Method, device, storage medium and electronic equipment for managing artificial intelligence generated content

Abstract

The application provides an artificial intelligence generated content management method and device, a storage medium and an electronic device, and relates to the field of information security. The method generates a watermark commitment, an on-chain index and a content binding value based on a preset watermark key, and uses zero-knowledge proof to prove that the three are derived from the same key, and the content binding value is further associated with the content integrity feature, thereby realizing trusted traceability, anti-transplantation and misuse, and multi-scale tampering positioning under limited watermark capacity, and reducing the risk of generated content being transplanted and misused.

Description

Technical Field

[0001] This application relates to the field of information security, and more specifically, to a method, apparatus, storage medium, and electronic device for managing artificial intelligence-generated content. Background Technology

[0002] In recent years, generative artificial intelligence technology has developed rapidly, and Artificial Intelligence Generated Content (AIGC), automatically generated by AI models, such as images, videos, audio, and text, has been widely disseminated on the internet. While AIGC has brought productivity improvements, it has also raised a series of information security and copyright issues: First, it is difficult to determine whether a piece of content was generated by artificial intelligence, leading to the proliferation of deepfakes and false information; second, the traceability information such as the generating entity, generation model, generation time, and generation parameters of AIGC content is difficult to reliably link to the content itself, making it difficult to determine copyright ownership; third, content is easily tampered with during dissemination through methods such as cropping, compression, and partial editing, making it difficult to determine whether the content has been modified and which specific areas have been modified.

[0003] To address the aforementioned issues, digital watermarking technology, blockchain-based evidence storage technology, and related technologies combining the two have been proposed. Digital watermarking technology embeds imperceptible watermark information into AIGC content during or after its generation, and extracts the watermark during the verification phase to determine the content's origin. Blockchain-based evidence storage technology writes the hash value of the content or related records to the blockchain, leveraging its immutability to achieve evidence storage.

[0004] Research has found that none of the aforementioned schemes cryptographically bind the watermark key, the watermark embedded in the content, the on-chain traceability certificate, and the content's perceptual hash Merkle root into a unified whole. Therefore, they still cannot simultaneously achieve watermark portability, stable correspondence between content and on-chain records, and trusted third-party verification and tamper location. Even simply overlaying digital watermarks with blockchain notarization—that is, embedding the watermark and then recording it on the chain—failure to solve the aforementioned problems due to the lack of strong cryptographic binding between the two, and still cannot prevent the watermark from being copied and misused. Summary of the Invention

[0005] To overcome at least one deficiency in the prior art, this application provides a method, apparatus, storage medium, and electronic device for managing artificial intelligence-generated content, which can achieve reliable traceability, prevention of unauthorized use, and multi-scale tampering location with limited watermark capacity, thereby reducing the risk of generated content being unauthorized.

[0006] Firstly, this application provides a method for managing artificial intelligence-generated content, the method comprising: Obtain the generated content and the corresponding source metadata; The generated content is subjected to content feature extraction to obtain content integrity features, and the source tracing metadata is stored to obtain the metadata storage address; A watermark commitment and on-chain index are generated based on a preset watermark key. A content binding value is generated based on the watermark key and the content integrity feature. A zero-knowledge proof is generated, wherein the zero-knowledge proof is used to prove that the watermark commitment, the on-chain index and the content binding value are associated with the watermark key, and that the content binding value is associated with the content integrity feature. Based on the on-chain index as the query identifier, the content integrity feature, the watermark commitment, the content binding value, the metadata storage address, and the zero-knowledge proof are written into the blockchain to form a traceability certificate. The on-chain index and the content binding value are embedded as watermark payloads into the generated content to obtain watermarked generated content.

[0007] Secondly, this application also provides a management device for artificial intelligence-generated content, the device comprising: The raw data module is used to acquire the generated content and the source metadata corresponding to the generated content; The content processing module is used to extract content features from the generated content to obtain content integrity features, store the source traceability metadata, and obtain the metadata storage address; The traceability information module is used to generate a watermark commitment and an on-chain index based on a preset watermark key, generate a content binding value based on the watermark key and the content integrity feature, and generate a zero-knowledge proof, wherein the zero-knowledge proof is used to prove that the watermark commitment, the on-chain index and the content binding value are associated with the watermark key, and that the content binding value is associated with the content integrity feature; The traceability information module is also used to write the content integrity feature, the watermark commitment, the content binding value, the metadata storage address and the zero-knowledge proof into the blockchain based on the on-chain index as the query identifier, thereby forming a traceability certificate; The watermark generation module is used to embed the on-chain index and the content binding value as watermark payload into the generated content to obtain watermarked generated content.

[0008] Thirdly, this application provides a storage medium storing a computer program that, when executed by a processor, implements the aforementioned method for managing artificial intelligence-generated content.

[0009] Fourthly, this application provides an electronic device, which includes a processor and a memory. The memory stores a computer program, which, when executed by the processor, implements the aforementioned method for managing artificial intelligence-generated content.

[0010] Compared with the prior art, this application has the following beneficial effects: The management method, apparatus, storage medium, and electronic device for AI-generated content provided in this application synchronously generate a watermark commitment, on-chain index, and content binding value based on a preset watermark key, and uniformly verify with zero-knowledge proof that all three originate from the same key. Furthermore, the content binding value is further associated with the content integrity feature, thereby achieving trusted traceability, prevention of porting and misuse, and multi-scale tampering location under limited watermark capacity, reducing the risk of generated content being ported and misused. Attached Figure Description

[0011] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 One of the flowcharts for a method of managing artificial intelligence-generated content provided in an embodiment of this application; Figure 2 A second schematic flowchart illustrating the management method for AI-generated content provided in this application embodiment; Figure 3 A third flowchart illustrating the management method for AI-generated content provided in this application embodiment; Figure 4 A schematic diagram of the structure of the management device for artificial intelligence-generated content provided in the embodiments of this application; Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0013] To make the objectives, technical solutions, and advantages of the embodiments of this application (hereinafter referred to as "the embodiments") clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0014] Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0015] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0016] In the description of this application, it should be noted that the terms "first," "second," "third," etc., are used only for distinguishing descriptions and should not be construed as indicating or implying relative importance. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0017] To make the following description of this embodiment easier to understand, the relevant technical data that may affect the subsequent understanding will be explained first: Blockchain is a decentralized distributed ledger where data is recorded and verified by all nodes on the network. Therefore, once written, it cannot be unilaterally altered or deleted, ensuring the openness, transparency, and non-repudiation of information.

