightweight classifier is used to determine the type of the virtual image to identify it as one of at least two preset image types, such as "cartoon image" or "realistic portrait".

[0011] S3: Based on the type discrimination result of S2, adaptively select the refined verification module that best matches the image type for processing to extract the most distinctive key verification features. This step specifically involves:

[0012] If the image is identified as a "cartoon character," the contour feature verification module is invoked. This module further extracts a subset of contour features from the initial feature representation and generates a contour map using edge detection algorithms such as Canny. Next, key feature points such as corners and endpoints are identified on the contour map, and a contour topology graph is constructed based on the spatial connections between these points. This graph uses key points as nodes and the attributes of connecting line segments (such as length, curvature, and relative angle) as edges, thus abstracting the image's contour information into structured graph data.

[0013] If the image is identified as a "realistic portrait," the micro-expression feature verification module is invoked. This module processes the image data (especially consecutive frames of the facial region) into a temporal image sequence and inputs it into a pre-defined spatiotemporal neural network model that incorporates an attention mechanism (such as a combination of 3D-CNN or CNN-LSTM with Attention). This model can automatically learn and focus on key areas of the face that are most likely to expose forgery (such as the corners of the eyes and mouth) and key time points, thereby extracting a micro-expression feature sequence that deeply integrates spatial texture and temporal dynamic information.

[0014] S4: The initial feature representation obtained in S1 (or the preliminary true / false probability generated from it) is fused with the key verification features obtained in S3 (the outline topology of the cartoon character or the micro-expression sequence of a realistic portrait). This step employs a multimodal fusion model, preferably a Conditional Random Field (CRF) model. Using the aforementioned features as the observation sequence in the CRF model, the model calculates the posterior probability distribution of the true state ("true" or "false") of the image given the observations, based on the learned feature functions and transition probabilities. Finally, the state with the highest posterior probability is selected as the final true / false judgment result.

[0015] Furthermore, in order to establish a reliable traceability system and achieve system self-evolution, this method also includes:

[0016] S5: If the final judgment result of S4 is "real", the system generates a unique verification log for the virtual avatar. This log contains the avatar's unique identifier (such as user ID + avatar ID), a verification timestamp, and a feature fingerprint generated from its core features (such as the structural hash value of the outline topology or a compact representation of the micro-expression feature sequence). Subsequently, through a smart contract calling the blockchain interface, this verification log is recorded as a new transaction on a dedicated avatar genealogy chain, thereby achieving secure, permanent, and tamper-proof storage of the verification result.

[0017] S6: The system has a built-in self-optimizing scheduler that periodically (e.g., every morning at midnight) securely retrieves recent historical verification log data from the character genealogy chain. An analysis engine statistically analyzes this data and calculates key performance indicators (KPIs), such as average verification time for different character types, high-confidence rejection rate, and fuzzy decision rate. If a certain indicator (such as the fuzzy decision rate for cartoon characters) is found to be consistently higher than a preset threshold, it indicates that the current model's defense capability against a certain type of attack has decreased. At this time, the system will start a self-optimizing algorithm (such as a genetic algorithm), encoding the hyperparameters of the model to be optimized (such as the kernel size and learning rate of a CNN, or the feature weights of a CRF) as "genes," generating a new candidate model population through crossover, mutation, and other operations, and iteratively optimizing using the improvement of this performance indicator as the fitness function. The optimal parameter combination will be used to update the online model, thus completing a closed-loop self-evolution.

[0018] On the other hand, this application provides a method for verifying the authenticity of cloned virtual images, including the following steps:

[0019] Step 1: Obtain the virtual avatar data to be verified and perform preprocessing including size normalization and pixel value standardization.

[0020] Step 2: Analyze the preprocessed virtual character data using a preset character style classification model to determine its style type; the style type includes at least cartoon style and realistic style.

[0021] Step 3: Based on the style type, initiate the corresponding differential feature extraction pipeline, while simultaneously extracting common appearance features:

[0022] Extract the general appearance features of the virtual image, including a first texture feature set and a first color distribution feature set;

[0023] If the style type is cartoon style, then start the cartoon character feature extraction pipeline to extract its structured contour features;

[0024] If the style type is realistic, then the realistic portrait feature extraction pipeline is activated to extract its spatiotemporal dynamic micro-expression features.

[0025] Step 4: Use a Conditional Random Field (CRF) model to integrate the general appearance features and differential features (structured contour features or spatiotemporal dynamic micro-expression features) into a multimodal feature set to generate the final authenticity judgment result; wherein, the node potential function of the CRF model is constructed based on the confidence of each feature stream.

[0026] Step 5: Based on the final authenticity judgment result, if it is determined to be authentic, the identity identifier and characteristic summary of the virtual image will be recorded in the preset image genealogy blockchain through a smart contract, forming an immutable and traceable verification log.

[0027] Thirdly, the present invention provides a system for verifying the authenticity of cloned virtual images, comprising:

[0028] The feature extraction module is used to acquire the image data of the virtual image to be verified and extract its initial feature expression through a preset feature extraction network.

[0029] The type discrimination module is used to perform type discrimination on the virtual image based on the initial feature expression, so as to determine that it is one of at least two preset image types;

[0030] The refined verification module is used to select the module corresponding to the image type to process the initial feature expression based on the type discrimination result, so as to obtain key verification features; wherein, the refined verification module includes a contour feature verification module for cartoon images and a micro-expression feature verification module for realistic portraits;

[0031] The result integration module is used to fuse the initial feature representation and the key verification features, and generate the final true / false judgment result through a multimodal fusion model.

[0032] Fourthly, the present invention provides a system for verifying the authenticity of cloned virtual avatars, characterized in that it includes:

[0033] The data preprocessing module is used to acquire and preprocess the virtual avatar data to be verified;

[0034] A style classification module is used to determine the style type of the virtual avatar data, wherein the style type includes at least cartoon style and realistic style;

[0035] The differential feature extraction module is used to extract structured contour features or spatiotemporal dynamic micro-expression features based on the style type.

[0036] The multimodal fusion module is used to fuse common appearance features with the differentiated features and generate the final authenticity judgment result.

[0037] This invention discloses a method and system for identifying the authenticity of virtual avatars based on multimodal feature fusion. Addressing the business scenario problem of difficulty in distinguishing between genuine and fake cartoon avatars and realistic human portraits in virtual avatar data, it achieves accurate identification through deep feature extraction and multimodal integration. The invention extracts initial texture and color features from virtual avatar data using a convolutional neural network and performs dimensionality reduction using principal component analysis to generate compressed feature vectors. Then, the system uses these vectors to determine the type of the virtual avatar and activates a type-adaptive verification path: for cartoon avatars, it constructs a contour topology map and performs structural consistency matching; for realistic human portraits, it analyzes micro-expression sequences using a spatiotemporal attention network to capture dynamic artifacts. Subsequently, a conditional random field model is used to fuse the preliminary judgment results with the type-adaptive verification features to generate the final authenticity judgment. This application also includes: recording the verified avatar identifier and feature fingerprint to a blockchain to construct a traceable avatar family chain; and periodically analyzing the on-chain data and adaptively optimizing the verification model using a genetic algorithm to form a closed-loop self-evolving system. This invention significantly improves the accuracy and robustness of identifying the authenticity of virtual characters through deep feature extraction and multimodal fusion, and is applicable to scenarios such as virtual reality and digital entertainment.

