Orthogonalized visual fingerprint and lightweight matching based generative ai data chain provenance method, system and medium

Abstract

The application discloses a generative AI data chain provenance method and system based on orthogonal visual fingerprint and lightweight matching, and a medium, comprising: extracting initial features by using a visual model; zero-meaning and singular value decomposition are performed on the initial features to obtain an initial projection matrix; the initial features are reduced to feature vectors of a target dimension by using the orthogonal projection matrix; an orthogonal visual fingerprint in the form of an integer array is generated; the similarity of query image fingerprints and database image fingerprints is calculated, and a candidate image set is screened; multi-scale features of candidate images and query images are extracted and a reliability graph is constructed, and low-confidence key points are filtered out; local matching scores are calculated and aggregated to obtain global similarity scores; the global similarity scores are filtered by using a preset threshold to obtain a verification result; the verification result is input into a blockchain smart contract, a distribution ratio is calculated by using a preset distribution function, and a payment process of a digital asset is automatically triggered according to the distribution ratio.

Description

Technical Field

[0001] This invention belongs to the field of blockchain technology for generative artificial intelligence data governance and traceability, specifically involving a generative AI data on-chain traceability method, system, and medium based on orthogonal visual fingerprints and lightweight matching. Background Technology

[0002] Generative AI, as a cutting-edge technology in the field of artificial intelligence, is rapidly changing the way digital content is created, especially in text-to-image synthesis models, which rely on billions of images scraped from the internet for training. However, this large-scale data collection often occurs without the explicit consent of creators, resulting in the unilateral removal of metadata from the original source. This allows AI developers to reap huge economic benefits from the generated assets, while the original artists are excluded from the economic value chain. Therefore, establishing a transparent and traceable data source verification mechanism to ensure informed consent from creators and fair royalty distribution has become a critical issue that urgently needs to be addressed in the field of generative AI.

[0003] Currently, there are many generative AI data on-chain traceability methods, but they all have their limitations. Traditional traceability based on centralized databases is difficult to scale due to single point of failure risk and high maintenance costs; methods relying on traditional visual features are not robust enough to deal with common modifications such as image compression and cropping, leading to matching failures; although blockchain traceability can ensure the immutability of data, it is limited by on-chain storage costs and consensus mechanism efficiency, making it difficult to directly process high-dimensional visual features and meet real-time requirements; although deep learning models can extract high-dimensional semantic features, their high computational resource consumption and black-box nature result in poor interpretability, and they are difficult to deploy in resource-constrained environments. Summary of the Invention

[0004] To address the shortcomings of existing technologies, the present invention aims to provide a generative AI data chain tracing method, system, and medium based on orthogonalized visual fingerprints and lightweight matching, thereby overcoming the aforementioned deficiencies in existing technologies.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] Generative AI data chain tracing methods based on orthogonalized visual fingerprinting and lightweight matching include:

[0007] Fingerprint extraction steps: Acquire the input image and extract initial features using a visual model; perform zero-mean and singular value decomposition on the initial features to obtain an initial projection matrix; perform matrix decomposition on the initial projection matrix to construct an orthogonal projection matrix, and use the orthogonal projection matrix to reduce the initial features to feature vectors of the target dimension; perform normalization and quantization packaging on the feature vectors of the target dimension in sequence to generate an orthogonal visual fingerprint in integer array format;

[0008] Geometric verification steps: Calculate the similarity between the fingerprint of the query image and the fingerprint of the database image, and filter the candidate image set; extract multi-scale features of the candidate images and the query image and construct a reliability map, filtering out low-confidence key points; use the nearest neighbor matching algorithm and neural network to predict and retain the sub-pixel-level offset between key points, calculate the local matching score and aggregate it to obtain the global similarity score; use a preset threshold to filter the global similarity score to obtain the verification result;

[0009] On-chain allocation steps: Store the orthogonalized visual fingerprint and the verification result in a distributed file system and generate a content identifier; input the verification result into the blockchain smart contract, calculate the allocation ratio through a preset allocation function, and automatically trigger the digital asset payment process according to the allocation ratio.

[0010] Preferably, the step of performing matrix decomposition on the initial projection matrix to construct an orthogonal projection matrix, and using the orthogonal projection matrix to reduce the dimensionality of the initial features to a feature vector of the target dimension, includes:

[0011] Perform an orthogonalization decomposition operation on the initial projection matrix constructed by truncating a specified number of columns from singular value decomposition to obtain a strictly orthogonal orthogonal projection matrix; use the orthogonal projection matrix to perform a linear transformation projection on the initial features to compress the initial features in the high-dimensional space into feature vectors of the target dimension.

