A method for generating a watermark for an RWA asset physical state off-chain photograph uploading

Abstract

The application provides a RWA asset physical state off-chain photograph uploading watermark generation method, comprising: S1. Data acquisition and algorithm pre-judgment; S2. Photograph shooting and watermark generation; S3. Algorithm interaction and iterative optimization; S4. Hash calculation and on-chain evidence storage. The application aims to solve the problems of algorithm coordination loss, insufficient creativity, weak authenticity verification and risk prevention and control capability in the existing RWA asset photograph uploading scheme, and realizes the authenticity, security and efficiency of RWA asset physical state record by constructing an algorithm fusion interaction system and an optimization basic module, meets the business compliance and risk prevention and control requirements, and improves the scene adaptability, resource utilization efficiency and operation convenience, and is suitable for multiple types of RWA assets and complex business scenarios.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence, blockchain, image processing and data security, and in particular to a method for generating watermarks by uploading photos of RWA assets in the physical state chain. Background Technology

[0002] As RWA's business diversifies, asset physical status verification has become crucial for compliance and risk control, and off-chain photo uploading is the core means of connecting off-chain assets with on-chain information.

[0003] Existing solutions mostly adopt a "simple upload + storage" model, incorporating one or more AI algorithms, but these are merely conventional combinations, lacking effective interaction and closed-loop collaboration mechanisms. This leads to problems such as photos being easily tampered with, insufficient traceability information, unreasonable watermark design, serious resource waste, difficulty in verifying authenticity, and weak scene adaptability. These solutions struggle to address the complex risks of RWA asset verification and fail to meet business compliance and efficient prevention and control requirements.

[0004] 1. Photos are easily tampered with: The lack of a collaborative anti-tampering mechanism makes it difficult for a single anti-tampering algorithm to resist advanced tampering methods such as AI synthesis and reuse of historical photos. Malicious users can modify the content of photos or replace real photos in various ways, resulting in the distortion of asset status verification results. 2. Insufficient traceability information: Only photo files and basic association information are stored, without combining multi-dimensional time-series data and cross-scene verification information. In case of disputes, it is impossible to trace the complete scene chain of photo shooting through collaboration, making it difficult to verify the authenticity of asset status. 3. Unreasonable watermark design: The watermark embedding parameters are fixed or only a single algorithm is adjusted, without combining risk prediction and health monitoring for dynamic optimization. This can easily lead to problems such as watermark obscuring core details, insufficient anti-attack capabilities, or extraction failure during the verification stage. 4. Serious waste of resources: There is no precise triggering mechanism with algorithmic collaboration, and the processing resource allocation is not dynamically adjusted based on risk level and scene characteristics. All photos are processed indiscriminately, resulting in a large amount of unnecessary computing and storage resources being wasted. 5. Difficulty in verifying authenticity: The verification stage is mostly a one-way process of detection, without reverse correction and collaborative verification mechanisms. When third-party institutions or on-chain contracts are executed, they cannot quickly and accurately verify the tamper-proof attributes of the photos, which affects the progress of business processes. 6. Insufficient adaptability: It only adjusts for single-dimensional features and does not combine multimodal data and cross-time period scene features for collaborative adaptation. When facing complex scenarios (such as cross-border network environment and dynamic changes in asset status), the adaptability and robustness are insufficient. 7. Lack of algorithm collaboration: Multiple AI algorithms operate independently without interaction or closed-loop linkage mechanisms, making it impossible to achieve a technical effect of 1+1>2 and difficult to cope with the increasingly complex risks of RWA asset verification. Summary of the Invention

[0005] This invention provides a method for generating watermarks by uploading photos of RWA assets under the physical state chain. It aims to solve the problems of lack of algorithm collaboration, insufficient creativity, weak authenticity verification and risk control capabilities in existing RWA asset photo uploading solutions. By constructing an algorithm fusion and interaction system and optimizing basic modules, it achieves authenticity, security and efficiency in recording the physical state of RWA assets, meets the needs of business compliance and risk control, and improves scenario adaptability, resource utilization efficiency and operation convenience, adapting to multiple types of RWA assets and complex business scenarios.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: A method for generating watermarks by uploading photos of RWA assets under the physical state chain, including: S1. Initiate an RWA asset status update request, collect scene preview data and multimodal data. Multimodal data includes asset operation status data, asset historical time series data, and network environment data. Combined with preset asset basic information, which includes asset identification information and associated asset historical data, output scene similarity results through scene consistency verification algorithm, and output preliminary risk assessment results and scene adaptation parameters through multimodal risk prediction algorithm. S2. Trigger shooting based on scene similarity results and preliminary risk assessment results, acquire asset photos and preprocess them. After verification by the fake photo detection algorithm, embed watermark data based on scene adaptation parameters through an adaptive invisible watermark generation algorithm to obtain watermarked photos. The watermark data includes asset identification information and real-time collected related shooting information, and retains the original watermark information simultaneously. S3. Initiate bidirectional interaction and iterative optimization of the multimodal risk prediction algorithm, watermark health monitoring algorithm, and scene consistency verification algorithm, and record iterative interaction logs simultaneously. The multimodal risk prediction algorithm outputs real-time risk assessment results and scene adaptation parameters to guide the other two types of algorithms to adjust their running parameters and optimize watermark generation and embedding parameters. The other two types of algorithms output watermark health status, scene consistency results, and abnormal information based on watermarked photos, original watermark information, asset historical data, and real-time risk assessment results, respectively. They then reverse-correct the risk assessment results and optimize the model parameters. The watermark generation, health monitoring, and scene consistency verification are repeatedly executed until the results meet the standards or reach the preset iteration limit, resulting in optimized watermarked photos. The optimized output results are then summarized. S4. Perform hash calculation on the optimized watermarked photo to obtain the hash value, package the hash value, asset identification information, original watermark information, optimized output results and iteration interaction log, generate a notarization serial number and complete on-chain notarization through a blockchain node.

[0007] In this specification, the method for generating watermarks by uploading photos of RWA assets in physical state off-chain also includes: S5. The verifier submits materials to be verified, including photos to be verified, asset identification information, and notarization serial number. Through the AI ​​two-way verification mechanism, combined with on-chain notarization information, algorithm iteration logic, and optimized output results, the authenticity, watermark integrity, and scenario consistency are verified. A differentiated strategy is implemented with reference to the final risk assessment results, and a verification report containing verification conclusions and parameter optimization basis is output.

[0008] In this specification, the bidirectional interaction and iterative optimization of the multimodal risk prediction algorithm, watermark health monitoring algorithm, and scene consistency verification algorithm are as follows: the higher the real-time risk assessment result output by the multimodal risk prediction algorithm, the higher the monitoring dimension of the watermark health monitoring algorithm and the verification strength of the scene consistency verification algorithm, and the higher the optimization frequency of the watermark generation embedding parameters; when the watermark health monitoring algorithm outputs watermark anomaly information and the scene consistency verification algorithm outputs scene anomaly information, the abnormal feature data is simultaneously fed back to the multimodal risk prediction algorithm to correct the risk assessment result and perform weighted iterative optimization of the model parameters of the multimodal risk prediction algorithm, and simultaneously trigger the corresponding level of watermark generation re-execution action.

[0009] In this specification, the multimodal risk prediction algorithm generates risk assessment results and scenario adaptation parameters based on feature fusion and temporal correlation analysis of multimodal data. After performing feature standardization and dimension normalization processing on asset operation status data, asset historical time-series data, and network environment data, the algorithm extracts multi-dimensional temporal correlation features and asset risk features, combines them with asset historical data in the asset basic information to allocate risk weights, and outputs corresponding preliminary risk assessment results, real-time risk assessment results, and scenario adaptation parameters adapted to different risk levels.

[0010] In this specification, the scene consistency verification algorithm extracts scene spatial features and temporal features based on scene preview data, scene information in watermarked photos, and asset historical data, and performs feature matching calculations to output scene similarity results and scene consistency results. During the iterative optimization process, the algorithm adjusts the weight ratio of feature matching according to the real-time risk assessment results. The higher the risk assessment result, the higher the weight of temporal feature matching of asset historical data.

[0011] In this specification, the watermark health monitoring algorithm is based on the watermarked photo and the original watermark information. It comprehensively detects the integrity, anti-tampering ability and visual imperceptibility of the watermark data, and outputs the watermark health status and abnormal information. During the iterative optimization process, the algorithm adjusts the judgment threshold of the watermark health status according to the real-time risk assessment results. The higher the risk assessment result, the more stringent the judgment threshold. In addition, the abnormal information is synchronously associated with the corresponding area features of the watermarked photo and fed back to the watermark generation stage.

[0012] In this specification, the adaptive invisible watermark generation algorithm embeds watermark data in non-critical feature areas of asset photos based on scene adaptation parameters. The asset identification information in the watermark data is extracted from the basic asset information, and the associated shooting information is the shooting time and shooting environment data collected synchronously when the asset photo was taken. During the iterative optimization process, the algorithm adjusts the embedding position, embedding depth and encryption level of the watermark data according to the parameter optimization instructions output by the multimodal risk prediction algorithm, and generates optimized watermarked photos.