[0038] Compared to existing technologies, this invention employs a type-adaptive verification framework, utilizing optimal feature extraction and comparison strategies (contour topology vs. spatiotemporal micro-expressions) for different virtual avatars. This deeply uncovers their core identity features, significantly outperforming traditional "one-size-fits-all" methods. It effectively resists high-precision cloning attacks, improving verification accuracy and robustness. The innovative introduction of blockchain technology transforms each verification result into an immutable link in the avatar's family tree chain. This not only provides users with reliable identity verification but also offers robust technical support for the platform to handle identity disputes and trace copyrights, constructing a trustworthy identity traceability system. By mining historical verification data on the blockchain, a closed-loop adaptive mechanism of "verification-feedback-optimization" is established. This allows the system to learn from past experience, dynamically adjusting the model to cope with unknown new attacks, achieving self-evolution and continuous learning of the verification system, and ensuring the long-term advancement and effectiveness of the technical solution. Attached Figure Description

[0039] Figure 1 This is an overall flowchart of a method for verifying the authenticity of a cloned virtual image provided in an embodiment of this application.

[0040] Figure 2 This is a schematic diagram of the differential feature extraction and fusion framework in the embodiments of this application.

[0041] Figure 3 This is a schematic diagram of a graph neural network (GNN) used for structured contour feature extraction in a cartoon style in an embodiment of this application.

[0042] Figure 4 This is a schematic diagram of the spatiotemporal dynamic micro-expression feature extraction network structure used in the embodiments of this application for realistic style.

[0043] Figure 5 This is a structural block diagram of a system for verifying the authenticity of a cloned virtual image, as provided in an embodiment of this application.

[0044] Figure 6 This is a schematic diagram comparing the key performance indicators of the method in the embodiments of this application with those of traditional methods.

[0045] Figure 7 This is an overall flowchart of a method for verifying the authenticity of a cloned virtual image, provided in another embodiment of this application.

[0046] Figure 8 This is a schematic diagram of a differential feature extraction and fusion framework in another embodiment of this application.

[0047] Figure 9 This is a schematic diagram of a graph neural network (GNN) for structured contour feature extraction in a cartoon style, as described in another embodiment of this application.

[0048] Figure 10 This is a schematic diagram of a spatiotemporal dynamic micro-expression feature extraction network structure for realistic style in another embodiment of this application.

[0049] Figure 11 This is a flowchart of a blockchain-based system dynamic optimization closed-loop process in another embodiment of this application.

[0050] Figure 12 This is a structural block diagram of a system for verifying the authenticity of a cloned virtual image, provided in another embodiment of this application.

[0051] Figure 13 This is a micro-expression analysis diagram of a spatiotemporal attention network for a realistic human face provided in Embodiment 1 of this application. Detailed Implementation

[0052] The technical solutions of the embodiments of the present invention will be clearly and thoroughly described below with reference to the accompanying drawings. The described embodiments are merely some embodiments of the present invention.

[0053] Currently, verification technologies for virtual avatars have several limitations. On the one hand, most existing methods employ uniform, standardized feature comparison strategies, such as directly comparing pixel differences or the distance of common features (e.g., color histograms, texture descriptors). This "one-size-fits-all" approach ignores the essential differences in core features between different types of virtual avatars (e.g., the clean lines of a cartoon character versus the rich detail of a realistic portrait). The identity of a cartoon character lies in its stable outline structure, while the realism of a realistic portrait is reflected in its delicate skin texture and vivid micro-expression changes. A uniform verification standard is difficult to simultaneously address, resulting in low recognition rates for specific types of cloning attacks and a high likelihood of false positives and false negatives.

[0054] On the other hand, existing verification processes are typically isolated, one-off events. Verification results are held solely by the platform, lacking a public, transparent, and tamper-proof third-party verification mechanism, making it difficult to build trust among users. When identity theft disputes arise, an effective and traceable chain of evidence cannot be provided. Furthermore, verification systems are usually static; their model parameters, once set, remain unchanged, unable to learn and evolve from massive daily verification practices. This makes them ill-suited to adapt to emerging cloning technologies, and the system's robustness and sophistication decline over time. Therefore, how to efficiently and accurately identify the authenticity of virtual avatars and build a trustworthy, traceable, and self-improving verification ecosystem is a critical technical problem urgently needing to be solved in this field.

[0055] Example 1

[0056] This embodiment describes in detail... Figure 1The complete execution flow of the method for verifying the authenticity of a cloned virtual avatar is shown. Suppose a virtual social platform needs to verify whether a newly uploaded virtual avatar "Avatar_Clone," which a user claims to be their original avatar, is a clone of another well-known avatar "Avatar_Original" on the platform.

[0057] Step S101: Obtain image data and extract initial features

[0058] First, the system retrieves the avatar data for "Avatar_Clone". If the avatar is a static image, the data is a 2D image, such as a PNG or JPG format, with a size of 1024x1024 pixels. If the avatar is animated, a 5-second video clip at 30fps is retrieved.

[0059] The system calls a pre-trained convolutional neural network (CNN), whose structure can be similar to ResNet-50 but optimized for texture and color features.

[0060] Texture feature extraction: The input image (or each frame of the video) is fed into the first few convolutional layers of the CNN to extract a 2048-dimensional texture feature vector. These features capture the details, lighting, and material information of the image's surface.

[0061] Color feature extraction: The input image is quantized in color, for example, using the K-Means algorithm (K=16) to cluster its dominant color into 16 center colors, and the RGB values ​​of these 16 center colors are then extracted. The 16 dimensions and their proportion in the image are combined to form a 64-dimensional color distribution feature vector.

[0062] Dimensionality Reduction and Fusion: The 2048-dimensional texture vector and the 64-dimensional color vector are concatenated into a 2112-dimensional initial vector. To improve efficiency and robustness, Principal Component Analysis (PCA) is used to reduce the dimensionality of this vector, retaining 99% of the variance, ultimately generating a 256-dimensional compressed feature vector as the initial feature representation of the image. This vector is both compact and retains most of the identity information.

[0063] Specifically, when extracting pixel-level features from virtual avatar data, image processing algorithms can be used to decompose the input RGB image (with a resolution of 256x256) into pixels and extract the color value (R, G, B components, range 0-255) of each pixel.

[0064] For example, using the OpenCV library's `cv2.imread()` function to read an image generates a three-dimensional array (256, 256, 3), where each pixel contains R, G, and B values, such as (123, 45, 67). To ensure data consistency, the image needs to be normalized, scaling the pixel values ​​to [0, 1] using the formula x' = x / 255, resulting in a normalized pixel-level feature matrix. Next, a convolutional neural network (CNN) is used to extract the first texture feature set. A network with three convolutional layers can be designed, with a kernel size of 3x3, a stride of 1, same padding, and ReLU activation. The first layer has 32 kernels, outputting a feature map of 256x256x32, capturing edge and texture information. The second and third layers use 64 and 128 kernels respectively, progressively extracting more complex texture patterns. The pooling layer uses 2x2 max pooling with a stride of 2 to reduce the feature map size to 64x64x128 while preserving significant texture features. The analysis process evaluates the saliency of texture features by calculating the L2 norm of the feature map; feature maps with an L2 norm > 0.5 are considered valid texture regions. Next, the first color distribution feature set is extracted. This can be done by calculating the color histogram of the normalized RGB image, setting 256 bins, and statistically analyzing the pixel distribution of each channel to obtain a probability density vector. For example, the R channel distribution is [0.02, 0.03, ..., 0.01]. To enhance the representation of color features, K-means clustering (K=8) is used to cluster the color distribution, generating 8 representative color clusters with center points such as (0.48, 0.22, 0.15). The initial feature representation is formed by concatenating texture features (flattened to 64x64x128=524288 dimensions) and color features (8x3=24 dimensions), resulting in a 524312-dimensional feature vector. To reduce dimensionality, Principal Component Analysis (PCA) is used, retaining 95% of the variance, reducing the dimensionality to approximately 1000 dimensions, thus forming the final initial feature representation. Logically, pixel-level features provide raw data for subsequent extraction, CNN texture extraction focuses on spatial structure, color distribution clustering highlights color patterns, and PCA dimensionality reduction ensures computational efficiency. Together, they constitute the feature representation of the virtual avatar, suitable for subsequent classification or generation tasks.