[0018] A hash value is an algorithm that can map data of any length to a string of fixed length. Therefore, it is extremely sensitive to the input. Even changing just one punctuation mark will result in a completely different output. It is often regarded as the "electronic fingerprint" of data.

[0019] Perceptual hashing is a hashing technique that focuses on visual similarity. Compared to traditional hash values, it does not care about pixel-level subtle differences (such as brightness or compression). As long as the content of the images appears the same to the human eye, they are considered identical. Conversely, if the content is substantially modified, the differences will immediately become apparent.

[0020] Merkle trees are tree-like data structures that pair and re-hash the hash values ​​of a large number of data blocks layer by layer, ultimately aggregating them into a "root". When using Merkle trees to verify whether any single block of data has been tampered with, only a very small number of path nodes need to be checked, without downloading all the data.

[0021] The watermark key, the unique password (private key) for embedding and extracting watermarks, is similar to the key to a safe and must be absolutely confidential; once leaked, attackers can forge, tamper with, or erase watermarks at will, rendering the entire copyright protection system ineffective.

[0022] Watermarking commitment is a cryptographic encapsulation of the watermark key, used to generate a public and irreversible value, much like locking a key in a transparent safe. Anyone can confirm that the safe is intact, but cannot retrieve the key, thus solving the problems of key escrow and trust.

[0023] Zero-knowledge proof is a cryptographic protocol that allows the prover to demonstrate to the verifier that the secret does exist and has been used correctly without revealing any information about the secret itself, thus achieving ultimate privacy by "proving the truth while keeping the secret."

[0024] The on-chain index, a unique identifier derived from the watermark key, can guide users to accurately locate the evidence record of the content in the massive amount of data on the blockchain, achieving efficient retrieval without exposing the key itself.

[0025] The root binding value is the result of a cryptographic operation that forcibly binds the watermark to the content fingerprint. It involves encrypting the Merkle root using a key. If the watermark is moved to another graph, the calculated binding value will immediately become invalid, thus reducing the risk of misuse.

[0026] Semi-fragile watermarks are a type of watermark that combines robustness and sensitivity. They can tolerate normal format conversions or compression and keep the watermark alive; however, when faced with malicious tampering such as partial erasure or cropping, they will react immediately and accurately mark the damaged area.

[0027] Based on the above-mentioned technical terms, as introduced in the background section, in order to solve a series of information security and copyright issues caused by AIGC content, digital watermarking technology, blockchain evidence storage technology, and related technologies that combine the two have been proposed.

[0028] Digital watermarking technology embeds imperceptible watermark information into AIGC content during or after its generation, and extracts the watermark during the verification stage to determine the source of the content. However, it has the following limitations: 1. Due to the constraints of imperceptibility and robustness, the effective information that a watermark can carry is usually only tens to hundreds of bits, making it difficult to fully carry traceability metadata such as generation model, generation time, copyright holder, and generation parameters.

[0029] 2. Watermark verification usually relies on the key held by the embedding party for comparison. The verifier and the embedding party are the same trusted entity, which has a "self-verification" trust defect. It is difficult for a third party to independently verify without trusting the embedding party.

[0030] 3. Once a watermark is extracted or copied, it can be transplanted into other people's content and misused, lacking anti-counterfeiting capabilities.

[0031] Blockchain-based evidence storage technology writes the hash value of content or related records into the blockchain, utilizing its immutability to achieve evidence storage, but it also has the following shortcomings: 1. Cryptographic hashes are highly sensitive to any minor changes to the content. The hash recalculated after the content is compressed, scaled, or converted in format will be inconsistent with the on-chain record, making it difficult for the content to correspond stably with the on-chain record.

[0032] 2. It can only record the fact that a certain hash was uploaded to the chain at a certain moment. It cannot locate the specific area where the content was tampered with, nor can it trace the original version of the tampered content.

[0033] In view of the above-mentioned shortcomings, a related technology that partially combines the two types of technologies is proposed. One type of scheme registers the perceptual hash of the content as a fingerprint on the blockchain and verifies the source through similarity comparison, but it does not embed a watermark into the content, nor does it involve watermark keys, zero-knowledge proofs, or tamper location. Another approach uses zero-knowledge proofs to publicly verify the watermark without disclosing the watermark parameters, but this only applies to text watermarks, is not bound to on-chain traceability records, and does not involve perceptual hashing or tamper location.

[0034] Therefore, none of the above solutions cryptographically bind the watermark key, the watermark embedded in the content, the on-chain traceability certificate, and the content's perceptual hash Merkle root into a unified whole. Consequently, they still cannot simultaneously achieve watermark portability, stable correspondence between content and on-chain records, and trusted third-party verification and tamper location. Even simply overlaying digital watermarks with blockchain notarization—that is, recording the data on the chain after embedding the watermark—failure to solve the aforementioned problems due to the lack of strong cryptographic binding between the two, and the inability to prevent watermark copying and misuse, remains unresolved.

[0035] It should be noted that the defects in the solutions in the prior art are the result of practice and careful research. Therefore, the discovery process of the above problems and the solutions proposed by the embodiments of this application in the following text should be regarded as contributions to this application in the process of invention and creation, and should not be understood as technical content known to those skilled in the art.

[0036] Based on the discovery of the aforementioned technical problems, this embodiment provides a method for managing artificial intelligence-generated content. For example... Figure 1 As shown, the method includes: S1, obtain the generated content and the corresponding source metadata.

[0037] S2 extracts content features from the generated content to obtain content integrity features, stores source metadata, and obtains the metadata storage address.

[0038] S3 generates a watermark commitment and on-chain index based on a preset watermark key, generates a content binding value based on the watermark key and content integrity features, and generates a zero-knowledge proof.

[0039] Zero-knowledge proofs are used to prove that the watermark commitment, on-chain index, and content binding value are associated with the watermark key, and that the content binding value is associated with the content integrity feature.

[0040] S4 uses on-chain indexes as query identifiers to write content integrity features, watermark commitments, content binding values, metadata storage addresses, and zero-knowledge proofs into the blockchain, forming traceability credentials.