[0012] Preferably, the step of performing normalization and quantization packaging on the feature vector of the target dimension to generate an orthogonal visual fingerprint in integer array format includes: performing norm normalization calculation on the target dimension feature vector after dimensionality reduction to map the feature data to a unit hypersphere; multiplying the normalized feature vector by a preset numerical factor and performing a rounding operation to convert it into a fixed-precision integer array; and encapsulating multiple feature integer variables into a single storage word according to the memory word length rules of the blockchain virtual machine, and outputting the orthogonal visual fingerprint.

[0013] Preferably, the calculation of the similarity between the query image fingerprint and the database image fingerprint, and the screening of the candidate image set, includes: calculating the cosine similarity between image fingerprints, and extracting images with a cosine similarity greater than a first threshold as strong matching images; counting the number of the strong matching images, and automatically triggering a retrieval rollback mechanism when the number is lower than a preset safety lower limit; expanding the retrieval matching range of the database through the retrieval rollback mechanism to obtain a larger order of magnitude of candidate images compiled into the candidate image set.

[0014] Preferably, the step of extracting multi-scale features from candidate images and query images and constructing a reliability map to filter out low-confidence key points includes: extracting multi-scale feature maps of the images at a preset level; upsampling and fusing the multi-scale feature maps according to a preset weight distribution to generate a dense feature map and a spatial reliability map; calculating a data-driven percentile filtering threshold based on the statistical distribution of values ​​in the spatial reliability map; and using the percentile filtering threshold to perform a masking operation on the spatial reliability map to remove low-confidence key point descriptors with values ​​lower than the threshold.

[0015] Preferably, the step of using the nearest neighbor matching algorithm and neural network prediction to retain the sub-pixel-level offset between key points, calculating local matching scores, and aggregating them to obtain a global similarity score includes: performing a mutual nearest neighbor matching algorithm on the filtered high-confidence matching feature set to establish a spatial topological correspondence between key points of the query image and candidate images; inputting the key points with established spatial topological correspondence into a multilayer perceptron neural network to output sub-pixel-level geometric offsets in the feature space; combining the spatial reliability value and the confidence weight corresponding to the geometric offset to calculate the local matching score of each key point pair; calculating the local matching scores in both directions respectively, and summing and averaging the bidirectional local matching scores to generate a global similarity score with symmetry attributes.

[0016] Preferably, storing the orthogonalized visual fingerprint and the verification result in a distributed file system and generating a content identifier includes: encapsulating the acquired orthogonalized visual fingerprint and the verification result, together with metadata information from the digital content anti-counterfeiting standard, into a structured data object; calculating the cryptographic hash value of the structured data object containing the current timestamp using a cryptographic hash function; and converting the cryptographic hash value into a unique content identifier code according to the addressing specifications of the distributed file system, serving as an anchor pointer between off-chain data and on-chain smart contracts.

[0017] Preferably, the verification result is input into a blockchain smart contract, the allocation ratio is calculated using a preset allocation function, and the digital asset payment process is automatically triggered based on the allocation ratio, including:

[0018] The smart contract introduces a global similarity score. When the global similarity score exceeds a preset truncation threshold, the relative contribution weight is obtained by subtracting the truncation threshold from the score. Normalization calculation is performed on the relative contribution weights of each data provider participating in this traceability verification to determine the specific royalty allocation ratio. When the global similarity scores of all candidates do not exceed the truncation threshold, the smart contract automatically resets to a uniform allocation fallback strategy. The blockchain node computing unit is called to execute the allocation logic of the royalty allocation ratio and to perform cross-ledger transfer of digital assets to the wallet address of the data provider.

[0019] A generative AI datachain-based on-chain tracing method system based on orthogonalized visual fingerprinting and lightweight matching is configured to execute the generative AI datachain-based on-chain tracing method. The system includes:

[0020] The fingerprint extraction module is used to acquire the input image and extract the initial features. It constructs an orthogonal projection matrix through zero mean, singular value decomposition and matrix decomposition to perform feature dimensionality reduction. The target dimension feature vector generated by dimensionality reduction is normalized and quantized and packaged to generate an orthogonal visual fingerprint in integer array format.

[0021] The geometric verification module is used to screen candidate image sets based on fingerprint similarity, extract multi-scale features to construct a reliability map and filter out low-confidence key points, call the nearest neighbor matching algorithm and neural network model to predict sub-pixel level offsets, calculate local matching scores and aggregate them to obtain global similarity scores for threshold filtering verification.

[0022] The on-chain allocation module is used to store the orthogonalized visual fingerprint and verification result in a distributed file system and generate a content identifier, and input the verification result into the blockchain smart contract to trigger automated digital asset payment through the allocation function.

[0023] A non-volatile computer-readable storage medium storing computer program instructions, which, when executed by a computer processor, enable the implementation of steps in a generative AI datachain tracing method based on orthogonalized visual fingerprints and lightweight matching.