[0013] In this specification, the fake photo detection algorithm extracts image authenticity features and identifies forgery traces in the preprocessed asset photos. The detection dimensions include image pixel texture features, color distribution features, synthetic edge features, and asset physical feature matching degree. Only when the results of each detection dimension meet the preset authenticity standard is the asset photo determined to have passed the verification and enter the watermark generation process.

[0014] In this specification, the hash calculation of the optimized watermarked photo is an irreversible one-way hash operation. The generated hash value is bound to the asset identification information before being included in the packaged data. The on-chain notarization involves synchronously uploading all the packaged data to multiple blockchain nodes to complete distributed notarization. The notarization serial number forms a unique correspondence with the hash value and the asset identification information.

[0015] In this specification, the differentiation strategy is to classify risk levels based on the final risk assessment results in the optimized output. High-risk photos to be verified require additional verification using historical asset data and multimodal data. Medium-risk photos to be verified require verification of the watermark integrity and scene consistency. Low-risk photos to be verified only require basic hash value comparison and watermark extraction verification. The parameter optimization in the verification report is based on the watermark generation parameters, algorithm running parameters adjustment records, and optimization results recorded in the iteration interaction log.

[0016] In summary, this invention offers at least the following advantages: Significantly enhanced anti-tampering and fraud detection capabilities, effectively resisting advanced fraudulent techniques such as AI synthesis and historical photo reuse, ensuring the authenticity of asset status records. Dual protection of watermark health and robustness, transforming watermarks from passive repair to proactive early warning, avoiding extraction failures during the verification phase without affecting the display of core asset information. Enhanced scenario adaptability and robustness, flexibly adapting to complex scenarios such as cross-border networks and dynamic asset changes, covering multiple types of RWA assets. Pre-emptive and refined risk management, reducing manual intervention costs and improving verification and business progress efficiency through differentiated management strategies. Optimized resource utilization efficiency, avoiding waste of computing and storage resources caused by indiscriminate data processing, improving system performance. Enhanced end-to-end compliance and security, forming a complete and tamper-proof evidence chain, meeting data security and cross-border business compliance requirements. Low operational threshold and strong scalability, requiring no professional knowledge to operate, supporting the addition of AI modules and business scenarios, and adapting to more RWA asset types. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the process for generating watermarks by taking photos and uploading them under the RWA asset physical state chain involved in this invention.

[0018] Figure 2 This is a schematic diagram of the AI ​​adaptive invisible watermark embedding process involved in this invention.

[0019] Figure 3 This is a schematic diagram of the AI ​​two-way verification process involved in this invention.

[0020] Figure 4 This is a schematic diagram of the logical flow of the AI ​​dynamic risk assessment algorithm involved in this invention.

[0021] Figure 5 This is a schematic diagram of the multi-scenario AI parameter adaptive adjustment process involved in this invention. Detailed Implementation

[0022] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0023] like Figure 1 As shown, this embodiment provides a method for generating watermarks by uploading photos of RWA assets under the physical state chain, including: S1. Initiate an RWA asset status update request, collect scene preview data and multimodal data. Multimodal data includes asset operation status data, asset historical time series data, and network environment data. Combined with preset asset basic information, which includes asset identification information and associated asset historical data, output scene similarity results through scene consistency verification algorithm, and output preliminary risk assessment results and scene adaptation parameters through multimodal risk prediction algorithm. S2. Trigger shooting based on scene similarity results and preliminary risk assessment results, acquire asset photos and preprocess them. After verification by the fake photo detection algorithm, embed watermark data based on scene adaptation parameters through an adaptive invisible watermark generation algorithm to obtain watermarked photos. The watermark data includes asset identification information and real-time collected related shooting information, and retains the original watermark information simultaneously. S3. Initiate bidirectional interaction and iterative optimization of the multimodal risk prediction algorithm, watermark health monitoring algorithm, and scene consistency verification algorithm, and record iterative interaction logs simultaneously. The multimodal risk prediction algorithm outputs real-time risk assessment results and scene adaptation parameters to guide the other two types of algorithms to adjust their running parameters and optimize watermark generation and embedding parameters. The other two types of algorithms output watermark health status, scene consistency results, and abnormal information based on watermarked photos, original watermark information, asset historical data, and real-time risk assessment results, respectively. They then reverse-correct the risk assessment results and optimize the model parameters. The watermark generation, health monitoring, and scene consistency verification are repeatedly executed until the results meet the standards or reach the preset iteration limit, resulting in optimized watermarked photos. The optimized output results are then summarized. S4. Perform hash calculation on the optimized watermarked photo to obtain the hash value, package the hash value, asset identification information, original watermark information, optimized output results and iteration interaction log, generate a notarization serial number and complete on-chain notarization through a blockchain node. S5. The verifier submits the materials to be verified, including the photo to be verified, asset identification information and the notarization serial number. Through the AI ​​two-way verification mechanism, combined with the on-chain notarization information, algorithm iteration logic and optimized output results, the authenticity, watermark integrity and scenario consistency are verified. The verifier implements differentiated strategies with reference to the final risk assessment results and outputs a verification report containing the verification conclusion and parameter optimization basis.

[0024] In some embodiments, the cross-time-period scene consistency verification algorithm also incorporates a causal correlation verification model between asset physical state and shooting environment data. This model is built on a Bayesian network and uses the historical causal relationship between asset IoT data (such as device temperature and real estate energy consumption) and shooting environment data (such as light intensity and ambient temperature and humidity) as training basis to output the causal correlation confidence score. Specifically, the model first filters out strong causal correlation feature pairs (such as device operating load and shooting scene light stability, real estate energy consumption and personnel flow data during the shooting period) through mutual information calculation, and then obtains the causal correlation threshold based on historical samples (such as a confidence score ≥ 0.85 for normal causal relationship). During scene consistency verification, if the causal correlation confidence score between the currently captured asset IoT data and environmental data is lower than the threshold, even if the spatial and temporal feature similarity meets the standard, it is still judged as a scene abnormality, and the deep detection of the watermark health monitoring algorithm is triggered simultaneously. This technology breaks through the limitations of traditional verification that only relies on spatiotemporal features, and resists advanced fake methods such as "physical state simulation in fake scenes" through causal logic.

[0025] In some embodiments, the embedding rules of the adaptive invisible watermark generation algorithm are dynamically bound to the real-time block height of the blockchain. Specifically, before watermark generation, the system obtains the current block height (e.g., Ethereum block height) through the blockchain node interface, performs a SHA-256 hash operation on the block height, and takes the first 8 bits as the "dynamic embedding factor." Based on this factor, the core parameters of watermark embedding are adjusted: the offset of the embedding position (offset = dynamic embedding factor × photo width / 256), the attenuation coefficient of the embedding depth (attenuation coefficient = 1 - dynamic embedding factor × 0.1), and the encryption key fragment of the watermark data (key fragment = the XOR result of the dynamic embedding factor and the asset identification information). Simultaneously, this dynamic embedding factor and the original watermark information are synchronously stored on the blockchain. During the verification phase, the block height at the time of generation of the corresponding storage serial number must be obtained first, and the above hash operation must be repeated to obtain the same embedding factor before the watermark can be successfully extracted. This technology, through the immutability and real-time change characteristics of the blockchain block height, makes the watermark embedding rules "one-time" and "traceable," resisting targeted cracking of fixed embedding rules.

[0026] In some embodiments, the bidirectional interaction of the three core algorithms adopts a layered architecture of "edge node pre-interaction + on-chain final confirmation" to adapt to cross-border low network quality scenarios. The specific process is as follows: The terminal device (edge ​​node) deploys a lightweight version of the three algorithms locally. It first performs pre-interaction based on locally collected multimodal data, watermarked photos, etc., and outputs the pre-interaction results (including temporary risk assessment results, preliminary judgment of watermark health status, and scenario consistency pre-verification score). If the pre-interaction results meet the preset conditions (such as temporary risk assessment result < 60 points, preliminary judgment of watermark health status as normal, and scenario consistency pre-verification score ≥ 0.9), the pre-interaction results are directly synchronized to the chain to complete the final confirmation without repeating the complete interaction process. If the pre-interaction results do not meet the conditions (such as temporary risk assessment result ≥ 60 points), the complete data collected locally is uploaded to the core server, and the full algorithm interaction is performed before the results are uploaded to the chain. At the same time, the parameters of the edge node pre-interaction (such as the feature extraction dimension of the lightweight algorithm and the pre-interaction threshold) are updated periodically by the output results of the dynamic risk prediction algorithm stored on the chain (such as synchronizing once every 24 hours). This technology reduces reliance on network transmission in cross-border scenarios and improves algorithm interaction efficiency through layered interaction.