[0041] S5 embeds the on-chain index and content binding value as the watermark payload into the generated content to obtain the watermarked generated content.

[0042] This embodiment can be understood as follows: by synchronously generating watermark commitment, on-chain index and content binding value based on preset watermark key, and uniformly verifying with zero-knowledge proof that all three originate from the same key, and the content binding value is further associated with content integrity features, thus achieving trusted traceability, prevention of porting and misuse and multi-scale tampering location under limited watermark capacity, reducing the risk of generated content being ported and misused.

[0043] It should be noted that the devices involved in the AI-generated content management method provided in this implementation include an embedding end on the content generation side, blockchain network nodes, off-chain storage systems (e.g., IPFS nodes), and a verification end on the verification side. The embedding end can be a server deployed on the AIGC content generation service side, used to perform operations such as content acquisition, block division, perceptual hash calculation, Merkle tree construction, zero-knowledge commitment generation, watermark embedding, and on-chain calls. The verification end device can be any third-party terminal or server, used to perform watermark extraction, credential reading, zero-knowledge proof verification, and tamper location.

[0044] Taking a server as an example, the server can be a single server or a group of servers. The server group can be centralized or distributed (e.g., the servers can be a distributed system). In some embodiments, the server can be local or remote relative to the user terminal. In some embodiments, the server can be implemented on a cloud platform; by way of example only, a cloud platform can include private cloud, public cloud, hybrid cloud, community cloud, distributed cloud, inter-cloud, multi-cloud, etc., or any combination thereof. In some embodiments, the server can be implemented on an electronic device having one or more components.

[0045] To make the solution provided in this embodiment clearer, the following will continue to use a server as an example, and combine it with... Figure 1 Each step of the method is described in detail. However, it should be understood that the operations in the flowchart may not be implemented in sequence, and steps without logical contextual relationships may be reversed in order or implemented simultaneously. Furthermore, those skilled in the art, guided by the content of this application, may add one or more other operations to the flowchart, or remove one or more operations from the flowchart. Figure 1 As shown, the method includes: S1, obtain the generated content and the corresponding source metadata.

[0046] In practical applications, the server can obtain the generated content from the artificial intelligence model and simultaneously collect the source metadata directly associated with that content. Here, source metadata refers to key attributes that are determined and verifiable during the content generation process, including at least one of the following: generation model identifier, generation timestamp, copyright holder identifier, generation parameters, and prompt word hash.

[0047] For example, the generated model identifier can be the hash value of the model name and version number, the copyright subject identifier can be the creator's decentralized identity identifier (DID), and the generation parameters cover specific configurations that affect the output results, such as random seed, number of sampling steps and guidance coefficient; the prompt word hash is used to replace the plaintext prompt word, ensuring traceability while also protecting privacy.

[0048] Based on the above embodiments' description of generated content and traceability metadata, the following will continue to discuss... Figure 1 Step S2 will be explained in detail: S2 extracts content features from the generated content to obtain content integrity features, stores source metadata, and obtains the metadata storage address.

[0049] As an optional implementation, this embodiment provides the following optional implementation of step S2: S2-1, divide the generated content into multiple content blocks, calculate the perceptual hash of each content block, and obtain a list of perceptual hashes for the content blocks.

[0050] In practical applications, after obtaining the generated content, the server needs to divide the content into multiple content blocks and calculate the perceptual hash for each block, thereby establishing an independent content fingerprint for each block. It should be understood that different types of generated content require different block segmentation methods, and the server will select the appropriate segmentation strategy based on the content type.

[0051] When the generated content is an image, the server... A grid divides an image into multiple image blocks evenly. For example, it can be used to... This involves dividing an image into 64 image blocks. When generating video content, the server first extracts representative frames based on keyframes or fixed intervals, then performs the same grid division as for the image on each representative frame. Each content block is identified by a tuple consisting of a frame number and block coordinates, thus refining the accuracy of subsequent tamper location down to a specific region within a specific frame. When generating text content, the server divides the text by sentence or paragraph, using each text segment as a content block.

[0052] After completing the chunking, the server calculates a perceptual hash for each content chunk. Taking image content as an example, the server performs a hash calculation on each image chunk with a size of [size missing]. The two-dimensional discrete cosine transform (DCT) converts the content block from the spatial domain to the frequency domain, obtaining the frequency domain coefficient matrix. For pixel coordinates and With frequency domain coordinates and The frequency domain coefficients are calculated according to the following expression:

[0053] In the formula, and From 0 to , and These are the normalized coefficients of the discrete cosine transform, when the frequency domain coordinates are 0. When the frequency domain coordinates are taken from 1 to... hour Similarly, when the frequency domain coordinates take values ​​from 1 to... Time to take ,Right now .

[0054] During this process, the server retrieves the frequency domain coefficient matrix. The set of low-frequency coefficients in the upper left corner that does not contain DC components The coefficients in this set reflect the main structural information of the content blocks and exhibit strong stability against conventional propagation processes such as compression and scaling. The server calculates the mean of the coefficients in this set using the following formula:

[0055] Then the set Each low-frequency coefficient within the range is compared bit by bit with the mean and binarized to obtain a value of length [length missing]. Bit-perceptual hashing (Also known as Robust-aware hashing), in which equals set Number of elements:

[0056] The resulting perceptual hash maintains a small Hamming distance when the content undergoes imperceptible processing such as compression, scaling, or brightness adjustment, indicating robustness to such processing. However, when the content block is substantially replaced or altered, the perceptual hash will produce distinguishable differences.

[0057] For video content, the server performs the same block division and perceptual hash calculations on each representative frame as for the images described above. Content blocks are identified by a tuple of "frame number + block coordinates". For text content, the server divides the content into sentences or segments as content blocks and uses semantic fingerprints instead of the aforementioned perceptual hashing based on discrete cosine transform. These semantic fingerprints are calculated based on Locality Sensitive Hashing (LSH) of sentence vectors, remaining stable against paraphrasing, whitespace, and punctuation normalization, while producing detectable changes to substantial semantic alterations.

[0058] The server aggregates the perceptual hashes of all content blocks into a content block perceptual hash list. Each perceptual hash is associated with a corresponding content block.