[0024] The beneficial effects of this invention are as follows: The use of orthogonalized visual fingerprinting and lightweight matching significantly improves the accuracy and efficiency of generative AI data tracing. During visual fingerprint extraction, orthogonal principal component analysis is used to compress features for a compact representation, effectively suppressing hardware-induced numerical drift and ensuring feature consistency and stability across different hardware environments. Simultaneously, the compressed feature vectors are quantized and packaged into fixed-size integer arrays for efficient storage on the blockchain. In the two-stage hybrid retrieval process, lightweight matching first quickly retrieves candidate image sets with coarse granularity and reduces unnecessary computational overhead through a high-confidence filtering mechanism. Then, fine granular matching effectively filters out low-confidence matching points, achieving high-precision sub-pixel-level alignment. This design improves verification accuracy and significantly reduces computational latency, meeting real-time requirements. Attached Figure Description

[0025] Figure 1 This is an overall flowchart of the method of the present invention.

[0026] Figure 2 This is a schematic diagram of the orthogonal visual fingerprint extraction process of the method of the present invention.

[0027] Figure 3 This is a schematic diagram of the lightweight geometric verification process of the method of the present invention.

[0028] Figure 4 This is a schematic diagram of the on-chain traceability and royalty distribution process of the method of the present invention. Detailed Implementation

[0029] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0030] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0031] The embodiments of the present invention will be further described in detail below with reference to the accompanying drawings:

[0032] This invention relates to a generative AI data chain traceability method based on orthogonal visual fingerprinting and lightweight matching. This method achieves end-to-end traceability management and automated copyright allocation of generative AI training data through three core stages: fingerprint extraction, geometric verification, and on-chain allocation. The overall execution flow of this invention starts with the input image, and the data flow between each step forms a tight processing chain. The output of the previous step directly serves as the input of the next step, ensuring the determinism and reproducibility of the entire process. The specific embodiments of this invention are described in detail below.

[0033] In the fingerprint extraction step, the input image is first acquired, and initial features are extracted using a visual model. The visual model employs a self-supervised pre-trained backbone network based on a visual transformer architecture. This network is pre-trained on large-scale unsupervised data and is capable of extracting semantically rich high-dimensional feature representations from the input image. Specifically, after applying standardization preprocessing to the input image, the backbone network extracts a 384-dimensional classification token feature vector, denoted as the initial features. This refers to a vector in a 384-dimensional real space. The fundamental reason for employing a self-supervised pre-trained backbone network is that supervised networks have limited generalization ability when faced with image transformations outside the training distribution, while the semantic embedding vectors extracted by the self-supervised model can maintain higher robustness under common network perturbations such as JPEG compression, Gaussian blur, and center cropping. To avoid introducing the risk of data leakage, the backbone network remains frozen throughout the entire system operation, without any parameter updates. All trainable parameters are limited to the projection matrix construction process in the subsequent offline stage.

[0034] After obtaining the initial features, zero-mean normalization and singular value decomposition are performed on the initial features to obtain the initial projection matrix. Zero-mean normalization is a necessary pre-operation of principal component analysis, and its purpose is to eliminate the interference of feature distribution center shift on subsequent directional decomposition. Specifically, the offline training sample set contains N samples, each corresponding to a d-dimensional feature vector. First, calculate the mean vector of all sample features. Then, the mean vector is subtracted from the entire feature matrix to obtain the centered feature matrix. The calculation formula is:

[0035]

[0036] in, Let N be the sample mean vector, and N be the total number of training samples. Let X be the original feature vector of the i-th sample, and let X be the feature matrix formed by stacking the features of all samples. This is the centered eigenvalue matrix after zero-mean normalization. Then, the centered eigenvalue matrix... Performing singular value decomposition (SVD) decomposes the matrix into the product of three submatrices, where each column of the right singular vector matrix corresponds to the principal direction of variance in the eigenspace, arranged in descending order. The decomposition result is denoted as:

[0037]

[0038] Where U is the left singular vector matrix, V is a singular value diagonal matrix, and V is a right singular vector matrix. This is the transpose of V. The initial projection matrix is ​​constructed by truncating the first 256 columns of V, which retains the 256 principal directions with the largest variance contribution in the original feature space. The samples used in the above offline training process come from a large-scale image dataset with semantic similarity to the registration database. Cosine similarity is used to ensure the distribution quality of the training samples, thereby establishing a stable and reliable projection benchmark for subsequent online fingerprint extraction.