[0027] In some embodiments, the verification dimensions of the AI ​​two-way verification mechanism are dynamically adjusted based on the asset lifecycle stage. The asset lifecycle stage is automatically determined by the system based on basic asset information (such as acquisition time, depreciation rate, and historical maintenance records) and is divided into the new asset stage, stable operation stage, and aging stage. The specific adjustment rules are as follows: For assets in the new asset stage (acquisition time ≤ 6 months, depreciation rate < 10%), the verification dimensions focus on the integrity of the watermark and the authenticity of the photographed information (e.g., improving the matching accuracy between GPS location and asset registration address to ≤ 3 meters, and verifying the temporal consistency between associated photographed information and asset entry time); for assets in the stable operation stage (6 months < acquisition time ≤ 3 years, 10% ≤ depreciation rate < 30%), the basic verification dimensions (authenticity + watermark integrity + scene consistency) are maintained; for assets in the aging stage (acquisition time > 3 years, depreciation rate ≥ 30%), a dimension for verifying the consistency of changes in the asset's physical state is added (e.g., the matching degree between the asset's appearance wear characteristics in the watermarked photo and historical maintenance records, and the correlation verification between the watermark embedding area and easily damaged parts of the asset). This technology sets different verification focuses based on the risk characteristics of assets at different lifecycle stages, improving verification accuracy.

[0028] In some embodiments, the watermark health monitoring algorithm is linked in real time with an IoT data anomaly pattern library, which is built through offline training: collecting historical anomaly samples of asset IoT data (such as sudden voltage drops in devices, sudden changes in temperature and humidity in real estate), extracting anomaly pattern features (such as the inflection point of anomaly occurrence, duration of anomaly, and range of anomaly value fluctuations), classifying anomaly patterns into several categories based on the K-means clustering algorithm (such as sudden anomalies, gradual anomalies, and periodic anomalies), and recording the historical correlation probability between each type of anomaly pattern and watermark tampering (such as the correlation probability between sudden voltage drops in devices and local watermark tampering ≥ 0.7). During watermark health monitoring, the system compares the current asset IoT data with the anomaly pattern library in real time. If a certain anomaly pattern is matched and the correlation probability is higher than a preset threshold (such as ≥ 0.65), a deep watermark integrity verification is still triggered even if no watermark anomaly is detected (such as increasing the frequency of bit error rate detection of watermark data and expanding the frequency domain feature scanning range). This technology realizes the shift from "post-anomaly detection of watermarks" to "tampering prediction based on IoT data," proactively resisting watermark tampering during periods of abnormal physical state of assets.

[0029] In some embodiments, the cross-time-period scene consistency verification algorithm also incorporates attitude sensor data (such as acceleration and gyroscope data) from the shooting device. Specifically, the terminal device synchronously collects attitude sensor data (sampling frequency 10Hz) during shooting and extracts key features (such as the device tilt angle during shooting, the device movement trajectory before and after shooting, and the peak acceleration at the moment the shutter is pressed). The system constructs a "trustworthy attitude feature library" based on historical shooting attitude sensor data (such as historical shooting tilt angle fluctuation range ≤5° and movement trajectory conforming to physical movement laws). During scene consistency verification, the current shooting attitude sensor features are compared with the trusted attitude feature library. If the tilt angle fluctuation range is >8°, the movement trajectory has non-physical abrupt changes (such as instantaneous movement distance >1 meter), or the peak acceleration deviates from the historical average by ±30%, the scene is determined to be abnormal, and a manual review prompt is triggered simultaneously. This technology verifies the authenticity of the scene through the physical attitude data of the shooting device, resisting "fixed-position imitation shooting".

[0030] The technical concept of this invention is as follows: Core fusion algorithm construction: A multimodal temporal fusion dynamic risk prediction algorithm (AI dynamic risk assessment algorithm, the logical flow of which is as follows) Figure 4 ): (1) Algorithm objective: Risk assessment is “pre-judgment of the whole process”. Through the fusion of multimodal time series data, real-time risk prediction score and scenario adaptation parameter suggestions are output to guide other algorithms to dynamically adjust and realize risk prevention and control and precise resource allocation.

[0031] (2) Input data system ; (3) Algorithm model construction The Transformer + LightGBM hybrid model architecture is adopted, as follows: 1. Feature Preprocessing Layer: Numerical features (e.g., temperature, voltage, delay): Z-Score standardization is used, with the formula Xnorm=σX-μ (where μ is the mean and σ is the standard deviation); Temporal features (e.g., historical shooting time distribution, IoT data time series curves): a sliding window method (window size 7 days) is used to extract temporal statistical features (mean, variance, peak value); Categorical features (e.g., asset value level, scene type): one-hot encoding is used to convert them into vector form; Feature selection: based on mutual information (MI) value, features with MI≥0.3 are retained to reduce dimensionality redundancy.

[0032] 2. Transformer Temporal Feature Extraction Module: Model Structure: 6 encoder layers, 8 multi-head attention heads, 512 hidden layer dimensions; Position Encoding: Sine wave position encoding is used, the formula is... , (pos is the time-series position, i is the dimension index, (Feature dimension); Function: Capture long-term temporal dependencies in multimodal data (such as trend changes in asset IoT data, and periodic distribution of historical shooting time).

[0033] 3. LightGBM Feature Weight Optimization Module: Model parameters: learning rate 0.05, number of decision trees 100, maximum tree depth 8, number of leaf nodes 32; Loss function: adopts mean squared error (MSE) loss function, the optimization objective is to minimize the prediction error of risk prediction score; Function: based on the time series features extracted by Transformer, further optimize the weight of each feature to improve the accuracy of risk prediction.

[0034] 4. Output Layer: First Output: Real-time risk prediction score (0-100 points), obtained by normalization using the Sigmoid function, formula: Second output: Suggestions for scene adaptation parameters, including watermark embedding strength (0.2-0.4 range), GPS verification threshold (1-10 meters), and scene verification strength (high / medium / low).

[0035] (4) Model training process; 1. Dataset construction: Collect 500,000 multimodal time series samples (covering RWA types such as real estate, equipment, and commodities), each sample contains complete input data items and manually labeled risk levels (0-100 points); 2. Dataset partitioning: Divide the dataset into training set, validation set, and test set in a 7:2:1 ratio, and use stratified sampling to ensure that the samples of each RWA type are evenly distributed; 3. Training process: Pre-training: First, train the Transformer module independently, using the Adam optimizer (learning rate 0.001, decay rate 0.001). Pre-training is stopped when the accuracy of temporal feature extraction on the validation set is ≥95% (with a reduction rate of 0.0001) and 50 training rounds. Joint training involves training the pre-trained Transformer module and the LightGBM module together using a gradient descent optimizer (learning rate 0.05, momentum 0.9) for 100 training rounds, with validation every 10 rounds. Training is stopped when the risk prediction error on the validation set is ≤2%. Incremental learning automatically starts incremental training every 10,000 new samples to update model parameters and ensure that the model adapts to changes in business scenarios.

[0036] (5) Algorithm Application Process; 1. Data Acquisition: Collect various types of data in the input data system in real time and transmit them to the algorithm module through a TLS1.3 encrypted channel; 2. Feature Preprocessing: Standardize the data, extract temporal features, and filter features according to the above preprocessing rules; 3. Model Inference: Input the preprocessed feature vectors into the hybrid model and output real-time risk prediction scores and scenario adaptation parameter suggestions; 4. Parameter Distribution: Distribute the scenario adaptation parameter suggestions to modules such as watermark generation, scenario verification, and health monitoring in real time to guide them to dynamically adjust their operating parameters. The multi-scenario AI parameter adaptive adjustment process is as follows: Figure 5 As shown.

[0037] Active watermark health monitoring and early warning algorithm (1) Algorithm objective: To shift from “passive repair” to “active monitoring + early warning repair”, construct a watermark health index system, monitor the watermark status in real time, detect the attenuation of watermark caused by compression, transmission and slight tampering in advance, avoid watermark extraction failure in the verification stage, and provide feedback data for risk prediction algorithm.

[0038] (2) Input data system (3) ; (4) The algorithm model is constructed using a hybrid architecture of convolutional neural network (CNN) + Bayesian probability model, as detailed below: 1. CNN Feature Extraction Module; Model Structure: Input Layer (3×224×224 image tensor) → Convolutional Layer 1 (64 3×3 convolutional kernels, stride 1, padding with SAME) → Pooling Layer 1 (2×2 max pooling, stride 2) → Convolutional Layer 2 (128 3×3 convolutional kernels, stride 1, padding with SAME) → Pooling Layer 2 (2×2 max pooling, stride 2) → Convolutional Layer 3 (256 3×3 convolutional kernels, stride 1, padding with SAME) → Fully Connected Layer 1 (1024 dimensions) → Fully Connected Layer 2 (256 dimensions); Activation Function: ReLU function is used in the convolutional layers ( The fully connected layer uses the LeakyReLU function. Function: Extracts frequency domain features, spatial features, and local features of the watermark embedding region from real-time frames of watermarked photos.

[0039] 2. Construction of Watermark Health Index System: Invisibility SSIM value: Calculated using the Structural Similarity Index (SSIM), the formula is as follows: (Where x is the original photo, and y is the photo with watermark,) The mean, Standard deviation, (This is a constant), with a value range of 0-1, and ≥0.98 is considered excellent; Robustness score against attacks: Calculated based on the results of simulated attack tests, covering 70% compression of JPG, 15% cropping of edges, and Gaussian filtering (radius ≤ 2 pixels). Successful watermark extraction after each attack earns 33.3 points, with a total score of 0-100. Data integrity check value: The bit error rate (BER) between the extracted watermark binary data and the original on-chain evidence storage data, calculated as follows: The check value is 1 - BER × 100, and the range is 0-100 points.