[0059] Based on the above embodiments regarding content blocks and perceptual hashes, step S2 further includes: S2-2, construct a Merkle tree by using each perceptual hash in the content block perceptual hash list as a leaf node, and use the Merkle root of the Merkle tree as the content integrity feature.

[0060] In practical applications, the server uses each aware hash in the content block aware hash list. As a leaf node, a cryptographic hash function is used. The Merkle tree is constructed by calculating layer by layer from bottom to top. Specifically, the server concatenates two adjacent leaf nodes into a bit string and performs a hash operation to obtain the value of the parent node in the next higher level. The same concatenation and hashing operation is then repeated for the parent node, aggregating layer by layer upwards until a unique Merkle tree root is calculated. This layer-by-layer calculation process can be represented as:

[0061] In the formula, This indicates bit string concatenation. and These are the left and right child nodes of the current node, respectively. After all leaf nodes have participated in the calculation, the Merkle root is finally obtained:

[0062] In this process, cryptographic hash functions The national cryptographic algorithm SM3 or SHA-256 can be used. In this Merkle tree construction, any change in the perceptual hash of any content block will propagate upwards along its Merkle path, eventually reaching the Merkle tree root. Changes occur. Therefore, during the verification phase, a specific content block can be efficiently verified for tampering using only a Merkel path from a leaf node to the root of the tree, without having to recalculate the entire tree. This has a computational efficiency advantage when there are a large number of content blocks.

[0063] Based on the above embodiments' description of Merkel trees and Merkel tree roots, step S2 further includes: S2-3, store the traceability metadata in the off-chain storage system to obtain the metadata storage address.

[0064] It should be understood that the amount of source metadata usually far exceeds the capacity that watermarks can carry. Embedding it all into the content is neither practical nor will it improve the visual quality of the content. Therefore, it is necessary to use an off-chain storage system to carry this complete information.

[0065] In practical applications, the server will provide complete traceability metadata. The storage is located in the InterPlanetary File System (IPFS), and its content addressing feature is used to obtain the storage address. The source metadata here. This can include a generation model identifier, a generation timestamp, a copyright holder identifier, generation parameters, and a hash of the prompt word. The generation model identifier is the hash value of the model name and version number; the copyright holder identifier is the creator's decentralized identifier (DID); and the generation parameters include a random seed, the number of sampling steps, and a guidance coefficient. Here, the prompt word hash is used instead of the plaintext prompt word for storage to balance traceability and privacy protection.

[0066] During this process, to support the tracing of the original version of the subsequently tampered area, the server can also extract each original content block or abbreviated original content. Figure 1 And it is stored in an off-chain storage system to obtain the corresponding storage address. Thus, when the verification phase detects that a content block has been tampered with, the corresponding original content block can be read through that address, enabling a comparison of the content before and after the tampering.

[0067] Based on the above embodiments' descriptions of content integrity features and source metadata, please refer to... Figure 1, next, the description of step S3 in the figure will be continued: S3, generating a watermark commitment and an on-chain index based on a preset watermark key, generating a content binding value based on the watermark key and a content integrity feature, and generating a zero-knowledge proof.

[0068] The zero-knowledge proof is configured to prove that the watermark commitment, the on-chain index and the content binding value are associated with the watermark key, and the content binding value is associated with the content integrity feature.

[0069] It should be understood that the verification of traditional digital watermarks usually relies on the key held by the embedding party itself for comparison, and the verifying party and the embedding party belong to the same trust subject. This verification mechanism has a self-authentication trust defect. When a third party does not trust the embedding party, it cannot independently verify the authenticity of the watermark. It should be pointed out that once a watermark is extracted or copied, it can be transplanted to other people's content for fraudulent use, while the existing mechanism lacks cryptographic constraints on the legality of the watermark source, making it difficult for the verifying party to distinguish whether the watermark is originally embedded or copied and transplanted.

[0070] To solve the above problems, this embodiment further provides the following optional implementation manners of step S3: S3-1, generating a random blinding factor, and generating a watermark commitment based on the watermark key and the random blinding factor.

[0071] In practical applications, the server holds the watermark key , and the watermark key can be derived from the private key of the copyright subject. The server additionally selects a random blinding factor , based on the watermark key and the random blinding factor to jointly generate the watermark commitment . Specifically, the server uses an elliptic curve-based Pedersen commitment function for calculation, and its expression is as follows:

[0072] In the formula, and are two generators on the elliptic curve group. The Pedersen commitment function here makes it impossible to inversely deduce the watermark key only from the watermark commitment , so as to ensure that the watermark key is not leaked in the public verification process. The introduction of the random blinding factor makes even if the same watermark key is used in different scenarios, the generated watermark commitments are different, avoiding the security association risk caused by repeated commitment values.

[0073] Based on the above description of the watermark commitment in the embodiment, step S3 further includes: ​S3-2, perform a one-way hash operation on the combination of the watermark key and the preset separator string to obtain the on-chain index.

[0074] In practical applications, the server uses a watermark key. As input, through a one-way function Exporting on-chain indexes Its calculation expression is:

[0075] In the formula, It is the national cryptographic hash algorithm. This indicates bit string concatenation. This is a predefined, fixed delimiter string. The on-chain index here... This is the query identifier for the traceability certificate on the blockchain. The server will then use this index as the key to write the traceability certificate into the blockchain smart contract.

[0076] During this process, due to the one-way function It is irreversible, meaning it can only be reversed by on-chain indexes. Unable to deduce the watermark key The value; the preset delimiter string. The introduction of watermark keys By concatenating it with a fixed string before performing a hash operation, the watermark key is avoided. It is reused elsewhere, causing a collision.

[0077] Based on the above embodiments regarding the setting of index identifiers and on-chain indexes, step S3 further includes: S3-4: Use the watermark key to perform message authentication code calculation on the Merkle root tree to obtain the content binding value.

[0078] In practical applications, the server uses a watermark key. As the key, for Merkel root Perform message authentication code calculation to obtain content binding value. Its calculation expression is:

[0079] In the formula, For Hash-based Message Authentication Code (HMAC) is a key-based authentication method. The Merkel root represents the integrity fingerprint of the entire generated content.