[0065] Step S102: Virtual Avatar Type Determination

[0066] To enable subsequent adaptive verification, the system needs to determine the type of "Avatar_Clone". This step uses a lightweight Support Vector Machine (SVM) classifier. This classifier is trained using a large number of labeled initial feature representations of cartoon and realistic images. The 256-dimensional feature vector generated in step S101 is input into this SVM, and the classifier outputs the discrimination result. Let's assume the output is "Realistic Human Portrait".

[0067] Step S103B: Call the micro-expression feature verification module to process the initial feature expression and obtain key verification features.

[0068] Since the judgment result is "realistic portrait", the system calls the micro-expression feature verification module, such as Figure 2 The path on the right is shown.

[0069] 1. Data preparation: Decode the acquired 5-second video clip into a 150-frame image sequence.

[0070] 2. Model Processing: Input the image sequence into... Figure 3 The spatiotemporal attention network shown. Figure 3 This is a schematic diagram of a graph neural network (GNN) used for structured contour feature extraction in a cartoon style, as described in this application embodiment. Preferably, a spatiotemporal attention network is employed. As shown, the input is a temporal image sequence, i.e., consecutive frames in a video. Each frame in the sequence is first processed by a convolutional neural network (CNN, such as MobileNetV2) with shared weights. This CNN is responsible for extracting the spatial features of each frame, generating a spatial feature sequence. Subsequently, this feature sequence is fed into a spatiotemporal attention module, which includes parallel spatial and temporal attention modules. The spatial attention module analyzes the feature map of each frame, assigning higher weights to key areas such as facial muscles; the temporal attention module analyzes the entire sequence, assigning higher weights to dynamic frames where expressions change. After attention weighting, the resulting weighted feature sequence is fed into a long short-term memory network (LSTM) for temporal modeling to capture deep patterns of feature evolution over time. Finally, the final hidden state of the LSTM passes through a fully connected layer, outputting a highly condensed micro-expression feature sequence vector. In this process, each frame in the CNN encoder sequence has its spatial feature map extracted using a shared CNN (e.g., MobileNetV2, to ensure speed). LSTM modeling: The attention-weighted feature sequence is fed into a Long Short-Term Memory (LSTM) network. The LSTM network can capture the patterns of feature evolution over time, thus modeling the entire dynamic process.

[0071] Figure 4 This is a schematic diagram of the spatiotemporal dynamic micro-expression feature extraction network structure used in this application embodiment. The diagram details the process of transforming the two-dimensional outline information of a cartoon character into structured data that a GNN can process. First, the outline is extracted from the original image using an edge detection algorithm. Then, in the graph construction stage, key points on the outline, such as endpoints, intersections, or high-curvature points, are abstracted as nodes in the graph. Figure 4 As shown on pages P1 to P7. The outline segments connecting these nodes are abstracted as edges of the graph, such as... Figure 4 As shown in e1 to e8. Importantly, each node and edge carries rich attributes. For example, the attributes of a node P1 may include its coordinate position (x, y) and degree (i.e., the number of edges it connects to). The attributes of an edge e1 may include its length L1, mean curvature C1, and angle A1 relative to the horizontal axis. The resulting topology graph is then fed into a graph neural network (such as a GIN network), which learns node attributes and graph adjacency relationships to output a feature vector that characterizes the overall skeletal structure of the image, i.e., structured contour features.

[0072] Spatiotemporal attention: After obtaining a sequence of 150 feature maps, the spatial attention mechanism analyzes each feature map, assigning higher weights to facial muscle regions (such as the orbicularis oculi and orbicularis oris muscles); simultaneously, the temporal attention mechanism analyzes the entire sequence, assigning higher weights to frames where subtle changes in expression occur (e.g., an unnatural blink or a twitch of the mouth), resulting in the following final effect: Figure 13 As shown.

[0073] 3. Feature Generation: The last hidden state of the LSTM is extracted, passed through a fully connected layer, and finally outputs a 512-dimensional micro-expression feature sequence. This sequence highly condenses all suspicious artifacts in the image during the dynamic process. The system calculates the cosine similarity between this sequence and the pre-stored baseline micro-expression sequence "Avatar_Original". Assuming the calculated similarity is 0.75, this value itself is a key verification feature.

[0074] Step S104: Multimodal feature fusion (CRF) to generate the final true / false judgment result.

[0075] Now, the system has information from multiple dimensions:

[0076] Preliminary assessment: Calculate the Euclidean distance between the multi-dimensional information and the pre-stored reference vector of Avatar_Original, and then convert this distance value into a probability value within a certain range using a pre-defined mapping function (such as the Sigmoid function). For example, calculating the Euclidean distance between the initial feature representation (256 dimensions) of "Avatar_Clone" and the initial feature representation of "Avatar_Original" can be mapped to a preliminary true / false probability, such as P(true) = 0.6.

[0077] Key validation feature: Similarity score of 0.75 obtained from the micro-expression module.

[0078] The system uses a Conditional Random Field (CRF) model to fuse this information. The CRF is modeled as a linear chain containing two state nodes (true and false). The observed data includes preliminary probabilities and micro-expression similarity. Using a pre-trained CRF model, the posterior probabilities of the state being "true" and the state being "false" under the current observation are calculated. Due to the low similarity of micro-expressions (0.75, while the similarity of real-life images is typically above 0.95), the CRF model ultimately outputs P(true|observation) = 0.2 and P(false|observation) = 0.8.

[0079] Specifically, the implementation method for generating the final true / false judgment result by integrating multimodal features using a Conditional Random Field (CRF) model based on the preliminary judgment result and contour feature map or micro-expression feature sequence is as follows: First, assume that the preliminary true / false result comes from image analysis of video frames, including facial contour and micro-expression data. The preliminary true / false result is extracted through a Convolutional Neural Network (CNN). The input is a video frame with a resolution of 1920x1080, and the output is a true / false probability vector, for example, [0.7, 0.3], indicating that 70% are true and 30% are false. The first contour feature map is generated by a face detection algorithm (such as MTCNN), extracting the coordinates (x, y) of 68 key points, such as the nose tip coordinates (500.5, 600.3), forming a feature vector with a dimension of 136. The micro-expression feature sequence is analyzed by optical flow method for 10 consecutive video frames, calculating the displacement vectors of areas such as eyebrows and corners of the mouth, obtaining a 12-dimensional feature vector for each frame, with a total sequence length of 120 dimensions. Next, multimodal features are integrated using a Conditional Random Field (CRF) model. The CRF input includes preliminary true / false probabilities, contour feature vectors, and micro-expression sequences, with a total feature dimension of 1 + 136 + 120 = 257 dimensions. The CRF model defines states as {true, false}, and transition probabilities are calculated based on the training data; for example, the transition probability from true to true is 0.8, and from false to false is 0.75. The node potential function uses a Gaussian kernel to calculate the matching degree between features and states, such as exp(-||x-μ||^2 / 2σ^2), where μ is the mean of the training set, and σ = 0.1. The CRF is optimized using a belief propagation algorithm, iterating 5 times to obtain posterior probabilities; for example, the probability of a true state is 0.85, and that of a false state is 0.15. Finally, the true state with the highest probability is selected as the judgment result. Analysis shows that contour features provide static structural information, while micro-expression sequences capture dynamic changes. CRF effectively integrates the spatiotemporal characteristics of both, improving judgment accuracy. In terms of business applications, the system can output the results to a security monitoring platform, triggering an alarm threshold (e.g., probability > 0.8) to achieve automated true / false detection.