[0040] 3. Bayesian probability model: Prior probability: Based on historical watermark health data, a prior probability distribution (normal distribution) is set for each indicator. Posterior probability calculation: Combining features extracted by CNN with watermark health index data, using Bayes' theorem... Calculate the posterior probability of watermark health. Health score fusion: The scores of the three indicators are fused using a weighted summation method, with the following weights: SSIM value 0.4, anti-attack robustness score 0.35, and data integrity check value 0.25. The final health score = 0.4 × SSIM value × 100 + 0.35 × anti-attack robustness score + 0.25 × data integrity check value (0-100).

[0041] (5) Model training process; Dataset construction: Collect 300,000 watermarked photo samples (including normal samples, compressed and attenuated samples, slightly tampered samples, and damaged transmission samples). Each sample contains complete input data items and manually labeled health scores; Dataset division: Divide the dataset into training set and validation set in an 8:2 ratio to ensure a balanced distribution of various types of samples; Training process: Pre-training: First train the CNN feature extraction module using the Adam optimizer (learning rate 0.0001, decay rate 0.00001), with 80 training rounds. Stop pre-training when the feature extraction accuracy of the validation set is ≥96%; Joint training: Jointly train the pre-trained CNN module with the Bayesian probability model. The optimization objective is to minimize the absolute error between the predicted health score and the manually labeled score. Train for 120 rounds, and validate every 15 rounds. Stop training when the average absolute error of the validation set is ≤1.5 points; Online calibration: For every 5,000 samples processed, calibrate the prior probability distribution of the Bayesian model based on the real health results to improve monitoring accuracy.

[0042] (6) Algorithm application process; Data acquisition: Collect real-time frames of watermarked photos and related input data according to dynamically adjusted sampling frequency; Feature extraction: Extract multi-dimensional features of photos and watermarks through CNN module; Health score calculation: Calculate real-time health score based on Bayesian probability model and health index system; Early warning judgment: Health score ≥80 is normal, 60-79 is early warning, <60 is serious abnormality; When early warning or serious abnormality occurs, trigger corresponding processing mechanism (such as parameter adjustment, secondary evidence storage, manual early warning); Data feedback: Feed back the health score and abnormality type to the dynamic risk prediction algorithm in real time for risk score correction.

[0043] Cross-time scenario consistency verification algorithm (1) Algorithm objective: Upgrade scene recognition from "single static verification" to "cross-time dynamic verification". By comparing the spatiotemporal characteristics of the current scene with those of historical scenes, resist advanced fake scenes such as "simulated scene shooting" and "reuse of historical real photos", and provide scene anomaly feedback for risk prediction algorithm.

[0044] (2) Input data system ; (3) The algorithm model is constructed using a hybrid architecture of Siamese network and temporal attention mechanism, as follows: 1. Siamese network spatial feature extraction module; Model structure: Two-branch shared weight architecture, each branch contains: Input layer (3×224×224 image tensor) → Convolutional layer 1 (64 3×3 convolutional kernels) → Pooling layer 1 (2×2 max pooling) → Convolutional layer 2 (128 3×3 convolutional kernels) → Pooling layer 2 (2×2 max pooling) → Convolutional layer 3 (256 3×3 convolutional kernels) → Fully connected layer (512-dimensional feature vector); Loss function: Triplet Loss function is used, the formula is as follows: (Where a is the anchor point sample, p is the positive sample, n is the negative sample, d is the Euclidean distance, and α is the marginal value, set to 0.5); Function: Extract spatial feature vectors (such as environment layout, asset location, and landmark distribution) between the current scene and historical scenes, and calculate spatial feature similarity.

[0045] 2. Temporal Attention Mechanism Module: Model structure: Input layer (temporal feature sequence, dimension 512×T, where T is the temporal length) → Attention weight calculation layer (using the Softmax function to calculate the weight of each temporal position) → Weighted summation layer (to obtain the temporal feature aggregation vector); Attention weight formula: ,in ( (The time-series feature is represented by MLP, which stands for Multilayer Perceptron). Function: Capture the matching degree between the current scene's temporal features (such as shooting time and light changes) and the historical scene's temporal feature curves, and assign higher weight to important temporal positions.

[0046] 3. Similarity fusion calculation: Spatial feature similarity: calculated using cosine similarity, the formula is as follows: ( This is the current scene space feature vector. (For historical scene spatial feature vectors). Temporal feature similarity: Based on the aggregated vector output by the temporal attention mechanism, it is calculated using Euclidean distance normalization, as shown in the formula below. ( This is the current scene time-series aggregation vector. This is a time-series aggregation vector of historical scenes. (for the maximum Euclidean distance). Overall similarity: A weighted summation method is used, with spatial feature similarity weighted at 0.55 and temporal feature similarity weighted at 0.45. The formula is as follows: (Value range: 0-1).

[0047] (4) Model training process; 1. Dataset construction: Collect 400,000 scene samples (including real scene samples, fake scene samples, and historical photo reuse samples). Each sample contains current scene data, historical scene data, and manually labeled similarity tags; 2. Dataset division: Divide the dataset into training set and validation set in a 7:3 ratio to ensure comprehensive coverage of various fake scene samples; 3. Training Process: Pre-training: First, train the Siamese network module using the SGD optimizer (learning rate 0.001, momentum 0.9) for 60 training epochs. Pre-training stops when the spatial feature similarity recognition accuracy on the validation set is ≥97%. Joint Training: Jointly train the pre-trained Siamese network with the temporal attention mechanism module. The optimization objective is to minimize the squared error between the comprehensive similarity prediction value and the manually labeled value. The training epochs are 90, with validation every 12 epochs. Training stops when the similarity prediction error on the validation set is ≤0.02. Scene Adaptation Optimization: Fine-tune the attention mechanism weights for different RWA types (real estate, equipment, bulk commodities) to improve the verification accuracy for specific scenarios.

[0048] (5) Algorithm application process; Data acquisition: Collect the current shooting scene preview image and asset IoT time series data, and retrieve the historical scene feature library of the corresponding asset from the database; Feature extraction: Extract the spatial feature vectors of the current and historical scenes through the Siamese network, and extract the temporal aggregation vector through the temporal attention mechanism; Similarity calculation: Combine spatial and temporal similarity to obtain a comprehensive similarity score; Scene determination: A comprehensive similarity ≥ 0.85 indicates scene consistency (real scene), and < 0.85 indicates scene abnormality (fake scene); Result feedback: Feedback the scene consistency score and abnormal temporal features to the dynamic risk prediction algorithm in real time for risk score correction and subsequent processing.

[0049] The core algorithm features a two-way bidirectional interaction mechanism. Dynamic risk prediction algorithm ↔ Active watermark health monitoring and early warning algorithm (1) Interaction 1: Dynamic risk prediction algorithm → Active watermark health monitoring and early warning algorithm The "real-time risk prediction score" output by the dynamic risk prediction algorithm determines the monitoring frequency, warning threshold, and monitoring depth of the watermark health monitoring algorithm. 1. High Risk (≥80 points): Monitoring frequency: increased from "sampling every 10 minutes" to "real-time frame-by-frame monitoring" (monitoring every frame when the terminal uploads, and monitoring one frame for every frame retrieved when the database is called); Warning threshold: increased from 80 points to 85 points (stricter health requirements); Monitoring depth: additional frequency domain feature stability verification (calculating the variance of the frequency domain features of the watermark in 5 consecutive frames, variance > 0.05 is judged as abnormal), and output "watermark protection priority" (divided into 3 levels according to asset value level, with level 1 being the highest), and watermarks of level 1 priority assets are monitored in a focused manner.

[0050] 2. Medium risk (60-79 points): Monitoring frequency: maintain "sampling every 5 minutes" (sampling 1 frame every 5 frames when the terminal uploads, and sampling 1 frame every 5 frames when the database is called); Warning threshold: maintain 80 points; Monitoring depth: routine monitoring, increase the frequency of watermark data integrity verification (an additional BER test is performed every 2 samplings).

[0051] 3. Low risk (<60 points): Monitoring frequency: Maintain "sampling every 10 minutes" (sampling 1 frame every 10 frames when the terminal uploads, and sampling 1 frame every 10 frames when the database is called); Warning threshold: Reduced to 75 points; Monitoring depth: Basic monitoring, only performing core health index calculation (SSIM value + anti-attack robustness score).

[0052] (2) Interaction 2: Active watermark health monitoring and early warning algorithm → Dynamic risk prediction algorithm The "watermark health score" output by the watermark health monitoring algorithm and the "anomaly type" inversely correct the risk prediction score of the dynamic risk prediction algorithm and trigger corresponding optimization suggestions: 1. Health score 75-79 (slight decay, no signs of tampering): Risk score correction: Dynamic risk prediction score +3 points; Optimization suggestion: Trigger "Watermark parameter optimization suggestion", guiding the dynamic risk prediction algorithm to increase the watermark embedding strength by 0.02-0.03 in the next parameter distribution.

[0053] 2. Health score 60-74 (moderate attenuation, no obvious signs of tampering): Risk score correction: Dynamic risk prediction score +8 points; Optimization suggestion: guide the dynamic risk prediction algorithm to output "watermark embedding depth adjustment suggestion" (expand the frequency domain range to the first 1 / 3), and trigger the "secondary encryption storage of photos" mechanism.