[0080] Thus, values ​​are bound to content. By cryptographically binding the watermark key to the content integrity feature, only those who possess the watermark key at the same time can access it. And have the same Merkel tree roots Only one side can calculate the matching side. Value. When an attacker attempts to copy the watermark from the original content and transplant it to another Merkel tree root, it is... ( When verifying the content, the Merkle root of that content will be recalculated during the verification phase and compared with the records on the chain. A comparison was performed because The binding relationship prevents attackers from creating new tree roots. To forge a legitimate binding value.

[0081] Based on the above embodiments regarding on-chain indexes, step S3 further includes: S3-5 uses the watermark commitment, on-chain index, content binding value, and Merkle root as public inputs, and the watermark key and random blinding factor as secret inputs to generate a zero-knowledge proof.

[0082] This can be understood as the server generating watermark commitments separately. On-chain index and content binding value Next, a zero-knowledge proof needs to be generated so that a third party can verify the legality of the association between the above values ​​without obtaining the watermark key.

[0083] In practical applications, the public input for this proof is a watermark commitment. On-chain index Content binding value And Merkel roots The secret input is the watermark key. With random blinding factor Its generating expression is:

[0084] In the formula, This indicates the generation of a proof that the proposition within the curly braces is true. and This is a private secret of the prover and will not be disclosed. The existence of a certain watermark key can be proven using the above expression. and random blinding factor This makes the watermark commitment On-chain index and content binding value All three public values ​​are derived from the same watermark key. Export.

[0085] Therefore, through zero-knowledge proof This can prove the watermark commitment. Confirmed by watermark key With random blinding factor Calculated via the Pedersen commitment function; on-chain index Confirmed by watermark key The combination of the string and the preset delimiter is derived through a one-way hash operation; content is bound to the value. Confirmed by watermark key For Merkel's roots The message authentication code is obtained by performing a computation, thus cryptographically binding the watermark commitment, on-chain index, content binding value, and Merkle tree root together, while the watermark key... and random blinding factor It was kept completely secret throughout the entire proof process.

[0086] Based on the above embodiments' explanation of the zero-knowledge proof in step S3, please refer to... Figure 1 Next, we will continue to explain step S4 in the diagram: S4 uses on-chain indexes as query identifiers to write content integrity features, watermark commitments, content binding values, metadata storage addresses, and zero-knowledge proofs into the blockchain, forming traceability credentials.

[0087] In practical applications, the server invokes a blockchain smart contract using on-chain indexes. As a query identifier, content integrity features Watermark Promise Content binding value Metadata storage address and zero-knowledge proofs Write it into the blockchain to create a traceability certificate. The data organization format of this traceability certificate can be represented as follows:

[0088] During this process, the smart contract will first perform zero-knowledge proofs when receiving a write request. Perform verification to confirm the on-chain index. Content-bound values Confirmed by watermark The corresponding watermark key is exported. Once verification is successful, the smart contract uses an on-chain index. Store traceability credentials for indexing This allows the credential to be indexed on the chain. It is uniquely located on the blockchain.

[0089] In this way, even if an attacker attempts to construct a false evidence record on the blockchain, they will not be able to forge a zero-knowledge proof that matches the legitimate watermark commitment, thus ensuring the uniqueness and unforgeability of the traceability certificate.

[0090] Based on the above embodiments' explanation of on-chain indexes and content binding values, please refer to [link to previous document]. Figure 1 Next, we will continue to explain step S5 in the diagram: S5 embeds the on-chain index and content binding value as the watermark payload into the generated content to obtain the watermarked generated content.

[0091] It should be understood that while the previously constructed robust perceptual hash remains stable against imperceptible processing such as compression and scaling, when tampering involves only a very small range of local pixels within the content, the change in the perceptual hash may be insufficient to exceed the detection threshold, thus creating a detection blind spot. Furthermore, it should be noted that content faces diverse tampering methods during dissemination, ranging from large-area replacement or erasure to fine-grained editing targeting only a few pixels. A single detection mechanism cannot simultaneously achieve both the ability to detect large-area tampering and the sensitivity to detect small-area fine-grained tampering.

[0092] To address the above problems, this embodiment also provides the following optional implementation methods for step S5: S5-1 generates corresponding block-level verification information based on the perceptual hash and watermark key of each content block.

[0093] In practical applications, the server can use a watermark key. Using the key, the perceptual hash of each content block is encrypted to obtain the block-level verification information corresponding to each content block. Specifically, the server takes the first few bits of the perceptual hash of each content block and uses them as a watermark key. Encrypt it to obtain the block-level verification information of the block. .

[0094] During this process, block-level verification information Each perceptual hash corresponds one-to-one with a specific content block, representing a key protection digest that describes the content characteristics of that block. This is due to block-level verification information. The generation depends on the watermark key. Only the party possessing the legitimate watermark key can calculate the correct verification value, thus providing a key-level verification basis for determining whether the content block has been tampered with.

[0095] For example, when a content block suffers a small-scale, subtle alteration, its perceived hash may change slightly. This subtle change might not be detected in robust perceived hash Merkle verification because it does not exceed a threshold, but the block-level verification information... Inconsistencies can arise due to changes in the perceptual hash, thus compensating for the insufficient sensitivity of robust perceptual hash in detecting small-area tampering.

[0096] The server will ultimately process the block-level verification information for each content block. On-chain index in watermark payload and content binding value The joint organization forms the complete watermark information. Its data organization can be represented as:

[0097] In the formula, This represents bit string concatenation. Thus, the watermark information embedded in each content block includes both on-chain indexes and content binding values ​​for traceability, as well as block-level verification information for block-by-block integrity checks. .

[0098] Based on the above embodiments regarding the on-chain index and content binding value, step S5 further includes: S5-2 performs redundant encoding on-chain indexes and content binding values, and embeds them as robust watermark blocks into each content block.

[0099] In practical applications, the server first indexes the chain. Content-bound values Redundant encoding is employed to give these two key pieces of information a certain degree of fault tolerance within the content. Even if the content undergoes conventional processing such as compression or format conversion, the watermark payload can still be correctly recovered. This results in a watermarked image. Visually, it is similar to It is indistinguishable from other images and is robust to common image processing methods; therefore, it is called a robust watermark.