[0039] After obtaining the initial projection matrix, matrix decomposition is performed on the initial projection matrix to construct an orthogonal projection matrix. This orthogonal projection matrix is ​​then used to reduce the dimensionality of the initial features to eigenvectors of the target dimension. It should be noted that standard principal component analysis (PCA) operations, when performing linear projection, can destroy the L2 norm of the eigenvectors, causing the projected vectors to deviate from the unit sphere, resulting in geometrical damage during subsequent quantization operations. More critically, when the system is deployed on heterogeneous decentralized worker nodes, the different floating-point precisions of the hardware between nodes can cause numerical drift in the orthogonality of the projection matrix, leading to arithmetic overflow in the fixed-point operations of the blockchain virtual machine. To fundamentally eliminate these hidden dangers, this invention modifies the initial projection matrix... Perform orthogonalization operations based on QR decomposition, i.e., on... After performing QR matrix decomposition, the orthogonality factor Q is taken as the orthogonal projection matrix P, obtaining an orthogonal projection matrix that strictly satisfies the orthogonality constraint. Its mathematical expression is:

[0040] , making

[0041] in, Let P be the orthogonalization function based on QR decomposition, and let P be the orthogonalized projection matrix. This is the transpose of P. The initial features are a 256-order identity matrix. This constraint requires that each column of the projection matrix be pairwise orthogonal and consist of unit vectors. This orthogonalization step strictly constrains the Frobenius norm error to an extremely low level, ensuring that fixed-point operations in the subsequent blockchain virtual machine will not cause arithmetic overflow, thus guaranteeing the computational security of the system in a heterogeneous distributed hardware environment. After orthogonalization, the initial features are linearly transformed and projected using the orthogonal projection matrix, compressing the initial features in the high-dimensional space into feature vectors of the target dimension. The specific calculation method is as follows: Let... ,in Here, μ is the 384-dimensional original feature vector extracted by the backbone network during the online phase, μ is the mean vector calculated during the offline training phase, and P is the orthogonal projection matrix. This is the projected 256-dimensional feature vector.

[0042] After feature reduction, the feature vectors of the target dimension are sequentially normalized and quantized to generate an orthogonalized visual fingerprint in integer array format. An orthogonalized visual fingerprint refers to an image feature encoding represented as an integer array, generated after initial features are extracted using the aforementioned visual model and then sequentially processed through zero-mean normalization, singular value decomposition, orthogonalized matrix decomposition, dimensionality reduction, normalization, and quantization. Its design aims to simultaneously meet the stringent requirements of the blockchain virtual machine for fixed-point precision computation and the constraint of limited storage word length. The normalization operation performs L2 norm normalization calculation on the 256-dimensional feature vector after dimensionality reduction, mapping the feature data to a unit hypersphere to eliminate differences in feature vector magnitudes, thus providing a unified metric for subsequent cosine similarity calculations. The calculation method is as follows: ,in This is a normalized 256-dimensional unit vector. for The L2 norm. After normalization, the normalized eigenvectors are multiplied by a preset numerical factor and rounded to convert them into a fixed-precision integer array. The quantization formula is:

[0043]

[0044] in, The quantized 256-dimensional integer array, This is the rounding function. This is a preset numerical scaling factor. This represents a 256-dimensional integer vector space. Quantization converts the floating-point representation to fixed-precision integers because the blockchain virtual machine cannot handle the IEEE 754 floating-point format; all on-chain operations must be performed based on integer types. A preset numerical scaling factor ensures that the absolute value of all integer components is strictly less than 2 to the power of 63, satisfying the storage limit for 64-bit signed integers. Finally, according to the blockchain virtual machine's memory word length rules, multiple feature integer variables are encapsulated into a single storage word, outputting the orthogonalized visual fingerprint. Each memory word of the blockchain virtual machine contains 32 bytes. This method encapsulates four 64-bit signed integer variables into each storage word, thereby compressing the 256-dimensional feature vector into 64 storage words, fixing the on-chain storage footprint of each image at 2048 bytes. This integer array format is fully compatible with existing smart contract interfaces, requiring no modification to any on-chain code, achieving zero-intrusion integration.

[0045] After fingerprint extraction, the geometric verification step begins. This step first calculates the similarity between the query image fingerprint and the database image fingerprints to filter the candidate image set. The similarity calculation here uses the cosine similarity index. The orthogonalized visual fingerprint of the query image is compared with the cosine similarity of each registered image fingerprint in the database. Cosine similarity measures the degree of alignment between two vectors, with values ​​ranging from -1 to positive1. A value closer to positive1 indicates a closer similarity in semantic features between the two images. The system extracts the top 50 images based on cosine similarity to form an initial candidate set. Based on this, images with a cosine similarity greater than the first threshold are further extracted as strongly matching images. The selection criteria for the strongly matching subset H are as follows:

[0046]

[0047] Where H is a high-confidence strong-match subset. This is the candidate set of the top 50 based on cosine similarity. To query image fingerprints fingerprints of candidate images The cosine similarity between the images is calculated, with 0.7 as the first threshold. The number of strongly matching images is then counted. When this number falls below a preset safety threshold, a retrieval fallback mechanism is automatically triggered. The safety threshold is set to 5; that is, when the number of strongly matching images is less than 5, the system determines that the initial retrieval results are ambiguous and cannot provide sufficient candidate data for subsequent fine-grained geometric verification, thus automatically triggering the retrieval fallback mechanism. This fallback mechanism expands the retrieval matching range of the database, increasing the candidate set size from 50 to 100, obtaining a larger order of magnitude of candidate images to assemble into the candidate image set. This fallback mechanism ensures the stability of the recall rate when facing long-tail query scenarios. Simultaneously, for simple queries with clear similarity results, the system skips subsequent fine-grained geometric calculations, effectively saving computational resources.