[0080] Step S105: Generate the final true / false judgment result and make a decision.

[0081] Since the posterior probability of "fake" (0.8) is much higher than that of "true" (0.2), the system generates the final judgment: "Avatar_Clone" is a cloned image. This result will be returned to the platform, which can then process the user-uploaded content accordingly. Because the result is "fake," the process ends here and no blockchain record is triggered.

[0082] If the final authenticity judgment result is passed and the verified virtual image identifier is obtained, then blockchain technology is used to record it to the image genealogy chain, generate a traceable verification log, and execute step S106.

[0083] The input data is processed by a verification algorithm to determine the authenticity of the virtual avatar. Using pre-defined identifier extraction rules, the verified virtual avatar identifier is obtained from the verification results. The virtual avatar identifier is recorded in the avatar genealogy chain via a blockchain interface, generating an on-chain storage address. Based on the on-chain storage address, a traceable verification log containing a timestamp and identifier is generated. If the verification log is successfully generated, it is encrypted using a hash algorithm to obtain an encrypted log. The encrypted log is stored on the blockchain network through a distributed storage system, obtaining storage confirmation. Based on the storage confirmation, the log index in the avatar genealogy chain is updated, completing the traceability record.

[0084] Specifically, the final authenticity verification is achieved on the blockchain through smart contracts. The SHA-256 algorithm is used to encrypt the virtual avatar's feature data (such as a 256-bit hash value) to generate a unique identifier. Assume the input virtual avatar includes attributes such as hair color (RGB value: 255,0,0), height (1.75 meters), and creation time (2025-08-05 06:45:00). These attributes are concatenated into the string "255,0,0|1.75|2025-08-05 06:45:00", and the hash value "a1b2c3... (64 bits)" is calculated. This hash value is then compared with a pre-stored authenticity database. Verification is successful if the authenticity pass rate is above 95%. Verified identifiers (e.g., "Avatar_001") are recorded to the Avatar Genealogy Chain via a smart contract. Each block on the chain contains a timestamp (2025-08-05 06:45:00), the previous block hash (e.g., "d4e5f6..."), and identifier data. The Ethereum blockchain is used, with transaction gas fees of approximately 0.002 ETH. The recording process uses a Merkle tree structure to generate a root hash "m1n2o3...", ensuring data immutability. A traceable verification log is generated in JSON format, containing fields: identifier ("Avatar_001"), verification time (2025-08-05 06:45:00), hash value ("a1b2c3..."), and block height (e.g., 10086). The log is stored via IPFS, generating an access hash "QmXyz...", which is then stored on the blockchain for evidence. The analysis process verifies on-chain data to ensure 100% consistency between logs and block data. If inconsistencies are found, an alert is triggered to the management node. This closed-loop process ensures that the verification, recording, and log generation are fully automated, and that data is traceable and secure.

[0085] The method further includes step S107, which involves obtaining historical verification logs from the image genealogy chain, generating verification efficiency indicators through time series analysis, and determining the data basis for system optimization. Based on the verification efficiency indicators, a genetic algorithm is used to adjust the feature extraction parameters, generate an updated convolutional neural network model, and obtain an optimized feature extraction framework.

[0086] Historical verification logs are retrieved from the image genealogy chain, and the timestamps and verification results in the logs are parsed to obtain structured verification data. Time series analysis is used to process the structured verification data, calculating the statistical characteristics of verification time and success rate to obtain a verification efficiency index. If the verification efficiency index is lower than a preset threshold, data for system optimization is extracted to determine the range of feature extraction parameters to be adjusted. A genetic algorithm is used to iteratively optimize the feature extraction parameters, generating a candidate parameter set to obtain an optimized parameter combination. The convolutional neural network model is updated using the optimized parameter combination, and the model is retrained to obtain an updated feature extraction framework. The verification efficiency index of the updated feature extraction framework is obtained, and its trend is compared with that of historical verification efficiency indices to determine the optimization effect. If the optimization effect does not reach the preset threshold, the adjustment of the feature extraction parameter range is returned, and the genetic algorithm optimization and model update are repeated to obtain the final feature extraction framework.

[0087] Example 2

[0088] This embodiment describes Figure 5 The system architecture shown illustrates the functions and interactions of each module, which could be a service deployed in the cloud.

[0089] Feature Extraction Module: This is a core computational module that internally encapsulates the CNN model and PCA algorithm described in Example 1. It provides an API interface to receive image data (image or video URL) and returns the computed 256-dimensional initial feature representation. This module is typically supported by a GPU server cluster.

[0090] Type discrimination module: A lightweight service with a pre-trained SVM classifier deployed on-premises. It receives the initial feature representation generated by the feature extraction module and quickly returns a discrimination label of "cartoon character" or "realistic portrait".

[0091] Refined Validation Module: This is a parent module that contains two child modules:

[0092] Contour feature verification module: Activated when a request of type "cartoon" is received, it performs... Figure 2 The logic of the left path is as follows: call the OpenCV library to perform Canny edge detection, use a graph algorithm library (such as NetworkX) to build and store the contour topology map, and execute a graph matching algorithm (such as graph isomorphism or graph edit distance algorithm) to calculate the similarity with the original image reference map.

[0093] Micro-expression feature verification module: Activated when a request of type "realistic" is received. It has built-in... Figure 3 The spatiotemporal attention network model shown is implemented in TensorFlow or PyTorch and is responsible for processing temporal data and outputting micro-expression features.

[0094] Results Integration Module: This module implements the CRF model. It collects preliminary judgment probabilities and key verification features from various modules, organizes them into the format required by the CRF model, performs final probabilistic inference, and outputs high-confidence true / false judgments.

[0095] Blockchain Recording Module: This module acts as a bridge between the system and the blockchain network. It is triggered when the result integration module outputs a "true" judgment. It generates a JSON-formatted verification log, containing the image ID, timestamp, and a characteristic fingerprint calculated using the SHA-256 algorithm. Then, it uses Web3.js or a similar library to call the addVerificationRecord function of a smart contract pre-deployed on Ethereum (or a higher-performance consortium blockchain), sending the log as a transaction payload onto the blockchain.

[0096] Meanwhile, this module is responsible for sending transactions, waiting for confirmation (waiting for block packaging), and returning the transaction hash after a successful transaction, which serves as proof of the notarization.

[0097] System self-optimization module: This can be a background, periodically running daemon process. It includes:

[0098] Data retrieval: On a daily schedule, verification logs from the past 24 hours are retrieved in batches from the Image Genealogy Chain via a read-only function or event listener in the smart contract.

[0099] Analysis engine: Uses data analysis libraries such as Pandas to perform statistics on logs, such as calculating the average time for verifying cartoon characters and the proportion of fuzzy cases with similarity scores in the range of [0.6, 0.8] in realistic portraits.