[0100] In this process, the server can use a transform domain spreading method combining Discrete Wavelet Transform (DWT) and Discrete Cosine Transform (DCT) to embed the watermark into each content block. Specifically, the server performs a first-level Discrete Wavelet Transform on each content block, then performs block-based Discrete Cosine Transform on the low-frequency sub-band, and embeds the watermark bits... With watermark key The generated spread spectrum sequence Superimposed on the intermediate frequency coefficient The watermark coefficient is obtained from the above. Its calculation expression is:

[0101] In the formula, Embedding strength is used to control the degree to which the watermark affects the visual quality of the content. For watermark key This is a pseudo-random sequence generator for the seed. Thus, the spreading sequence... Watermark key The uniqueness of the key ensures that the watermark embedding process is closely linked to the watermark key, and only the party that possesses the legitimate watermark key can correctly extract the watermark information during the verification phase.

[0102] because and Merkel tree roots that have been bound to the original content with zero-knowledge proofs The verification phase requires recalculating the Merkel root of the content to be verified and connecting it with... The comparison showed that when an attacker copied the entire watermark to Merkel Root Cloud, When on other images, the recalculated tree roots and If the watermark does not match, the misuse of the watermark will be detected. For video content, the watermark is embedded frame by frame in representative frames according to the DWT-DCT spread spectrum method described above. For text content, the watermark payload is... W The text segments are encoded using probabilistic modulation of word output or synonym substitution. The remaining key binding, on-chaining, and verification processes are the same as those for images.

[0103] Based on the above description of block-level verification information in the embodiments, step S5 further includes: S5-3 embeds the block-level verification information as a semi-fragile watermark into the corresponding content block to obtain the watermarked generated content.

[0104] In practical applications, the server will store block-level verification information for each content block. This semi-fragile watermark is embedded within the corresponding content block. Compared to the robust watermark mentioned above, the robust watermark is more resistant to conventional processing such as compression and scaling, while the semi-fragile watermark is more sensitive to changes in any local pixels within the block. When only a small area of ​​fine-grained tampering occurs within the content block, the semi-fragile watermark will be destroyed.

[0105] In this process, the embedding of semi-fragile watermarks ensures that each content block carries a verification credential associated with its own perceptual hash. This is due to the block-level verification information. Based on the perceptual hash of the content block and the watermark key The encrypted watermark, after being extracted by the verification end, can recalculate the checksum based on the perceptual hash of the current content block and compare it with the extracted checksum information. When a content block suffers a small-area, subtle tampering, its perceptual hash may change slightly. This subtle change may not be detected in robust perceptual hash Merkle verification because the weighted mismatch distance does not exceed the threshold. However, the semi-fragile watermark will be damaged by changes in local pixels within the block, leading to block-level verification inconsistency, thus detecting the tampering.

[0106] Thus, robust watermarks are responsible for carrying traceability information such as on-chain indexes and content binding values, and are robust to routine processing; semi-fragile watermarks are responsible for carrying block-level verification information. It is sensitive to local modifications within the block.

[0107] In the above embodiments, the process of generating traceability credentials for the content and adding watermarks has been explained. Next, the verification process of the content to be verified will be described.

[0108] like Figure 2 As shown, the method for managing AI-generated content provided in this embodiment further includes: S6 extracts the on-chain index and the binding value of the content to be verified from the content to be verified.

[0109] S7 uses the on-chain index to be verified as the query identifier to read the corresponding target traceability certificate from the blockchain.

[0110] S8. Based on the watermark commitment, content integrity features, on-chain index to be verified, and content binding value to be verified in the target traceability certificate, verify the zero-knowledge proof therein, and compare the content binding value to be verified with the content binding value in the target traceability certificate to obtain the watermark authenticity verification result.

[0111] This embodiment can be understood as follows: through the verification mechanism of zero-knowledge proof, the verification end can confirm whether the watermark was generated by the legitimate embedder for the current content without obtaining the watermark key, thereby eliminating the self-proving trust defect. This means that even if an attacker copies the watermark from the original content and transplants it to other content, it will be detected in the verification stage due to the mismatch of content integrity features.

[0112] Similarly, to make the solution provided in this embodiment clearer, we will continue to use a server as an example below, and combine it with... Figure 2 Each step of the method is described in detail. For example... Figure 2 As shown, the method includes: S6 extracts the on-chain index and the binding value of the content to be verified from the content to be verified.

[0113] In practical applications, the server follows the same procedure as the embedding phase. The grid partitioning method divides the content to be verified into several content blocks. For example, when the content to be verified is an image, the server uses the same method as in the embedding stage. The grid divides it into 64 image blocks. The server also uses a watermark key. Irrelevant blind or semi-blind extraction methods are used to extract watermark payloads from each content block using a synchronous sequence structure.

[0114] In this process, the blind extraction method means that the extraction process does not depend on the watermark key used in the embedding stage. The verification end can recover the watermark information from the content block without prior knowledge of the watermark key. The server extracts and recovers the on-chain index to be verified from each content block. Content to be verified bound value And the block-level verification information of each content block. The extraction process can be represented as follows:

[0115] In the formula, This indicates content to be verified. This represents the watermark extraction function.

[0116] It should be noted that the extraction process here utilizes a synchronous sequence structure without needing to know the watermark key. In cases where the watermark bits are recovered, the embedded robust watermark can still be correctly extracted even after the content to be verified has undergone conventional propagation processing such as compression and scaling.

[0117] Based on the above embodiments, regarding the on-chain index to be verified Content to be verified bound value And the block-level verification information of each content block. For further explanation, please refer to [link / reference]. Figure 2 Next, we will continue to explain step S7: S7 uses the on-chain index to be verified as the query identifier to read the corresponding target traceability certificate from the blockchain.

[0118] In practical applications, the server uses the on-chain index to be verified extracted from the watermark. As a query identifier, the blockchain smart contract's read interface is invoked to retrieve the corresponding target traceability certificate from the blockchain. This read process can be represented as follows:

[0119] In the formula, This represents an on-chain data read function based on an on-chain index. The target traceability certificate that has been read.

[0120] Among them, the traceability certificate for the target includes Merkel root. This refers to the Merkle root tree, previously constructed by the server based on the perceptual hashes of each content block, representing the integrity benchmark of the original generated content; the watermark promises... This refers to the Pedersen commitment value generated by the server based on the watermark key and a random blinding factor; the metadata storage address of the content to be verified. This refers to the content addressing address of complete traceability metadata in the off-chain storage system; content binding value. That is, the binding value obtained by the server performing message authentication code calculation on the Merkle tree root using the watermark key as the key; and zero-knowledge proof. This refers to the zero-knowledge proof previously generated by the server to prove the legality of the association between various parameters.