[0048] After the candidate image set is determined, multi-scale features of the candidate images and query images are extracted and a reliability map is constructed to filter out low-confidence keypoints. The reliability map refers to the spatial confidence matrix generated synchronously with the dense feature map during the multi-scale feature extraction process. The value at each position reflects the reliability of the detected keypoint at that position and is used for subsequent filtering of low-confidence keypoints. Specifically, the geometric verification module performs 6-layer multi-scale convolutional neural network feature extraction on both the candidate and query images, obtaining corresponding multi-scale feature maps at multiple preset scale levels. The multi-scale feature maps are upsampled and fused according to a preset weight distribution. Feature maps from different scale levels are weighted and fused at ratios of 0.6, 0.3, and 0.1 to generate a dense feature map F and a spatial reliability map R. The dense feature map provides a high-dimensional descriptor for subsequent nearest neighbor matching, while the spatial reliability map records the detection confidence value at each position on the image plane in matrix form. Based on the statistical distribution of the values ​​in the spatial reliability map, a data-driven percentile filtering threshold is calculated. The threshold calculation method is as follows:

[0049]

[0050] in, The filtering threshold is... To query the 80th percentile of all values ​​in the image spatial reliability map R_q, The threshold is the 80th percentile of all values ​​in the spatial reliability map R_c of the candidate images. A data-driven percentile threshold is used instead of a fixed cutoff value to allow the filtering operation to adapt to the differences in feature distributions of different images, avoiding the over-removal of keypoints in high-quality images or the insufficient filtering of noisy keypoints in low-quality images due to a uniform rigid threshold. The spatial reliability map is masked using the percentile filtering threshold to remove low-confidence keypoint descriptors with values ​​below the threshold, ultimately retaining the high-confidence matching feature set. and , which correspond to the key point descriptor subsets filtered by reliability in the query image and candidate images, respectively.

[0051] After filtering out low-confidence keypoints, the nearest neighbor matching algorithm and neural network prediction are used to retain sub-pixel-level offsets between keypoints. Local matching scores are calculated and aggregated to obtain a global similarity score. Sub-pixel-level offsets refer to geometric displacement vectors predicted by the neural network during the keypoint matching stage, with precision exceeding single-pixel granularity. This precision ensures that the spatial correspondence of the image can still be accurately captured even with slight geometric transformations. The nearest neighbor matching algorithm is then applied to the filtered high-confidence matching feature set to establish the spatial topological correspondence between keypoints in the query image and candidate images. Nearest neighbor matching requires that the two keypoints participating in the pairing be each other's nearest neighbors; that is, only keypoint pairs satisfying the bidirectional nearest neighbor condition are retained, thereby eliminating one-way false matches and improving the reliability of the spatial correspondence. The keypoints for establishing the spatial topological correspondence are input into a multilayer perceptron neural network, which outputs sub-pixel-level geometric offsets in the feature space. The multilayer perceptron architecture used is 8×8, and its offset prediction results are normalized using the softmax function and used as the geometric confidence weight conf_refine_k. Combining the spatial reliability value and the confidence weight corresponding to the geometric offset, the local matching score of each keypoint pair is calculated using the following formula:

[0052]

[0053] in, Let be the local matching score of the k-th keypoint pair. To query key points in the image spatial reliability graph Location reliability values, Corresponding key points in the spatial reliability map of candidate images Location reliability values, This sets the confidence weights for sub-pixel offset predictions from the multilayer perceptron neural network output. Local matching scores are calculated in both directions, and the summation and averaging of these bidirectional local matching scores generate a global similarity score with symmetry properties. The global similarity score is a comprehensive evaluation index obtained by bidirectionally symmetrically aggregating the local matching scores of all matching keypoint pairs. Its value ranges from 0 to 1 and is used to measure the overall geometric similarity between the query image and the candidate image. The calculation formula is as follows:

[0054]

[0055] Where s is the global similarity score, and K is the total number of keypoint pairs that establish spatial correspondence through nearest neighbor matching. The local matching score for the k-th keypoint pair is calculated from the query image to the candidate image direction. The local matching score for the k-th keypoint pair is calculated from the candidate image to the query image. A bidirectional symmetric aggregation design ensures that the similarity score is unaffected by the image input order, enhancing the objectivity and stability of the scoring results. The global similarity score is filtered using a preset threshold to obtain the verification result. If the global similarity score *s* exceeds the preset truncation threshold, a verifiable geometric overlap is considered between the query image and the corresponding candidate image, and the verification result is marked as a match; otherwise, the verification result is marked as a mismatch.