[0100] Genetic Algorithm Solver: If the analysis engine detects a deterioration in performance metrics (e.g., the proportion of fuzzy cases exceeds 5%), it activates the genetic algorithm. The algorithm's "chromosomes" are a set of key hyperparameters of the CNN model, such as [number of convolutional layers, learning rate, dropout rate]. It iteratively generates new parameter combinations (population) and quickly evaluates the performance (fitness) of each combination using a small subset of validation data, ultimately finding the optimal parameter combination that improves the metric.

[0101] Model updater: Once the optimal parameters are found, it will automatically trigger a new model training / fine-tuning task and deploy the trained new model to the feature extraction module or the fine-tuning verification module to achieve seamless system upgrades.

[0102] Example 3

[0103] In a large metaverse platform called "MetaVerse-X", the economic system relies heavily on original, unique virtual avatar assets (sold as NFTs).

[0104] Scenario 1: Defending against high-precision dynamic cloning

[0105] A malicious user known as "Hacker" attempted to clone a realistic image of "Star," a top KOL on the platform. Hacker used advanced DeepFake software to generate a video in which the clone not only looked exactly like Star but also mimicked Star's signature smile.

[0106] When the hacker uploads the cloned image, MetaVerse-X's verification system kicks in:

[0107] 1. Initial detection: The system extracted the initial features of the clone image and found that it was very similar to the original features of Star in static terms. The initial judgment probability P(true) = 0.8, which is a high-risk fuzzy result.

[0108] 2. Type identification: The system determines that the image is a "realistic portrait" and forwards it to the micro-expression module.

[0109] 3. Refined Validation: The spatiotemporal attention network of the micro-expression module began analyzing the video. Although the signature smile was imitated, the network's temporal attention mechanism captured that in frames 3-5 of the smile's initiation, the clone's eye muscles did not exhibit natural, coordinated contraction—a flaw that is difficult for ordinary CNNs to detect. The model output a score with a micro-expression sequence similarity of only 0.68 to the benchmark Star.

[0110] 4. Final Fusion: The CRF model received two contradictory signals: the initial judgment leaned towards "true" (0.8), but the micro-expression dynamic verification strongly pointed to "false" (0.68). Since micro-expression features were given higher weights during CRF training, the final model overwhelmingly output P(false|observation) = 0.95.

[0111] 5. Result: Hacker's upload was rejected, and the platform's assets were secured.

[0112] Scenario 2: Confirmation and Tracing of Image Assets

[0113] An artist named "Creator" posted his meticulously designed original cartoon dragon character "Drake" on the platform, and it passed the system's authenticity verification.

[0114] 1. Verification passed: The system determined that "Drake" is a cartoon character, and its outline topology does not match any existing character in the database, so it was ultimately determined to be "real".

[0115] 2. On-chain Evidence Storage: The blockchain recording module is activated. It generates a log for "Drake": { "id": "creator-drake-v1", "timestamp": "2025-08-05T10:30:00Z", "fingerprint": "hash_of_topology_graph"}. This log is packaged into block 10086 of the image genealogy chain.

[0116] 3. Establishment of Rights: "Creator" now possesses public, verifiable, and tamper-proof evidence proving that he owned the original rights to the image at that point in time. Any image subsequently highly similar to "Drake's" outline topology will be easily identified as a clone by the system.

[0117] Compared to traditional verification methods, such as using a Siamese network to directly compare the feature vectors of two images, there are significant shortcomings:

[0118] For the dynamic clones in Scenario 1, the Siamese network can only perform static feature comparisons and cannot capture temporal flaws in micro-expressions, making it highly likely to misjudge the cloned image as "real". For cartoon characters, if the cloner only slightly changes the color or adds some textures that do not affect the outline, the Siamese network may also be misled. Figure 6 This diagram illustrates a comparison of key performance indicators between the method of this application and traditional methods. In the diagram, the diagonally filled bars represent the method of this application, while the gray-filled bars represent traditional methods (such as methods based on a single Siamese Network). As can be seen from the diagram, the method of this application (98.5% and 97.9%) significantly outperforms the traditional method (85.0% and 90.5%) in both "cartoon character accuracy" and "realistic portrait accuracy," thanks to the type-adaptive verification framework of this application. In terms of "resistance to dynamic attacks," the method of this application (95.0%) outperforms the traditional method (65.0%), demonstrating the powerful ability of spatiotemporal micro-expression analysis to capture dynamic artifacts. Regarding the "traceability and credibility" indicator, the method of this application achieves 100 points due to the introduction of blockchain technology, while the traditional method scores 0 due to data-driven centralized storage. This diagram visually demonstrates the advantages of this application over existing technologies.

[0119] In summary, the verification results of traditional solutions exist only in the platform's own database and lack credibility. The blockchain mechanism proposed in this application provides a high level of trust assurance. Traditional models cannot evolve on their own. However, the self-optimization module in this application enables the system to automatically patch itself by learning from historical data, maintaining long-term defense capabilities as hackers develop new attack methods.

[0120] Example 4

[0121] This embodiment details the specific steps of a method for verifying the authenticity of a cloned virtual image, referring to... Figure 7-8 .

[0122] Step S101: Obtain the virtual avatar data to be verified and perform preprocessing.

[0123] The purpose of this step is to standardize the raw data, which vary in format, size, and quality, into a unified format suitable for subsequent processing by artificial intelligence models.

[0124] Data source: The virtual avatar data to be verified can be a static image (e.g., a user-uploaded .jpg or .png format avatar) or a dynamic video (e.g., a short video in .mp4 or .mov format used for liveness detection when a user logs in).

[0125] Preprocessing Details: 1. Decoding and Frame Extraction: For video data, decoding is first performed using libraries such as FFmpeg, and image frame sequences are extracted according to a preset sampling rate (e.g., 15 frames per second). 2. Size Normalization: Both static images and video frames are scaled to a uniform resolution, such as 512x512 pixels, using bilinear interpolation. This ensures that the tensor dimensions input to subsequent network models are consistent. 3. Pixel Value Normalization: The RGB three-channel pixel values ​​of the image are transformed from the range of [0, 255] to the range of [-1, 1], calculated using the formula pixel_norm = (pixel_orig / 127.5) - 1.0. This operation helps prevent gradient vanishing or exploding, accelerating the convergence speed of model training. 4. Data Augmentation (Training Phase Only): During model training, to improve the model's generalization ability, a series of random data augmentation operations are applied to the training samples, including: random horizontal flipping, small-angle rotation (-10° to +10°), brightness and contrast adjustment (±20%), and the addition of a small amount of Gaussian noise. This allows the model to better cope with changes in lighting, angles, etc. in the real world.

[0126] Step S102: Determine the style type using a pre-defined image style classification model. Train an efficient convolutional neural network as a style classifier.

[0127] Model selection: To ensure low latency in the overall verification process, lightweight network architectures are preferred, such as Google's MobileNetV3-Small or EfficientNet-B0. These models exhibit excellent performance on both mobile and server-side devices.

[0128] Model Training: A large dataset of virtual avatar styles was constructed, containing over 500,000 images, covering two main categories: "Cartoon Style" (including subcategories such as Japanese, American, and Q-version) and "Realistic Style" (including subcategories such as Cyberpunk, Fantasy, and Modern Realism). Using this dataset, the pre-trained MobileNetV3 model was fine-tuned through transfer learning to focus on the style classification task of virtual avatars.

[0129] Decision Output: Given a preprocessed image as input, the model outputs a JSON object containing probabilities for various styles, such as {"style": "cartoon", "confidence": 0.998} or {"style": "realistic", "confidence": 0.995}. The system selects the style with the highest confidence as the image's type and uses it as a signal for the next step.

[0130] Step S103: Based on the style type, start the corresponding differential feature extraction pipeline.