[0054] 3. Health score < 60 (severely abnormal): Subtype 1: No obvious signs of tampering, only severe attenuation (such as multiple compressions / transmission damage): Risk score correction: Dynamic risk prediction score +15 points; Optimization suggestion: Guide the dynamic risk prediction algorithm to output "watermark re-embedding suggestion", triggering the watermark generation module to re-execute the watermark embedding process for the photo.

[0055] Subtype 2: Precursors to tampering exist (such as abnormal local frequency domain features, sudden increase in BER): Risk score correction: Dynamic risk prediction score +20 points; Optimization suggestion: Directly trigger "secondary evidence storage + manual warning", and at the same time guide the dynamic risk prediction algorithm to include "watermark anomaly" as a high-weight feature in the next incremental training.

[0056] Dynamic risk prediction algorithm ↔ Cross-time scenario consistency verification algorithm (1) Interaction 1: Dynamic risk prediction algorithm → Cross-time scene consistency verification algorithm The "real-time risk prediction score" and "scene type" output by the dynamic risk prediction algorithm determine the verification strength, similarity threshold and verification dimension of the scene consistency verification algorithm: 1. High-risk + cross-border scenarios: Verification intensity: deep verification; similarity threshold: increased from 0.85 to 0.92; verification dimensions: basic dimensions (spatial + temporal features) + additional dimensions (historical GPS trajectory consistency verification: whether the movement trajectory of maintenance personnel in the past hour conforms to physical movement patterns, such as speed ≤120km / h; asset IoT status consistency verification: the deviation between the asset's operating status at the time of shooting and the historical status at the same time is ≤10%).

[0057] 2. High-risk + regular scenarios (non-cross-border): Verification intensity: deep verification; similarity threshold: increased to 0.88; verification dimensions: basic dimensions + asset IoT status consistency verification.

[0058] 3. Medium-risk + cross-border scenarios: Verification intensity: standard verification; similarity threshold: maintain 0.85; verification dimensions: basic dimensions + historical GPS trajectory consistency verification (speed ≤150km / h).

[0059] 4. Medium-risk + standard scenarios: Verification strength: standard verification; Similarity threshold: maintain 0.85; 5. Verification Dimensions: Basic Dimensions. Low Risk (Regardless of Scenario Type): Verification Strength: Lightweight Verification; Similarity Threshold: Reduced to 0.80; Verification Dimensions: Simplified Dimensions (Compare only core spatial features, such as asset identification locations and key environmental markers).

[0060] (2) Interaction 2: Cross-time period scenario consistency verification algorithm → dynamic risk prediction algorithm The scenario consistency verification algorithm outputs "scenario consistency score" and "abnormal time series characteristics," which are used to inversely optimize the prediction model and risk score of the dynamic risk prediction algorithm. 1. Scene consistency score 0.80-0.84 (minor anomaly): Risk score correction: Dynamic risk prediction score +10 points; Model optimization: Incorporate "minor temporal anomaly" as a new feature into the incremental training of the dynamic risk prediction model, and adjust the weight of this feature to 0.05.

[0061] 2. Scene consistency score 0.70-0.79 (moderate anomaly, such as the current shooting time deviating from the historical peak shooting time by more than 3 hours): Risk score correction: Dynamic risk prediction score +15 points; Model optimization: Incorporate "moderate temporal anomaly" as a high-weight feature (weight 0.1) into the model training, and trigger "manual scene review prompt".

[0062] 3. Scene consistency score < 0.70 (severe anomaly, such as the same asset being photographed in two different locations within 1 hour, with completely mismatched spatial features): Risk score correction: Dynamic risk prediction score +30 points; Model optimization: Mark "severe scene anomaly" as a core risk feature (weight 0.2) and trigger the "photo upload interception + secondary verification of operation and maintenance personnel identity" mechanism, while updating the scene type identification parameters of the dynamic risk prediction model.

[0063] 4. Scene consistency score ≥ 0.95 (highly consistent, with highly overlapping scenes in 3 consecutive shots): Risk score correction: Dynamic risk prediction score -10 points; Model optimization: Marked as "trustworthy scene", the scene verification frequency for this asset will be reduced by 50% in the future, and "trustworthy scene identifier" will be included in the model as a negative risk feature.

[0064] Active watermark health monitoring and early warning algorithm ↔ Cross-time period scene consistency verification algorithm (1) Interaction 1: Active watermark health monitoring and early warning algorithm → Cross-time period scenario consistency verification algorithm When the watermark health monitoring algorithm detects a "local watermark anomaly" (such as a sudden decrease in watermark intensity or an increase in local BER in a certain area), it triggers the "directional temporal verification" mechanism of the scene consistency verification algorithm: 1. Anomaly Location: Precisely locate the coordinates of areas with localized watermark anomalies (such as equipment and instrument areas, or areas near property ownership certificates). 2. Data retrieval: Trigger the scene consistency verification algorithm to retrieve scene data from the last 10 shots of the asset, focusing on extracting historical scene features corresponding to abnormal areas (such as the location, lighting, and texture features of the area in historical shots). 3. Targeted Verification: Spatial Feature Targeted Comparison: Focuses on comparing the spatial feature similarity between abnormal areas in the current scene and corresponding areas in historical scenes (such as the positional offset of asset identifiers and texture consistency); Temporal Feature Targeted Analysis: Analyzes the temporal change trend of abnormal areas (such as whether the changes in lighting conform to the historical patterns of the same period, and whether there are abrupt changes in the shooting angle). 4. Result Determination: If the abnormal areas in the current scene and the historical scene highly overlap (similarity < 0.75), it is directly determined as "targeted tampering," and the "deep repair" of the watermark health monitoring algorithm and the "historical scene source tracing report" of the scene consistency verification algorithm are triggered simultaneously. If the abnormal area and the features of the historical scene do not have obvious conflicts (similarity ≥ 0.75), it is determined as "natural watermark decay," and the "local watermark repair" process of the watermark health monitoring algorithm is triggered.

[0065] (2) Interaction 2: Cross-time period scene consistency verification algorithm → Active watermark health monitoring and early warning algorithm When the scene consistency verification algorithm detects "scene temporal anomalies" (such as shooting the same asset in different locations within a short period of time, or a sudden change in the temporal characteristic curve), it triggers the "watermark integrity depth detection" of the watermark health monitoring algorithm. 1. Anomaly Notification: The scene consistency verification algorithm synchronizes the scene anomaly type and anomaly time sequence characteristic data to the watermark health monitoring algorithm; 2. Deep Detection Initiation: For photos taken in abnormal scenes, the watermark health monitoring algorithm performs the following additional checks: Deep verification of watermark data integrity: Compare the bit error rate (BER) of the on-chain stored watermark binary data with the currently extracted watermark data, improving the detection accuracy to 0.01%; Panoramic analysis of frequency domain features: Perform frequency domain feature scanning on the entire area of ​​the photo to detect whether there are traces of local tampering or replacement of the watermark; Multi-frame watermark comparison: If there are multiple photos of the same abnormal scene, compare the consistency of watermark features at the same location in different photos. 3. Result Feedback: If the bit error rate is >0.5%, the watermark health monitoring algorithm directly issues a warning that "the watermark has been tampered with" and feeds back the coordinates of the error area to the scene consistency verification algorithm. The scene consistency verification algorithm combines the error area with scene anomaly features to help locate "flaws in the counterfeit scene" (such as the lighting of the fake scene being inconsistent with the lighting when the watermark is embedded, or the abnormal area overlapping with the key forgery area of ​​the counterfeit scene). Joint Output: The two jointly generate a composite anomaly report of "watermark tampering + scene fraud", which is synchronized to the dynamic risk prediction algorithm and the back-end management system.

[0066] Basic functional module optimization Precise Scene Trigger Mechanism: A pre-defined "Asset Status Update Trigger Rule" is established, combining the initial scene identification of the cross-time-period scene consistency verification algorithm with the risk prediction results of the dynamic risk prediction algorithm to achieve dual precise triggering: 1. Initial Trigger: When maintenance personnel initiate an operation request, the terminal device collects a scene preview image, and the cross-time-period scene consistency verification algorithm performs preliminary scene identification (confidence ≥ 0.9 indicates a suspected real scene); 2. Risk Prediction: The dynamic risk prediction algorithm outputs a preliminary risk prediction score based on the initial scene data and basic asset information; 3. Final Trigger: The photo upload process is initiated only when "preliminary cross-time-period scene consistency verification passes + preliminary risk prediction score < 80 points (no serious risk)"; if the preliminary risk prediction score ≥ 80 points, secondary verification of the maintenance personnel's identity is required before initiating the upload process.

[0067] Multi-dimensional core information collection and verification: Based on the three core information items (precise timestamp, real-time GPS, and maintenance personnel ID), real-time data collection of asset IoT is added, and the verification process incorporates the following parameter suggestions for dynamic risk prediction algorithms: 1. Information collection: Real-time data of asset IoT (equipment temperature / voltage, real estate temperature and humidity, etc.) is collected synchronously with the core information; 2. Information verification: Based on the GPS verification threshold output by the dynamic risk prediction algorithm, the matching degree between the GPS location and the asset registration address is verified (e.g., a threshold of 10 meters for cross-border scenarios and 5 meters for regular scenarios); the reasonableness of the timestamp is verified in conjunction with the asset IoT data (e.g., the shooting time when the device is not running needs to be verified).