[0056] After the geometric verification step is completed and the verification result is obtained, the on-chain allocation step begins. The on-chain allocation step first stores the orthogonalized visual fingerprint and the verification result in a distributed file system and generates a content identifier. In this method, the distributed file system specifically refers to a peer-to-peer file storage network using a content addressing mechanism. Data is addressed using its cryptographic hash value as a unique identifier, thereby ensuring data immutability and decentralized storage. Specifically, the obtained orthogonalized visual fingerprint and the verification result, along with the metadata information of the digital content anti-counterfeiting standard, are encapsulated into a structured data object. The digital content anti-counterfeiting standard is an open technical specification developed by the Content Source and Authenticity Consortium. It defines a structured representation of the creation declaration, source information, and processing history of digital content. The metadata information includes, but is not limited to, fields such as content creator declaration, creation timestamp, content-related hash value, and digital signature. After encapsulating the above information into a structured data object, a cryptographic hash function is used to calculate the encrypted hash value of the structured data object, which includes the current timestamp. The purpose of introducing the current timestamp is to ensure that the structured data object generated by each encapsulation operation has temporal uniqueness. This ensures that even if the content itself remains unchanged, encapsulation results at different times will produce different encrypted hash values, thus preventing replay attacks. According to the addressing specifications of the distributed file system, the encrypted hash value is converted into a unique content identifier code, serving as an anchor pointer between off-chain data and on-chain smart contracts. The content identifier is generated by calculating the stored content using a cryptographic hash function, according to the addressing specifications of the distributed file system. It has content addressability, meaning there is a deterministic one-to-one correspondence between the content identifier and the file content. Any unauthorized tampering with the stored content will fundamentally change the content identifier, immediately severing its association with the on-chain smart contract pointer and ensuring the tamper-proof integrity of off-chain metadata.

[0057] After completing the storage of off-chain data and generating content identifiers, the verification results are input into the blockchain smart contract. The allocation ratio is calculated using a preset allocation function, and the digital asset payment process is automatically triggered based on this ratio. A smart contract is a self-executing program code deployed on a blockchain virtual machine. It can automatically trigger the transfer and allocation of digital assets when preset conditions are met, without manual intervention. Its execution process is entirely guaranteed by the on-chain consensus mechanism, possessing immutability and transparency. A global similarity score is introduced into the smart contract. When the global similarity score exceeds a preset truncation threshold, the relative contribution weight is obtained by subtracting the truncation threshold from the score. Let the global similarity score corresponding to the i-th data provider participating in this traceability verification be denoted as . Its unnormalized relative contribution weight The calculation method is as follows Where λ is a preset truncation threshold, the max(·,0) operation ensures that only data providers with a global similarity score exceeding the truncation threshold receive a non-zero weight; candidate results with a score below the truncation threshold do not participate in royalty allocation, thus preventing low-credibility forged content from encroaching on the funding pool from an algorithmic mechanism perspective. The relative contribution weights of each data provider participating in this source tracing verification are normalized to determine the specific royalty allocation ratio. The normalization formula is: ,in The final royalty allocation percentage for the i-th data provider. This is the sum of the unnormalized relative contribution weights of all data providers. When the global similarity scores of all candidates do not exceed the cutoff threshold, i.e., all... The initial values ​​are all close to zero. The smart contract automatically resets to a uniform rollback strategy, distributing royalties equally among all registered data providers to prevent payment failures due to the inability to identify major contributors. The blockchain node's computing unit executes the royalty allocation logic according to the stated proportions, performing cross-ledger transfers of digital assets to the data providers' wallet addresses. The blockchain node's computing unit executes the multi-party royalty transfer operation in an immutable manner based on the pre-coded allocation logic in the smart contract. The entire transfer process requires no trusted third-party intervention, ensuring the transparency and non-repudiation of the allocation results. Since the orthogonal visual fingerprints are quantized and packaged according to the blockchain virtual machine's memory word length rules, the on-chain storage usage of each image is strictly controlled within a fixed number of bytes, fundamentally limiting the blockchain virtual machine's gas consumption and effectively preventing denial-of-service risks caused by memory exhaustion.