[0121] In this way, the server obtains complete traceability credentials that can be independently verified by a third party through on-chain indexes and the immutability of blockchain.

[0122] S8. Based on the watermark commitment, content integrity features, on-chain index to be verified, and content binding value to be verified in the target traceability certificate, verify the zero-knowledge proof therein, and compare the content binding value to be verified with the content binding value in the target traceability certificate to obtain the watermark authenticity verification result.

[0123] In practical applications, after the server reads the target traceability certificate from the blockchain, it needs to confirm that the watermark extracted from the content to be verified was indeed generated by the legitimate embedding party for that content, and not copied or ported from other content. To this end, the server can verify the zero-knowledge proof carried in the target traceability certificate. This zero-knowledge proof can prove the existence of a watermark key and a random blinding factor, such that the watermark commitment, on-chain index, and content binding value are all derived from the same watermark key, and that the content binding value is associated with the content integrity feature.

[0124] Specifically, in the aforementioned content generation and watermark embedding stages, the watermark key has been utilized. And select random blinding factor Watermark commitments are generated according to the Pedersen commitment function based on elliptic curves. ,in , For group generators, For watermarking key, This is the random blinding factor.

[0125] The on-chain index is derived from the watermark key via a predefined one-way function, and is represented as follows: Here, "index" is a preset separator string used to prevent collisions when the watermark key is reused in other scenarios. The content binding value is obtained by calculating the content integrity feature using the watermark key as the message authentication code key. The content integrity feature is the Merkle root constructed from the perceptual hashes of each content block. The corresponding operation is represented as ,in For The keyed message authentication code is the key. The Merkel root represents the baseline for the integrity of the entire content.

[0126] Zero-knowledge proof The calculation process can be found in the above embodiments, and will not be repeated here. This proof process can demonstrate the existence of a certain... and This makes all three public values ​​determined by the same Export, but do not leak. and In itself, any third party can verify that the proof is valid and confirm that the on-chain index and content binding values ​​are derived by the legitimate embedding party that possesses the key corresponding to the watermark commitment.

[0127] During the verification phase, the server can use the on-chain index to be verified and the binding value of the content to be verified as the value extracted from the watermark of the content to be verified, and use the watermark commitment in the target traceability certificate. and Merkel roots These publicly available values, already registered on the blockchain, serve as the verification inputs for the zero-knowledge proof. The verification process can be represented as follows:

[0128] In the formula, For zero-knowledge proof, For the watermark commitment, For the on-chain index to be verified, Bind a value to the content to be verified. Merkel's root in the traceability certificate for the target.

[0129] If the verification passes, it indicates that both the on-chain index and the content binding value to be verified were derived by the legitimate embedding party possessing the key corresponding to the watermark commitment, and that the content binding value to be verified was indeed calculated from the Merkle root using that key. Since the binding relationship between the content binding value and the Merkle root has been confirmed by a zero-knowledge proof, the server does not need to obtain the watermark key to confirm that the binding is valid.

[0130] Simultaneously, the server compares the binding value of the content to be verified with the binding value recorded in the target traceability certificate. Only when the zero-knowledge proof verification passes and the two match is the watermark authenticity verification result considered passed, meaning the watermark was indeed generated by the legitimate embedding party for the current content. If the zero-knowledge proof verification fails, or the binding value of the content to be verified is inconsistent with the binding value in the on-chain certificate, it indicates that the watermark was not generated by the embedding party possessing the legitimate watermark key for the current content. For example, the watermark may have been copied or misused from other content. In this case, the watermark authenticity verification result is deemed failed, and the content source is untrustworthy.

[0131] After the watermark authenticity verification result indicates that the verification has passed, such as Figure 3 As shown, as an optional implementation, the method for managing AI-generated content provided in this embodiment further includes: S9, divide the content to be verified into multiple content blocks to be verified, and calculate the verification-aware hash of each content block to be verified. S10, Merkel path verification is performed based on the Merkel root in each perceptual hash to be verified and the target traceability certificate; S11, the content block to be verified through the Merkel path is taken as the target content block, and the semi-fragile watermark of the target content block is compared with the block-level verification information of the target content block to be verified, so as to obtain the block-level verification result, and the tampered content block is determined according to the block-level verification result.

[0132] This can be understood as follows: previously, watermark authenticity verification could only confirm whether the watermark was generated by a legitimate embedder for the current content, but it could not determine whether the content itself had been tampered with or where the tampering occurred.

[0133] During this process, the server re-divides the content to be verified into multiple content blocks using the same chunking method as in the embedding phase, and calculates the perceptual hash for each block. Since changes to any leaf node in the Merkle tree propagate upwards along the path, the server can determine whether a specific content block has been tampered with and pinpoint its location without rebuilding the entire tree. Simultaneously, the server compares the verification information of each block to be verified with the verification value regenerated based on the current perceptual hash. When a content block undergoes local modification, the block-level verification results will be inconsistent.

[0134] Similarly, to make the solution provided in this embodiment clearer, we will continue to use a server as an example below, and combine it with... Figure 2 Each step of the method is described in detail. For example... Figure 3 As shown, the method includes: S9 divides the content to be verified into multiple content blocks to be verified, and calculates the verification-aware hash of each content block to be verified.

[0135] Specifically, the server can divide the content to be verified into multiple blocks using the same chunking method as in the embedding phase. Taking images as an example, it can be divided into... The grid is divided into sections, and in this example, we take... This means dividing the image into 64 image blocks.

[0136] After segmentation, the server calculates a verification-aware hash for each content block to be verified. This verification-aware hash can be understood as a recalculation of the robust-aware hash used in the embedding stage, which remains stable against imperceptible processing but is sensitive to substantial tampering. Taking image content as an example, the server can perform verification on each image block... Perform the size is The two-dimensional discrete cosine transform (DCT) yields the frequency domain coefficient matrix. ; and further take the frequency domain coefficient matrix The set of low-frequency coefficients in the upper left corner that do not contain DC components. Calculate their mean, and then set Each coefficient within the hash is compared bit by bit with the mean and binarized to obtain the perceptual hash to be verified. The relevant calculation methods can be found in the above embodiments, and will not be repeated here.