[0068] AI Adaptive Hidden Watermark Generation Algorithm like Figure 2 The watermark embedding parameters are guided by the scenario adaptation parameter suggestions of the dynamic risk prediction algorithm, and optimized by the attention mechanism and reinforcement learning algorithm: 1. Embedding parameter initialization: Based on the watermark embedding strength and frequency range suggestions output by the dynamic risk prediction algorithm, the embedding parameters are initialized; 2. Non-critical area localization: Non-critical areas are located through the CBAM attention mechanism to avoid occluding core details; 3. Parameter dynamic adjustment: The reinforcement learning agent combines the real-time feedback of the watermark health monitoring algorithm (such as SSIM value and anti-attack score) to further optimize the embedding parameters to ensure that the watermark health meets the standards; 4. Watermark embedding: Based on the improved DCT algorithm, binary watermark data (including core information and algorithm interaction parameter identifiers) is embedded in the low-frequency coefficients of non-critical areas.

[0069] On-chain hash notarization and data synchronization mechanism: 1. Enhanced notarization data: Notarization data, the output results of the three core algorithms (risk prediction score, health score, and scenario consistency score), and interaction logs are packaged into a structure to generate a notarization serial number; 2. Multi-node notarization optimization: The number of notarization nodes is adjusted based on the dynamic risk prediction score. For high risk (≥80 points), consensus is adopted with ≥5 nodes, and consensus is adopted with ≥3 nodes for medium and low risk; 3. Encrypted data storage: Watermarked photos are stored using AES-256 encryption. The encryption key is bound to the watermark health score. The key is automatically updated when the health score is <60 points.

[0070] AI two-way verification and risk control mechanism: such as Figure 3 Combining three core algorithms and interaction results, the verification process and risk control are optimized: 1. Verification process: First, AI-synthesized photos are filtered through the GAN fake photo detection algorithm, and then the watermark is extracted through the AI ​​watermark repair algorithm. Combined with the watermark health monitoring results and scene consistency verification results, double verification is performed; 2. Risk control: Based on the risk level of the dynamic risk prediction algorithm, a differentiated verification strategy is implemented (high risk: manual review + multi-node secondary evidence storage; medium risk: hash secondary comparison; low risk: direct pass).

[0071] This invention addresses the shortcomings of insufficient collaboration and lack of creativity among existing AI algorithms by constructing three core fusion algorithms and a two-way bidirectional interaction mechanism. It forms a technical system of "algorithm fusion - interactive linkage - full-process adaptation - closed-loop risk control," possessing the following significant advantages: 1. Outstanding algorithmic creativity: Breaking through the limitations of conventional combinations of AI algorithms in existing technologies, three types of deeply integrated algorithms are constructed, and two-way interaction and closed-loop linkage are achieved. The algorithms guide each other and make reverse corrections, forming a synergistic and efficient technical effect, and significantly enhancing creativity; 2. Comprehensive upgrade in anti-tampering and fake detection capabilities: The system employs a cross-time-period scenario consistency verification algorithm to resist advanced fake scenarios (such as historical photo reuse and fake on-site shooting), a watermark health monitoring algorithm to detect early signs of tampering, and a dynamic risk prediction algorithm for proactive prevention. Combined with on-chain hash evidence storage, the tampering identification rate reaches 99.95%, and the fake photo interception rate reaches 99.8%. 3. Dual guarantee of watermark health and robustness: The proactive watermark health monitoring algorithm realizes the transformation from "passive repair" to "proactive early warning", with a watermark health compliance rate of ≥99.5%. Combined with the parameter guidance of the dynamic risk prediction algorithm, the watermark invisibility rate is 100%, and the anti-attack capability is further improved (supports 60% JPG compression, 20% edge cropping, and Gaussian filter radius ≤3 pixels). 4. Significantly enhanced scenario adaptability and robustness: The cross-time scenario consistency verification algorithm enables dynamic verification across time periods, and the dynamic risk prediction algorithm outputs scenario adaptation parameters based on multimodal data. The three types of algorithms interact to adapt to complex scenarios (cross-border networks, changes in asset status), achieving a multi-scenario adaptability of 99.8%. 5. Refined and proactive risk management: The dynamic risk prediction algorithm advances risk assessment and, combined with algorithmic feedback, dynamically adjusts risk scores, enabling differentiated risk management strategies (strict verification for high-risk, rapid approval for low-risk), reducing manual intervention rate to 0.1% and improving verification efficiency by 60%. 6. Optimized resource utilization efficiency: Based on precise triggering and dynamic resource allocation through algorithmic interaction, indiscriminate processing is avoided, resulting in a 50% improvement in system operating efficiency and a 40% reduction in storage resource consumption; 7. Comprehensive improvement in compliance and security: end-to-end data encryption storage, on-chain evidence storage, and algorithm interaction log retention form a complete chain of evidence, achieving a compliance audit pass rate of 99.9%, meeting the requirements of the Data Security Law, the Personal Information Protection Law, and cross-border business compliance. 8. Low operating threshold and strong scalability: The AI ​​algorithm runs automatically throughout the process, and maintenance personnel do not need professional knowledge; the algorithm architecture supports the addition of new AI modules and interaction logic, and can be adapted to more RWA asset types and business scenarios (such as art collateral and intellectual property rights confirmation). Detailed Implementation

[0072] Initial basic configuration: 1. Asset Information and Multimodal Data System Construction: Establish an RWA asset information database, input core fields (asset ID, name, type, registration address, value level, etc.) to form a four-dimensional association mapping of "asset-maintenance personnel-registration information-IoT devices"; connect asset IoT sensors (supporting MQTT protocol), configure data collection frequency (1 time / second for equipment, 1 time / minute for real estate); input visual feature samples of various assets (such as equipment appearance, real estate identification) and historical scene data (last 6 months) as training datasets for AI algorithms.

[0073] 2. Interface Integration and Authorization: Complete the integration of seven types of core interfaces (shooting equipment, blockchain nodes, encrypted database, NTP server, third-party data interface, IoT sensor interface, network quality detection interface) to ensure an interface call success rate of ≥99.9%; integrate with AI algorithm inference engine (supporting TensorFlow / PyTorch model deployment) and configure model inference resources (GPU ≥16GB, inference latency ≤300ms).

[0074] 3. Algorithm Model and Parameter Configuration: Three core fusion algorithm models are deployed: a multimodal temporal fusion dynamic risk prediction model (Transformer + LightGBM), an active watermark health monitoring and early warning model (CNN + Bayesian probability), and a cross-time-period scene consistency verification model (Siamese network + temporal attention mechanism). Preset basic algorithm parameters: Dynamic risk prediction model: 6 Transformer encoder layers, 100 LightGBM decision trees, learning rate 0.05; Watermark health monitoring model: 64 / 128 / 256 CNN convolutional kernels, Bayesian prior probability normal distribution parameters (μ=85, σ=5); Scene consistency verification model: 512 dimensions of the fully connected Siamese network, temporal attention mechanism window size 7 days; Preset interaction mechanism parameters: monitoring frequency threshold, similarity threshold, risk score correction coefficient, etc.

[0075] Algorithm training and optimization: 1. Training environment preparation: Hardware environment: GPU cluster (NVIDIA A100×4, 80GB VRAM / card), CPU 32 cores, 128GB RAM, 10TB storage; Software environment: Python 3.9+, TensorFlow 2.10+, PyTorch 1.13+, Scikit-learn 1.2+, LightGBM 3.3+.

[0076] 2. Model Training Execution: Following the training process of each algorithm, pre-training and joint training of the three core algorithms are completed in sequence to ensure that the validation set indicators meet the standards (dynamic risk prediction error ≤2%, watermark health monitoring average absolute error ≤1.5 points, scene consistency verification error ≤0.02); execute algorithm interaction and joint debugging to verify the correctness and response speed of the pairwise interaction logic (interaction response delay ≤100ms).

[0077] 3. Online incremental training configuration: Set incremental training trigger conditions: start when every 10,000 new samples are accumulated or the model prediction error is >3%; Configure dynamic allocation strategy for training resources: incremental training occupies no more than 50% of GPU resources to avoid affecting online business.