[0058] In summary, the generative AI data on-chain tracing method based on orthogonalized visual fingerprints and lightweight matching proposed in this invention fundamentally eliminates the numerical drift problem in heterogeneous hardware environments by introducing zero-mean, singular value decomposition, and QR decomposition-based orthogonalized matrix construction processes in the fingerprint extraction step, and compresses high-dimensional semantic features into an integer array format conforming to the blockchain virtual machine storage specifications. In the geometric verification step, a two-stage hybrid architecture combining candidate image set screening and multi-scale feature extraction is employed. Data-driven percentile thresholds are used to filter low-confidence key points, and sub-pixel-level geometric alignment is achieved using a multilayer perceptron neural network, significantly improving verification accuracy and efficiency under common image perturbation conditions. Furthermore, by combining cryptographically bound off-chain metadata storage with automated royalty distribution via smart contracts in the on-chain allocation step, a decentralized, trustless third-party creator rights protection mechanism is achieved. This system architecture features logical decoupling and clear module boundaries. Existing on-chain contract code can be smoothly integrated with the orthogonalized visual fingerprint output generated by this method without any modification, exhibiting good backward compatibility and engineering deployability.

[0059] A generative AI datachain-based on-chain tracing method system based on orthogonalized visual fingerprinting and lightweight matching is configured to execute the generative AI datachain-based on-chain tracing method. The system includes:

[0060] The fingerprint extraction module is used to acquire the input image and extract the initial features. It constructs an orthogonal projection matrix through zero mean, singular value decomposition and matrix decomposition to perform feature dimensionality reduction. The target dimension feature vector generated by dimensionality reduction is normalized and quantized and packaged to generate an orthogonal visual fingerprint in integer array format.

[0061] The geometric verification module is used to screen candidate image sets based on fingerprint similarity, extract multi-scale features to construct a reliability map and filter out low-confidence key points, call the nearest neighbor matching algorithm and neural network model to predict sub-pixel level offsets, calculate local matching scores and aggregate them to obtain global similarity scores for threshold filtering verification.

[0062] The on-chain allocation module is used to store the orthogonalized visual fingerprint and verification result in a distributed file system and generate a content identifier, and input the verification result into the blockchain smart contract to trigger automated digital asset payment through the allocation function.

[0063] A non-volatile computer-readable storage medium storing computer program instructions, which, when executed by a computer processor, enable the implementation of steps in a generative AI datachain tracing method based on orthogonalized visual fingerprints and lightweight matching.

[0064] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principle of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A generative AI data chain tracing method based on orthogonal visual fingerprinting and lightweight matching, characterized in that, include: Fingerprint extraction steps: Acquire the input image and extract initial features using a visual model; The initial features are decomposed to obtain the initial projection matrix; The initial projection matrix is ​​decomposed to construct an orthogonal projection matrix. The initial features are then reduced to feature vectors of the target dimension using the orthogonal projection matrix. Based on the feature vectors of the target dimension, an orthogonalized visual fingerprint in integer array format is generated. Geometric verification steps: Calculate the similarity between the fingerprint of the query image and the fingerprint of the database image, and filter the candidate image set; extract multi-scale features of the candidate images and the query image and construct a reliability map; use the nearest neighbor matching algorithm and neural network to predict and retain the sub-pixel level offset between key points, calculate the local matching score and aggregate it to obtain the global similarity score; use a preset threshold to filter the global similarity score to obtain the verification result; On-chain allocation steps: Store the orthogonalized visual fingerprint and the verification result in a distributed file system and generate a content identifier; input the verification result into the blockchain smart contract, calculate the allocation ratio through a preset allocation function, and automatically trigger the digital asset payment process according to the allocation ratio.

2. The generative AI data chain tracing method based on orthogonalized visual fingerprints and lightweight matching according to claim 1, characterized in that, The step of performing matrix decomposition on the initial projection matrix to construct an orthogonal projection matrix, and using the orthogonal projection matrix to reduce the dimensionality of the initial features to a feature vector of the target dimension, includes: Perform an orthogonalization decomposition operation on the initial projection matrix constructed by truncating a specified number of columns from singular value decomposition to obtain a strictly orthogonal orthogonal projection matrix; use the orthogonal projection matrix to perform a linear transformation projection on the initial features to compress the initial features in the high-dimensional space into feature vectors of the target dimension.

3. The generative AI data chain tracing method based on orthogonalized visual fingerprints and lightweight matching according to claim 1, characterized in that, The step of generating an orthogonalized visual fingerprint in integer array format based on the feature vector of the target dimension includes: performing norm normalization calculation on the target dimension feature vector after dimensionality reduction to map the feature data to a unit hypersphere; multiplying the normalized feature vector by a preset numerical factor and performing a rounding operation to convert it into a fixed-precision integer array; and encapsulating multiple feature integer variables into a single storage word according to the memory word length rules of the blockchain virtual machine, and outputting the orthogonalized visual fingerprint.