[0131] This is the core technical contribution of the present invention, which abandons the traditional one-size-fits-all approach and achieves accurate feature extraction.

[0132] 1. General appearance feature extraction: Regardless of style, the system runs a CNN-based basic feature extractor to obtain general appearance information of the image.

[0133] We use a pre-trained ResNet-50 model, removing its final classification layer, and use it as a feature extractor. The pre-processed image is input, and the output of its intermediate layers is extracted. After global average pooling, a 2048-dimensional vector is obtained. This vector is the first texture feature set.

[0134] Simultaneously, the three-channel color histogram of the image in the HSV color space is calculated, and the three histograms are concatenated into a vector as the first color distribution feature set.

[0135] The two feature sets are combined and dimensionality is reduced using principal component analysis (PCA), retaining 95% of the variance information to generate a low-dimensional (e.g., 256-dimensional) general appearance feature vector.

[0136] 2. Differentiated feature extraction: This includes two different scenarios.

[0137] Scenario 1: If the style is cartoon style

[0138] Reference Figure 9 The system activates the "cartoon feature pipeline" to extract structured contour features.

[0139] 1. Contour Extraction and Refinement: The Canny edge detection algorithm is applied to the image to obtain a binarized edge map. Morphological operations such as erosion and dilation are performed to remove noise and isolated lines.

[0140] 2. Contour Graph Structure Construction: Graph theory methods are used to model the contour. First, the Douglas-Peucker algorithm is used to approximate the contour lines into polygons, simplifying the lines. Then, the intersections of the vertices of the approximate polygons and line segments are defined as graph nodes, and the line segments connecting the nodes are defined as graph edges.

[0141] Node attributes: Each node contains its (x, y) coordinates and the degree of the node (the number of edges it connects).

[0142] Edge properties: Each edge includes its length, average curvature, and angle relative to the image horizon.

[0143] 3. Graph Neural Network (GNN) Analysis: The constructed contour map is fed into a pre-trained Graph Isomorphism Network (GIN). Through multiple rounds of neighborhood aggregation, GIN can learn deep topological information of the graph, such as high-level structural patterns like "a triangle connecting to a circle." The final output of GNN is a 512-dimensional graph embedding vector, i.e., a structured contour feature. This feature is not sensitive to easily imitated elements such as color and fill, but has a very strong discriminative power for the skeletal structure of the image.

[0144] Scenario 2: If the style type is realistic

[0145] Reference Figure 10 The system activates the "realistic feature pipeline" to extract spatiotemporal dynamic micro-expression features.

[0146] 1. Face Detection and Keypoint Tracking: For the input video frame sequence, a high-performance face detector (such as YOLOv5-Face) is used to locate faces. Within the located face region, the MediaPipe Face Mesh model is used to detect and track 468 3D facial keypoints in real time.

[0147] 2. Dynamic Feature Extraction: The positions of keypoints move with facial expressions. We focus on the relative displacement and velocity of these points, rather than their absolute coordinates. We calculate the normalized displacement vector of each keypoint relative to the center point of the face (e.g., the center of the bridge of the nose) in each frame.

[0148] 3. Spatiotemporal network modeling: Input the normalized displacement vector sequence of keypoints for 32 consecutive frames (sequence length is 32, one vector per frame) into a convolutional long short-term memory network (Conv-LSTM).

[0149] Conv-LSTM: This network replaces the fully connected operations in LSTM with convolutional operations, enabling it to capture temporal dependencies like LSTM and learn spatial correlations like CNN (e.g., the linkage pattern between the left corner of the eye and the left corner of the mouth).

[0150] The network's final hidden state is viewed as a 512-dimensional vector, representing spatiotemporal dynamic micro-expression features. It encodes unique, subtle, and subconscious facial expression dynamics that cannot be simulated by static cloning or simple animation loops.

[0151] Step S104: Use a Conditional Random Field (CRF) model to integrate multimodal features. Fuse multiple features extracted in previous steps and make a final judgment.

[0152] Model Construction: We use a linear CRF model, the goal of which is to predict a label (real / fake) for the input virtual avatar. The energy function of a CRF is defined as the sum of the node potential and the edge potential. In this scenario, we mainly focus on constructing the node potential.

[0153] Node potential function construction: The node potential function of CRF measures the probability of assigning a certain label (real / fake) to the current input data. We concatenate general appearance features and differential features (structured contours or spatiotemporal dynamic micro-expressions) into a total feature vector F.

[0154] The node potential Φ(y, F) = W_y · F, where y is the label (real or fake) and W_y is the weight vector corresponding to the label, which is learned by the model during the training phase.

[0155] Confidence weighting: To make the model more intelligent, we introduced confidence when constructing the total feature vector F.

[0156] F_g is the general appearance feature vector, and F_d is the differential feature vector. c_g and c_d are their respective confidence coefficients. c_d is directly taken from the confidence of the style classifier in step S102. c_g can be obtained by analyzing the stability of the general features (such as image sharpness assessment).

[0157] This approach allows CRF to rely more heavily on distinctive features when style classification is highly certain, and vice versa. It is an implicit, data-driven weighting mechanism that replaces complex attention networks.

[0158] Inference and Output: Through calculation, the model outputs a probability distribution, such as {"is_real": 0.985, "is_fake": 0.015}, thus obtaining the final true or false judgment.

[0159] Step S105: Blockchain Records and System Optimization.

[0160] 1. Image-based family tree blockchain record:

[0161] When the judgment result is "true", the system calls the blockchain recording module. This module uses the SHA-256 algorithm to perform hash operations on the core feature vector and metadata (such as creator ID) of the virtual character to generate a unique digital identity fingerprint.

[0162] By interacting with smart contracts deployed on consortium blockchains such as Hyperledger Fabric, the `registerOriginalAvatar(fingerprint, metadata)` function is invoked. The transaction is then packaged into a new block, forming a permanent, tamper-proof "birth certificate." Anyone can verify its authenticity by calling the `queryAvatar(fingerprint, metadata)` function.

[0163] 2. System dynamic optimization closed loop (adaptive evolution), such as Figure 11 As shown:

[0164] Performance monitoring: The smart contract also includes an updatePerformanceMetrics(accuracy,latency) function, which verifies that the system will upload the daily performance statistics to the blockchain.

[0165] Triggering mechanism: A separate monitoring contract periodically checks performance data. If the accuracy rate is found to be below the preset Service Level Agreement (SLA) of 99.5% for a consecutive week, it will automatically trigger an "OptimizationRequired" event.

[0166] Genetic Algorithm Optimization: This event invokes an offline optimization server. The server starts a genetic algorithm (GA) process to tune the hyperparameters of the core model (such as the number of layers in the GNN, the learning rate of the Conv-LSTM, etc.).

[0167] Population: Each individual is a set of hyperparameters.

[0168] Fitness function: The fitness is a weighted average of accuracy and speed on the most recently collected "difficult sample" test set.

[0169] Evolution: Through selection, crossover, and mutation operations, the optimal combination of hyperparameters is found after hundreds of generations of evolution.

[0170] Model update: The model is retrained using the found optimal parameters, and the new model is securely deployed to the online system via the CI / CD pipeline, completing an adaptive iteration against new attacks.

[0171] Figure 8 This is a schematic diagram of the differential feature extraction and fusion framework in the embodiments of this application, which intuitively illustrates the logical flow of the present invention for authenticity verification.

[0172] The process begins with receiving data on the virtual avatar to be verified. First, the data is fed into a pre-trained, lightweight style classification model (e.g., MobileNetV3). This model acts as the "intelligent scheduling hub" of the entire process, and its core task is to quickly and accurately determine the style of the input avatar, for example, classifying it as "cartoon style" or "realistic style".