[0078] Full process implementation steps Scene Triggering and Photo Acquisition: 1. Operation Request Initiation: Maintenance personnel initiate an asset status update request via terminal device (smartphone iOS 14.0+ / Android 10.0+ or ​​industrial inspection terminal with IP67 protection); 2. Initial Scene Verification: The terminal device automatically captures a scene preview image, and a cross-time period scene consistency verification algorithm performs preliminary scene recognition, outputting spatial feature similarity; 3. Risk Pre-judgment: A dynamic risk prediction algorithm collects basic risk features and real-time asset IoT data, outputting a preliminary risk prediction score; 4. Final Trigger Judgment: If "scene similarity ≥ 0.9 + preliminary risk prediction score < 80": the photo capture process is initiated; if "preliminary..." "Risk prediction score ≥ 80 points": Maintenance personnel are required to undergo secondary identity verification (facial recognition + employee ID verification). After successful verification, the shooting process will begin. 5. Photo shooting and preprocessing: Maintenance personnel take asset photos on-site (single / batch shooting is supported). The terminal device automatically completes the cropping of black borders and verifies the clarity (≥ 1080P). The photos are uploaded in segments through a TLS1.3 encrypted channel, supporting breakpoint resume. 6. GAN fake photo detection: The AI ​​inference engine starts the improved StyleGAN2 detection model to extract multi-dimensional features from the uploaded photos. Photos with a confidence score ≥ 0.9 are judged to be real photos and proceed to the next process; photos with a confidence score < 0.9 are required to be reshot (up to 3 times).

[0079] AI Adaptive Hidden Watermark Generation: 1. Core Information Collection and Verification: The system collects precise timestamps (UTC millisecond level), real-time GPS (latitude and longitude to 6 decimal places), operator IDs, and asset IoT data. Information consistency is verified based on the GPS verification threshold output by the dynamic risk prediction algorithm. 2. Watermark Parameter Initialization: The dynamic risk prediction algorithm outputs scenario-adaptive parameter suggestions (watermark embedding strength, frequency range) as initial values ​​for the embedding parameters. 3. Non-critical Area Location: The photo is preprocessed using the CBAM attention mechanism (RGB to YUV, retaining the Y channel) to generate an attention weight map, marking non-critical areas with pixel values ​​< 0.6. 4. Embedding Parameter Optimization: The reinforcement learning agent, combined with real-time feedback from the watermark health monitoring algorithm (SSIM value, anti-attack score), iterates three times to optimize the embedding parameters (k value 0.2-0.4, frequency range first 1 / 4-1 / 3). 5. Watermark Embedding Execution: An improved DCT algorithm is used to embed binary watermark data (core information + algorithm interaction parameter identifier) ​​into the low-frequency coefficients of non-critical areas. The embedding formula is... (C is the original coefficient, W is the watermark data); 6. Watermark validity verification: The system calls the AI ​​watermark extraction model to extract the watermark and compare it with the original information (error ≤ 0.1%). At the same time, a simulated attack test is performed (JPG 60% compression + edge 20% cropping). If the watermark extraction accuracy is ≥ 95%, it is judged as valid.

[0080] Algorithm interaction execution and on-chain evidence storage: 1. Algorithm Interaction: The dynamic risk prediction algorithm outputs real-time risk prediction scores and scene adaptation parameters to guide the operation parameters of the watermark health monitoring algorithm and the scene consistency verification algorithm; the watermark health monitoring algorithm monitors the watermark status in real time and feeds back the health score to the dynamic risk prediction algorithm to correct the risk score; the scene consistency verification algorithm performs cross-time period scene consistency verification and feeds back the scene consistency score to the dynamic risk prediction algorithm and the watermark health monitoring algorithm. 2. Packaging of evidence storage data: Package the hash value of the watermarked photo (SHA256 algorithm), asset ID, original watermark information, output results of the three major algorithms (risk prediction score, health score, and scene consistency score), and algorithm interaction logs into a structure to generate an evidence storage serial number; 3. On-chain storage: After signing via the blockchain node interface (Hyperledger Fabric 2.4+ / Ethereum Geth 1.10+), the data is sent to the blockchain network. High-risk (≥80 points) data requires consensus from ≥5 nodes, while medium-low risk data requires consensus from ≥3 nodes. Once verified, the data is written into a block, and the transaction hash and block height are returned.

[0081] Data Synchronization and Traceability Log Generation: 1. Structured Data Encryption Storage: Synchronizes watermarked photos (AES-256 encryption), watermark information, on-chain evidence storage information, algorithm processing results, and interaction logs to an encrypted database (MySQL 8.0+ + Redis 6.2+), establishing associated indexes; 2. Traceability Log Generation: Logs contain "evidence storage serial number + asset information + shooting information + algorithm processing details + algorithm interaction log + on-chain evidence storage information + verification result," with each log generating a SHA256 hash value to ensure immutability; 3. Multi-dimensional Query and Export: Supports multi-dimensional queries by asset ID, evidence storage serial number, risk score, etc., exporting reports in Excel / PDF format and automatically adding electronic signatures.

[0082] AI Two-Way Verification and Risk Management Execution: 1. Verification Application Submission: When a third-party institution or on-chain contract executes, it submits the photo to be verified, asset ID, and notarization serial number; 2. Permission Verification: Verifies whether the verifier has the corresponding asset verification permission. If not, a prompt is output; 3. Algorithm Collaborative Verification: GAN Fake Photo Detection: Re-detects the authenticity of the photo to be verified; AI Watermark Repair: If the photo damage ratio is ≤20%, the watermark is repaired based on the U-Net semantic segmentation model and Transformer feature completion technology; Watermark Information Comparison: Extracts the repaired watermark information and compares it with the on-chain notarization information (ID consistent, timestamp error ≤30 seconds, GPS error ≤5 meters); Tampering Location Detection: Based on differential hashing and the Siamese network, the tampered area is located; Hash Comparison: Calculates the hash value of the photo to be verified and compares it with the on-chain notarization; 4. Risk Level Judgment and Handling: High risk (≥80 points): Automatically triggers manual review + multi-node secondary evidence storage; only after the review is passed can it be determined as valid; Medium risk (60-79 points): Initiates on-chain hash secondary comparison to ensure the accuracy of the result; Low risk (<60 points): Directly passes verification; 5. Verification report generation: Includes "verification time + verifier information + algorithm processing details + algorithm interaction results + verification results + on-chain evidence storage details", supports multi-language export.

[0083] Implement environmental preparation Hardware environment: ; Software environment: ; Third-party dependencies: Location services: Gaode / Baidu / Google Maps API (GPS positioning accuracy ≤ 3 meters); AI model services: Third-party AI platforms that support online model fine-tuning and inference acceleration (optional); Compliance tools: Electronic signature services, log auditing tools; Network quality detection services: Third-party APIs that support real-time detection of latency and packet loss rate.

[0084] Example Test Data Example 1: Real Estate RWA Asset Inspection Scenario Implementation Process: 1. Input asset information for 500 office buildings, connect to real estate temperature and humidity IoT sensors, upload visual samples such as building exteriors and property ownership certificates, and historical scene data from the past 6 months, and complete the training and interactive debugging of the three core algorithms; 2. Configure algorithm interaction parameters: IoT data weight of 0.25 for the dynamic risk prediction algorithm, warning threshold of 80 points for the watermark health monitoring algorithm, and similarity threshold of 0.85 for the scene consistency verification algorithm; 3. 30 inspection personnel take inspection photos (3 photos per building) through a mobile APP to simulate some fake scenarios (such as taking network pictures from the office, reusing historical photos) and watermark attenuation scenarios (multiple compression and transmission); 4. Start the algorithm fusion interaction process, the dynamic risk prediction algorithm outputs risk scores and adaptation parameters, the watermark health monitoring algorithm monitors the watermark status in real time, and the scene consistency verification algorithm performs cross-time period verification.

[0085] Implementation Results: 1. Algorithm Interaction Response Speed: Average interaction latency of 85ms, meeting real-time requirements; 2. Scene Recognition Accuracy: 99.9%: Successfully filtered 15 fraudulent operation requests (including 8 historical photo reuse requests), with a false trigger rate of 0.03%; 3. Watermark Health Compliance Rate: 99.6%: Early warning of watermark decay in 12 photos, avoiding extraction failure during the verification stage, with a 100% watermark invisibility rate; 4. Inspection Efficiency Improvement: 50%: Inspection time for a single office building was reduced from 15 minutes to 7.5 minutes, with a manual intervention rate of 0.1%; 5. Compliance Audit Pass Rate: 99.9%: Complete algorithm interaction logs and evidence chains meet audit requirements.

[0086] Example 2: Industrial Equipment RWA Asset Inspection Scenario Implementation Process: 1. Input asset information for 2000 production equipment units, connect to equipment temperature and voltage IoT sensors, upload visual samples of equipment instruments and key components, and optimize algorithm interaction parameters (equipment photo texture feature weight 0.4). 2. Maintenance personnel take photos of the equipment (2 photos per unit) through industrial inspection terminals, simulating scene anomalies (photos taken at different locations within 1 hour on the same equipment) and partial watermark tampering (modifying the watermark in the instrument area). 3. The watermark health monitoring algorithm detects local watermark anomalies, triggering the directional time-series verification of the scene consistency check algorithm; the scene consistency check algorithm detects scene anomalies, triggering the deep detection of the watermark health monitoring algorithm.

[0087] Implementation Results: 1. 100% Anomaly Collaborative Recognition Rate: All 100 simulated anomaly data points (50 with partial watermark tampering + 50 with scene anomalies) were accurately identified without any missed or false positives; 2. 99.7% Accuracy in Tampering Location: The tampered areas of all 50 simulated tampered photos were accurately located (error ≤ 3 pixels), and a composite anomaly report was generated based on scene anomaly features; 3. 99.5% Success Rate in Watermark Restoration: The complete watermarks were successfully extracted from all 50 damaged photos (cropped by 20% + oil stains), which is 14.5% higher than traditional methods; 4. 65% Improvement in Verification Efficiency for Financial Institutions: The verification time for a single photo was reduced from 3 seconds to 1.05 seconds, and the review cycle was shortened by 40%.