4. The generative AI datachain tracing method based on orthogonalized visual fingerprints and lightweight matching according to claim 1, characterized in that, The calculation of the similarity between the query image fingerprint and the database image fingerprint, and the screening of the candidate image set, includes: calculating the cosine similarity between image fingerprints, and extracting images with a cosine similarity greater than a first threshold as strong matching images; counting the number of strong matching images, and automatically triggering a retrieval fallback mechanism when the number is lower than a preset safety lower limit; expanding the retrieval matching range of the database through the retrieval fallback mechanism to obtain a larger order of magnitude of candidate images to compile into the candidate image set.

5. The generative AI datachain tracing method based on orthogonalized visual fingerprinting and lightweight matching according to claim 1, characterized in that, The step of extracting multi-scale features from candidate images and query images and constructing a reliability map to filter out low-confidence key points includes: extracting multi-scale feature maps of the images at a preset level; upsampling and fusing the multi-scale feature maps according to a preset weight distribution to generate a dense feature map and a spatial reliability map; calculating a data-driven percentile filtering threshold based on the statistical distribution of values ​​in the spatial reliability map; and using the percentile filtering threshold to perform a masking operation on the spatial reliability map to remove low-confidence key point descriptors with values ​​lower than the threshold.

6. The generative AI datachain tracing method based on orthogonalized visual fingerprints and lightweight matching according to claim 1, characterized in that, The process of using the nearest neighbor matching algorithm and neural network prediction to retain sub-pixel-level offsets between keypoints, calculating local matching scores, and aggregating them to obtain a global similarity score includes: performing a mutual nearest neighbor matching algorithm on the filtered high-confidence matching feature set to establish a spatial topological correspondence between keypoints in the query image and candidate images; inputting the keypoints with established spatial topological correspondences into a multilayer perceptron neural network to output sub-pixel-level geometric offsets in the feature space; combining the spatial reliability value and the confidence weight corresponding to the geometric offsets to calculate the local matching score for each keypoint pair; calculating the local matching scores in both directions respectively, and summing and averaging the bidirectional local matching scores to generate a global similarity score with symmetry attributes.

7. The generative AI datachain tracing method based on orthogonalized visual fingerprinting and lightweight matching according to claim 1, characterized in that, The step of storing the orthogonalized visual fingerprint and the verification result in a distributed file system and generating a content identifier includes: encapsulating the acquired orthogonalized visual fingerprint and the verification result, together with the metadata information of the digital content anti-counterfeiting standard, into a structured data object; calculating the cryptographic hash value of the structured data object containing the current timestamp using a cryptographic hash function; and converting the cryptographic hash value into a unique content identifier code according to the addressing specification of the distributed file system, which serves as an anchor pointer between off-chain data and on-chain smart contracts.

8. The generative AI datachain tracing method based on orthogonalized visual fingerprints and lightweight matching according to claim 1, characterized in that, The step of inputting the verification result into the blockchain smart contract, calculating the allocation ratio through a preset allocation function, and automatically triggering the digital asset payment process based on the allocation ratio includes: The global similarity score is introduced into the smart contract. When the global similarity score exceeds a preset truncation threshold, the relative contribution weight is obtained by subtracting the truncation threshold from the score. The relative contribution weights of each data provider participating in this source tracing verification are normalized to determine the specific royalty allocation ratio. When the global similarity scores of all candidates do not exceed the truncation threshold, the smart contract automatically resets to a uniform distribution fallback strategy.

9. A generative AI data chain tracing method system based on orthogonal visual fingerprinting and lightweight matching, characterized in that, The system is configured to perform a generative AI datachain tracing method based on orthogonalized visual fingerprints and lightweight matching as described in any one of claims 1 to 8, the system comprising: The fingerprint extraction module is used to acquire the input image and extract the initial features. It constructs an orthogonal projection matrix through zero mean, singular value decomposition and matrix decomposition to perform feature dimensionality reduction. The target dimension feature vector generated by dimensionality reduction is normalized and quantized and packaged to generate an orthogonal visual fingerprint in integer array format. The geometric verification module is used to screen candidate image sets based on fingerprint similarity, extract multi-scale features to construct a reliability map and filter out low-confidence key points, call the nearest neighbor matching algorithm and neural network model to predict sub-pixel level offsets, calculate local matching scores and aggregate them to obtain global similarity scores for threshold filtering verification. The on-chain allocation module is used to store the orthogonalized visual fingerprint and verification result in a distributed file system and generate a content identifier, and input the verification result into the blockchain smart contract to trigger automated digital asset payment through the allocation function.

10. A non-volatile computer-readable storage medium storing computer program instructions thereon, characterized in that, When the computer program instructions are executed by a computer processor, they can implement the steps of the generative AI data chain tracing method based on orthogonalized visual fingerprints and lightweight matching as described in any one of claims 1 to 8.

Images