[0173] Based on the output of the style classification model, the verification process will enter two parallel, differentiated processing paths:

[0174] 1. General Appearance Feature Extraction: Regardless of the image style, the system initiates a general feature extraction pipeline (e.g., based on ResNet-50) to extract basic texture and color information from the image data, forming a general appearance feature vector. This path ensures a consistent evaluation of the basic visual information for all images.

[0175] 2. Differentiated Feature Extraction: If the style classification result is "Cartoon Style," the system will activate the Cartoon Style pipeline. This pipeline focuses on analyzing the structure and line art of the image, extracting its unique structured contour features by constructing and analyzing contour topology graphs. If the style classification result is "Realistic Style," the system will activate the Realistic Style pipeline. This pipeline is specifically designed for processing dynamic video data, extracting micro-expression features that reflect the dynamic realism by analyzing the spatiotemporal changes of facial micro-expressions.

[0176] Finally, the feature vectors from the general appearance feature extraction path and the feature vectors from the activated differential feature extraction path (i.e., structured contour features or spatiotemporal dynamic micro-expression features) are jointly fed into a multimodal fusion module. In this application, this module preferably employs a Conditional Random Field (CRF) model. The CRF model can intelligently integrate these feature information from different modalities and, based on pre-learned knowledge (including the importance of each feature under different styles), calculate the posterior probability of the image to be verified as "real" or "fake," thereby generating the final judgment result.

[0177] Example 5

[0178] This embodiment provides a system for verifying the authenticity of cloned virtual avatars, referring to... Figure 12 The system is designed as a microservice architecture and can be deployed in the cloud (such as AWS, Alibaba Cloud) or a private data center.

[0179] Data preprocessing module: As a standalone microservice, it receives HTTP requests from the API gateway, with the request body containing image or video data. It performs standardized processing using OpenCV and FFmpeg libraries, temporarily stores the results in object storage (such as S3), and returns a URI of the processed data.

[0180] The style classification module receives the URI of preprocessed data, downloads the data, and sends it to a TensorFlow Serving instance loaded with a MobileNetV3 model. Inference results (such as {"style": "realistic", "confidence":0.995}) are published to a message queue (such as RabbitMQ or Kafka) with the topic `style_classified`.

[0181] Differentiated Feature Extraction Module: This is a service scheduler or parent module that subscribes to the `style_classified` topic. Based on the message content, it distributes tasks to one of two different worker services:

[0182] Cartoon Feature Service: Subscribes to the cartoon_feature_request topic. It integrates the PyG (PyTorchGeometric) library to perform contour extraction, graph construction, and GNN inference, and publishes the generated 512-dimensional structured contour feature vectors to the feature_extracted topic.

[0183] Realistic Feature Service: Subscribes to the realistic_feature_request topic. Internally, it integrates the MediaPipe and PyTorch libraries to perform face tracking and Conv-LSTM inference, and publishes the generated 512-dimensional spatiotemporal dynamic micro-expression feature vectors to the feature_extracted topic.

[0184] Meanwhile, a general feature service retrieves data from object storage in parallel, extracts general appearance features using ResNet-50, and publishes them to the same feature_extracted topic.

[0185] Multimodal fusion module: This module subscribes to the `feature_extracted` topic. It collects common and differential features from the same request ID, concatenates and weights them, and then feeds them into a CRF inference engine (which can be based on Scikit-learn or a custom implementation). The final probability judgment result is returned to the initial requester via API callback or message queue.

[0186] Blockchain Record Module: When the fusion module determines that the transaction is genuine, it sends a request to this module. This module contains the Hyperledger Fabric client SDK, which is responsible for communicating with the blockchain network, constructing and submitting the transaction, and putting the illustrative "digital identity fingerprint" on the blockchain.

[0187] System Dynamic Optimization Module: This is a background scheduling system (such as Airflow or Kubeflow). It runs tasks periodically and queries performance data through a blockchain client. When the triggering conditions are met, it automatically launches a containerized training environment containing a genetic algorithm library (such as DEAP) and a machine learning framework (TensorFlow / PyTorch), performs model optimization, and registers the trained new model version to a model repository (such as MLflow Registry) for deployment.

[0188] Example 6:

[0189] Imagine "ArtVerse" as a trading platform focused on high-value, limited-edition virtual avatars (as NFT artworks). Each avatar is created by a renowned digital artist and commands a premium price. The platform's credibility is built on the guarantee that every NFT sold corresponds to a unique, inimitable original digital avatar.

[0190] The technical solution of this application is applied on this platform:

[0191] 1. Originality certification before "casting":

[0192] Renowned artist Seraphim has completed a new cartoon-style piece, StardustCat, which is set to be minted as a limited-edition NFT on ArtVerse.

[0193] Authentication Process: Seraphim uploaded high-resolution design images of "Stardust Cat" and a short animated video showcasing its signature tail-wagging motion. The platform's backend verification system received this data.

[0194] Step S102: The style classification module classifies it as "cartoon style" with a confidence level of 99.9%.

[0195] Step S103: The general feature extractor extracts its unique starry gradient fill (color feature) and fur texture (texture feature). Simultaneously, the cartoon feature pipeline is activated, analyzing the "Stardust Cat's" silhouette—particularly its iconic comet-shaped tail composed of three specific curves and the spiral patterns on its ears. The GNN encodes this unique topology into a 512-dimensional structured silhouette feature vector.

[0196] Step S104: The multimodal fusion module integrates all features. Due to the extremely high confidence level of the style classification, the CRF model assigns very high weights to the structured contour features when making decisions.

[0197] On-chain authentication: After successful authentication, the system packages its feature summary, artist signature, creation date, and other metadata, and uploads it to the blockchain through the blockchain recording module to generate a permanent, publicly searchable authentication ID. This ID is written into the metadata of the "Stardust Cat" NFT as its "authenticity certificate".

[0198] 2. Anti-counterfeiting measures during transactions and displays:

[0199] A collector purchased the "Stardust Cat" NFT and displayed it in their personal metaverse. A speculator attempted to create a highly similar-looking counterfeit "Nebula Cat" to mislead consumers.

[0200] Counterfeit detection: When "Nebula Cat" is uploaded to any metaverse platform that collaborates with "ArtVerse", its data will be sent to this verification system for routine scanning.

[0201] Differences revealed: Although "Nebula Cat" is very similar in color and general shape, the curvature of its tail and the topological structure of the spiral patterns on its ears are significantly different from the original GNN feature vector of "Stardust Cat" (the Euclidean distance is far beyond the threshold).

[0202] Judgment result: The fusion module (m4) determined that "Nebula Cat" was 99.8% likely to be a counterfeit. The system will not only block its display, but will also report the record of this counterfeiting attempt (including the counterfeiter's account information) to the risk control department of the "ArtVerse" platform.

[0203] Authentic Product Display: When a collector's authentic "Stardust Cat" artwork is displayed, a dynamic authentication badge will float next to it. Visitors who click on the badge will be directly redirected to a blockchain explorer to view its immutable authentication ID and artist information, experiencing the prestige and security of owning a genuine work of art.

[0204] 3. Evolution of Long-Term Defense Systems:

[0205] With the development of AI painting technology, a new cloning technique has emerged that can better mimic the line style of specific artists. The system's false positive rate (identifying counterfeits as genuine products) has seen a slight increase, from one in 100,000 to five in 10,000.