[0088] Example 3: Cross-border Commodity RWA Collateralization Scenario Implementation Process: 1. Input information on 1000 container assets and configure cross-border scenario parameters for the dynamic risk prediction algorithm (GPS verification threshold of 10 meters, network quality data weight of 0.15); 2. Port maintenance personnel take photos of the containers (2 photos per container) to simulate a cross-border network environment (latency of 800ms, packet loss rate of 15%) and advanced fake scenarios (imitating on-site port photography); 3. The dynamic risk prediction algorithm outputs risk scores and adaptation parameters based on multimodal data (IoT data, network quality data, historical time-series data), guiding the watermark health monitoring algorithm to increase monitoring frequency and the scenario consistency verification algorithm to increase the similarity threshold; 4. International financial institutions verify through the system, and the algorithm automatically completes interactive verification, generating a multilingual verification report.

[0089] Implementation Results: 1. Cross-border scenario adaptability: 99.8%; average latency between photo upload and algorithm interaction was 1.1 seconds, with no data loss, and a watermark health compliance rate of 99.4%; 2. Advanced fake scenario interception rate: 100%; 12 counterfeit on-site photos were successfully intercepted, and the overall similarity score of the scenario consistency verification algorithm was <0.70; 3. Dynamic risk prediction accuracy: 99.6%; high-risk containers (value ≥10 million RMB) triggered secondary verification, while low-risk containers passed quickly, improving verification efficiency by 70%; 4. International verification cycle shortened by 75%: from 15 days to 3.75 days, multilingual reports had no communication barriers, and the compliance pass rate was 99.9%.

[0090] Experimental data Core Algorithm Performance Comparison Test (Test Sample: 10,000 photos of various assets): ; Experimental data compared with existing technologies: ; High concurrency and extreme scenario test data: High-concurrency scenario test (16-core 32GB server, 50,000 photos uploaded per hour): ; Extreme scenario testing: ; Long-term operational stability data (180 consecutive days): .

Claims

1. A method for generating watermarks by uploading photos of RWA assets under the physical state chain, characterized in that, include: S1. Initiate an RWA asset status update request, collect scene preview data and multimodal data. Multimodal data includes asset operation status data, asset historical time series data, and network environment data. Combined with preset asset basic information, which includes asset identification information and associated asset historical data, output scene similarity results through scene consistency verification algorithm, and output preliminary risk assessment results and scene adaptation parameters through multimodal risk prediction algorithm. S2. Trigger shooting based on scene similarity results and preliminary risk assessment results, acquire asset photos and preprocess them. After verification by the fake photo detection algorithm, embed watermark data based on scene adaptation parameters through an adaptive invisible watermark generation algorithm to obtain watermarked photos. The watermark data includes asset identification information and real-time collected related shooting information, and retains the original watermark information simultaneously. S3. Initiate bidirectional interaction and iterative optimization of the multimodal risk prediction algorithm, watermark health monitoring algorithm, and scene consistency verification algorithm, and record iterative interaction logs simultaneously. The multimodal risk prediction algorithm outputs real-time risk assessment results and scene adaptation parameters to guide the other two types of algorithms to adjust their running parameters and optimize watermark generation and embedding parameters. The other two types of algorithms output watermark health status, scene consistency results, and abnormal information based on watermarked photos, original watermark information, asset historical data, and real-time risk assessment results, respectively. They then reverse-correct the risk assessment results and optimize the model parameters. The watermark generation, health monitoring, and scene consistency verification are repeatedly executed until the results meet the standards or reach the preset iteration limit, resulting in optimized watermarked photos. The optimized output results are then summarized. S4. Perform hash calculation on the optimized watermarked photo to obtain the hash value, package the hash value, asset identification information, original watermark information, optimized output results and iteration interaction log, generate a notarization serial number and complete on-chain notarization through a blockchain node.

2. The method for generating watermarks by taking photos and uploading them under the RWA asset physical state chain according to claim 1, characterized in that, Also includes: S5. The verifier submits materials to be verified, including photos to be verified, asset identification information, and notarization serial number. Through the AI ​​two-way verification mechanism, combined with on-chain notarization information, algorithm iteration logic, and optimized output results, the verifier completes the verification of authenticity, watermark integrity, and scenario consistency. The verifier implements differentiated strategies with reference to the final risk assessment results and outputs a verification report containing verification conclusions and parameter optimization basis.

3. The method for generating watermarks by taking photos and uploading them under the RWA asset physical state chain according to claim 1, characterized in that, The bidirectional interaction and iterative optimization of the multimodal risk prediction algorithm, watermark health monitoring algorithm, and scene consistency verification algorithm are as follows: the higher the real-time risk assessment result output by the multimodal risk prediction algorithm, the higher the monitoring dimension of the watermark health monitoring algorithm and the verification strength of the scene consistency verification algorithm, and the higher the optimization frequency of the watermark generation embedding parameters. When the watermark health monitoring algorithm outputs watermark anomaly information and the scene consistency verification algorithm outputs scene anomaly information, the abnormal feature data is simultaneously fed back to the multimodal risk prediction algorithm to correct the risk assessment result and perform weighted iterative optimization of the model parameters of the multimodal risk prediction algorithm, and simultaneously trigger the corresponding level of watermark generation re-execution action.

4. The method for generating watermarks by taking photos and uploading them under the RWA asset physical state chain according to claim 1, characterized in that, The multimodal risk prediction algorithm generates risk assessment results and scenario adaptation parameters based on feature fusion and temporal correlation analysis of multimodal data. After standardizing the features and normalizing the dimensions of asset operation status data, asset historical time-series data, and network environment data, the algorithm extracts multi-dimensional temporal correlation features and asset risk features, and performs risk weight allocation by combining the asset historical data in the asset basic information. The algorithm outputs the corresponding preliminary risk assessment results, real-time risk assessment results, and scenario adaptation parameters adapted to different risk levels.

5. The method for generating watermarks by taking photos and uploading them under the RWA asset physical state chain according to claim 1, characterized in that, The scene consistency verification algorithm extracts scene spatial features and temporal features based on scene preview data, scene information in watermarked photos, and historical asset data, and performs feature matching calculations to output scene similarity results and scene consistency results. During the iterative optimization process, the algorithm adjusts the weight ratio of feature matching according to the real-time risk assessment results. The higher the risk assessment result, the higher the weight of temporal feature matching of historical asset data.

6. The method for generating watermarks by taking photos and uploading them under the RWA asset physical state chain according to claim 1, characterized in that, The watermark health monitoring algorithm comprehensively detects the integrity, anti-tampering ability, and visual imperceptibility of watermark data based on watermarked photos and original watermark information, and outputs watermark health status and abnormal information. During the iterative optimization process, the algorithm adjusts the judgment threshold of watermark health status according to the real-time risk assessment results. The higher the risk assessment result, the more stringent the judgment threshold. In addition, abnormal information is synchronously associated with the corresponding regional features of the watermarked photo and fed back to the watermark generation stage.

7. The method for generating watermarks by taking photos and uploading them under the RWA asset physical state chain according to claim 1, characterized in that, The adaptive invisible watermark generation algorithm embeds watermark data in non-critical feature areas of asset photos based on scene adaptation parameters. The asset identification information in the watermark data is extracted from the basic asset information, and the associated shooting information is the shooting time and shooting environment data collected synchronously when the asset photo was taken. During the iterative optimization process, the algorithm adjusts the embedding position, embedding depth and encryption level of the watermark data according to the parameter optimization instructions output by the multimodal risk prediction algorithm, and generates optimized watermarked photos.

8. The method for generating watermarks by taking photos and uploading them under the RWA asset physical state chain according to claim 1, characterized in that, The fake photo detection algorithm extracts image authenticity features and identifies forgery traces in the preprocessed asset photos. The detection dimensions include image pixel texture features, color distribution features, synthetic edge features, and asset physical feature matching degree. Only when the results of each detection dimension meet the preset authenticity standard is the asset photo determined to be verified and enter the watermark generation process.

9. The method for generating watermarks by taking photos and uploading them under the RWA asset physical state chain according to claim 1, characterized in that, The optimized watermarked photos are subjected to an irreversible one-way hash operation. The generated hash value is bound to the asset identification information before being included in the packaged data. The on-chain notarization involves synchronously uploading all the packaged data to multiple blockchain nodes to complete distributed notarization. The notarization serial number forms a unique correspondence with the hash value and asset identification information.

10. The method for generating watermarks by taking photos and uploading them under the RWA asset physical state chain according to claim 2, characterized in that, The differentiation strategy involves classifying risk levels based on the final risk assessment results in the optimized output. High-risk photos undergo secondary verification by combining historical asset data and multimodal data. Medium-risk photos undergo verification by checking the watermark integrity and scene consistency. Low-risk photos only require basic hash value comparison and watermark extraction verification. The parameter optimization in the verification report is based on the watermark generation parameters, algorithm running parameters adjustment records, and optimization results recorded in the iteration interaction log.

Images