[0206] Performance alert: The monitoring smart contract deployed on the consortium blockchain detected this negative trend by analyzing daily verification reports and automatically triggered the optimization process.

[0207] Intelligent Evolution: The system's dynamic optimization module is activated. The genetic algorithm begins optimizing the GNN model in the cartoon feature pipeline. Through multiple generations of evolution, the algorithm discovers that increasing the number of GNN network layers from 3 to 5 and adding more attributes to its input graph structure (e.g., the angle between line segments) can more finely capture subtle differences in the "strokes" of the lines.

[0208] Seamless Upgrade: After rigorous testing, the optimized G-v2 model was automatically deployed online. The entire defense system achieved an "immunity" upgrade against new cloning technologies without human intervention, ensuring the platform's continued security and reputation.

[0209] In summary, compared to traditional technologies, the core advantage of this application lies in shifting from passive defense to proactive cognition and adaptive evolution. It not only identifies the style of virtual avatars and adopts optimal strategies, but also establishes a trusted digital identity ecosystem through blockchain. By introducing an avatar genealogy blockchain, it establishes an immutable "digital identity gene" for each verified original avatar, providing authoritative and traceable verification history. In the event of identity disputes or transaction conflicts, on-chain data can serve as legally valid third-party evidence, effectively protecting the rights of the creator and users. Furthermore, by recording system performance indicators on the blockchain and monitoring them through smart contracts, a closed-loop self-optimizing system is constructed. When system performance declines due to new attack methods, it automatically triggers a model optimization process guided by a genetic algorithm, enabling the system to continuously resist evolving cloning technologies and ensuring long-term effectiveness and advancement. The above embodiments are merely illustrative of the technical solutions of this invention and not intended to limit it. The invention has been described in detail with reference to preferred embodiments. Those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of this invention without departing from the spirit and scope of the invention, and all such modifications and substitutions should be covered within the scope of the claims of this invention.

Claims

1. A method of verifying authenticity based on a cloned avatar, characterized by, The method comprises the following steps: Obtaining image data of a virtual image to be verified, and extracting initial feature expression thereof through a preset feature extraction network; According to the initial feature expression, the type of the virtual image is determined to be one of at least two preset image types; According to the type determination result, a fine verification module corresponding to the image type is selected to process the initial feature expression to obtain key verification features; wherein the fine verification module includes a contour feature verification module for cartoon images and a micro-expression feature verification module for realistic portraits; Fusing the initial feature expression and the key verification features, and generating a final authenticity judgment result through a multi-modal fusion model; the method further comprises: If the final authenticity judgment result is true, a verification log containing image identification and feature fingerprint is generated for the virtual image; Through a block link interface, the verification log is recorded as a transaction record in a preset image genealogy chain, realizing traceable storage of the verification result; Periodically obtaining historical verification log data from the image genealogy chain; Based on the historical verification log data, the performance indicators of the current verification system are analyzed; If the performance indicators are lower than the preset threshold, a self-optimization algorithm is started to adjust the hyperparameters of the feature extraction network or the fine verification module, and retraining is performed to generate an updated verification model.

2. The method of claim 1, wherein, The extraction of the initial feature expression comprises: processing the image data through a convolutional neural network to extract a first texture feature set and a first color distribution feature set; using principal component analysis to perform dimensionality reduction processing on the first texture feature set and the first color distribution feature set to generate a compressed feature vector as the initial feature expression.

3. The method of claim 1, wherein, The processing method of the contour feature verification module for cartoon images comprises: extracting a contour feature subset from the initial feature expression, and generating a contour graph through an edge detection algorithm; identifying key feature points in the contour graph, and constructing a contour topology graph according to the connection relationship between the key feature points; performing graph matching calculation on the contour topology graph and a preset reference topology graph of the original image to obtain a structural similarity score as the key verification feature.

4. The method of claim 1, wherein, The processing method of the micro-expression feature verification module for realistic portraits comprises: processing the image data into a time sequence image sequence and inputting it into a preset spatio-temporal attention network model; the spatio-temporal attention network model performs weighted processing on the face key area and key time frame in the time sequence image sequence, and extracts a micro-expression feature sequence that integrates spatio-temporal dynamic information as the key verification feature.

5. The method of claim 1, wherein, The multi-modal fusion model is a conditional random field model; the fusion step comprises: inputting the initial feature expression corresponding to the preliminary authenticity probability and the key verification features as an observation sequence into the conditional random field model; calculating the posterior probability of the true and false states under the observation sequence through the conditional random field model, and selecting the state with the highest posterior probability as the final authenticity judgment result.

6. A clone avatar based authentication system for implementing the method of any one of claims 1-5, characterized in that, It comprises: The feature extraction module is configured to obtain image data of a virtual image to be verified, and extract initial feature expression of the virtual image through a preset feature extraction network. The type discrimination module is configured to discriminate the type of the virtual image according to the initial feature expression, so as to determine that the virtual image is one of at least two preset image types. The fine verification module is configured to select a module corresponding to the image type according to the type discrimination result, and process the initial feature expression to obtain key verification features. The result integration module is configured to fuse the initial feature expression and the key verification features, and generate a final authenticity judgment result through a multi-modal fusion model. The blockchain recording module is configured to, when the final authenticity judgment result is true, generate a verification log containing an image identifier and feature fingerprints for the virtual image, and record the verification log into a preset image genealogy chain through a blockchain interface. The system self-optimization module is configured to periodically obtain historical verification log data from the image genealogy chain, analyze performance indicators of a current verification system, and start a self-optimization algorithm to adjust hyperparameters of the feature extraction module or the fine verification module to generate an updated verification model when the performance indicators are lower than a preset threshold.

7. A method of verifying authenticity based on a cloned avatar, characterized by, The method comprises the following steps: Obtain virtual image data to be verified and pre-process the virtual image data; Analyze the pre-processed virtual image data using a preset image style classification model to determine the style type of the virtual image data, wherein the style type at least includes cartoon style and realistic style; According to the style type, a corresponding differential feature extraction pipeline is started: if the style type is cartoon style, the structured contour features are extracted; if the style type is realistic style, the spatiotemporal dynamic micro-expression features are extracted; Fuse the general appearance features of the virtual image to be verified with the differential features to generate a final authenticity judgment result; If the final authenticity judgment result is true, obtain the identity identifier and feature abstract of the virtual image; Record the identity identifier and feature abstract into a preset image genealogy blockchain through a smart contract to form a traceable verification log; Periodically monitor and record the performance indicators of the verification method through the smart contract; If the performance indicators are lower than a preset threshold, the smart contract automatically triggers an offline optimization process to update the artificial intelligence model in the image style classification model or the differential feature extraction pipeline, wherein the offline optimization process uses a genetic algorithm to search and optimize the hyperparameters of the artificial intelligence model with the improvement of the performance indicators as the optimization target.

8. A clone avatar-based authentication system for implementing the method of claim 7, characterized in that, The method comprises: A data pre-processing module is configured to obtain and pre-process virtual image data to be verified; A style classification module is configured to determine the style type of the virtual image data, wherein the style type at least includes cartoon style and realistic style; A differential feature extraction module is configured to extract structured contour features or spatiotemporal dynamic micro-expression features according to the style type. A multi-modal fusion module is configured to fuse the general appearance features and the differentiated features and generate a final authenticity judgment result. A blockchain recording module is configured to, when the final authenticity judgment result is true, record the identity of the virtual image and the feature digest to an image genealogy blockchain through a smart contract. And A system dynamic optimization module is configured to, in response to a trigger signal issued by the smart contract according to a system performance index, update the artificial intelligence model in the system.

Images