[0137] S10, Merkel path verification is performed based on the Merkel root in each perceptual hash to be verified and the target traceability certificate.

[0138] After obtaining the perceptual hash of each content block to be verified, the server can perform Merkel path verification with the Merkel root recorded in the target traceability certificate to determine whether each content block has been tampered with and to determine the location of the tampering.

[0139] It should be understood that the basic principle of Merkle path verification is that during the embedding phase, a Merkle tree is constructed from the bottom up using the perceptual hash of each content block as leaf nodes. Changes to any leaf node will propagate upwards along its Merkle path to the root of the tree. Therefore, the server does not need to rebuild the entire tree; it only needs to calculate upwards along its corresponding Merkle path for each content block to be verified, and compare the result with the Merkle tree root registered on the chain to determine whether the content block has been tampered with.

[0140] Considering the varying stability of different frequency coefficients under compression and scaling processes, the server can avoid using equal-weighted Hamming distance during the comparison process. Instead, it can introduce a reliability weight to each bit of the perceptual hash to be verified, and then sum the bit-by-bit differences in a weighted manner to obtain the weighted mismatch distance.

[0141]

[0142] In the formula, The first of the perceptual hashes of the content block to be verified bits, The first perceptual hash obtained through Merkel path verification bits, This represents the XOR operation. To perceive the length of the hash, For the first The reliability weight of each bit.

[0143] Reliability weight here It can be calculated using the following formula:

[0144] In the formula, For the content block to be verified A low-frequency coefficient, and These represent the mean and standard deviation of the low-frequency coefficients in the block, respectively. This weight characterizes the reliability of the corresponding bit. Coefficients far from the mean are stable in sign and have high weights under imperceptible processing such as compression, while coefficients close to the mean are prone to flipping and have low weights. Thus, the weighted mismatch distance weakens the low-reliability bits susceptible to compression, while strengthening the high-reliability bits that truly reflect content changes. This improves the sensitivity to tamper detection while suppressing false positives caused by conventional processing such as compression.

[0145] Based on this, the server will weight the mismatch distance. Adaptive threshold based on content type The comparison is performed, and block-level verification is used to determine tampering. The expression is:

[0146] In the formula, the threshold Values ​​are assigned based on content type, with separate values ​​for images, videos, and text. , , .when No more than If the block-level verification is consistent, the content block is determined to be unaltered, that is, it becomes the target content block through Merkel path verification; otherwise, the content block is determined to have been tampered with on a large scale.

[0147] Furthermore, to balance robustness to compression with sensitivity to small-area tampering, the server can also employ a hierarchical, multi-scale block strategy. Specifically, the image is first divided into... The coarse blocks are designed to resist compression and detect large-area tampering, and then each coarse block is further subdivided into... The second-level perceptual hash is calculated in fine blocks. When a small-area tampering occurs within a certain fine block, the tampering only affects the area occupying the entire block. Figure 1 The proportion of / 256 in the fine block is significantly increased, which is enough to make the perception hash of the fine block exceed the threshold and be detected, while refining the tamper location accuracy from coarse block to fine block.

[0148] S11, the content block to be verified through the Merkel path is taken as the target content block, and the semi-fragile watermark of the target content block is compared with the block-level verification information of the target content block to be verified, so as to obtain the block-level verification result, and the tampered content block is determined according to the block-level verification result.

[0149] After verification via the Merkel path, the server can classify each content block to be verified into two categories. Content blocks whose weighted mismatch distance does not exceed the content type adaptive threshold and are initially determined not to have been tampered with are designated as target content blocks. A second layer of tamper detection is then performed on these target content blocks.

[0150] It should be noted that although robust perceptual hashing remains stable against imperceptible processing such as compression and scaling, it is not sensitive to small-area fine-grained tampering that only modifies a few pixels. The amount of change in its perceptual hash may still be below the judgment threshold, thus causing missed detection.

[0151] Therefore, this embodiment introduces semi-fragile watermarking as a complementary means. It should be understood that, during the embedding stage, in addition to the on-chain index and content binding value, the watermark payload of each content block also carries block-level verification information. This verification information is embedded into the corresponding content block along with the semi-fragile watermark. The semi-fragile watermark is sensitive to any local pixel changes within the block. Once the content block suffers small-area alteration or fine editing, its semi-fragile watermark is destroyed.

[0152] During the verification phase, the server can extract the embedded semi-fragile watermark from each target content block and recover the block-level verification information, while also using the watermark key. The current perceived hash of the target content block is recalculated to obtain the correct checksum, and the two are compared. If the semi-fragile watermark of a target content block has been destroyed, making it impossible to correctly recover the checksum information, or if the recovered checksum information is inconsistent with the checksum recalculated based on the watermark key, then the content block is determined to have undergone small-area fine-grained tampering.

[0153] Robust-aware hash Merkle verification detects large-scale tampering without false positives from conventional processing like compression, while semi-fragile watermark block-level verification captures small-scale tampering missed by the former. The two layers complement each other. When either layer determines tampering, it records the grid coordinates of the content block. The server ultimately outputs the set of positions of all tampered content blocks within the content to be verified as the tampering location result. For example, the tampered area is marked with a highlight mask. For video content, the tampering location result can be accurate to a specific frame and its intra-frame region; for text content, it locates the tampered text segment.

[0154] As an optional implementation, the source metadata of the content to be verified is also associated with the original content or recovery data of each content block to be verified, after the content block to be verified is determined.

[0155] The server can also read the source metadata of the content to be verified based on the metadata storage address of the content to be verified; read the original content or restored data corresponding to the tampered content block based on the source metadata of the content to be verified; and generate a source traceability output result containing the content before and after the tampering based on the original content or restored data corresponding to the tampered content block.

[0156] Specifically, the server can use the metadata storage address recorded in the target traceability certificate. Read complete traceability metadata from off-chain storage system For content blocks identified as tampered with, the server can also trace their original versions for comparison. It should be understood that, during the embedding phase, to support tracing the original versions of tampered areas, each original content block or original content thumbnail can be stored in an off-chain storage system to obtain the corresponding original block storage address. This address is also recorded in the tracing certificate. After identifying the tampered content block, the server can, based on... The corresponding original content block is read from off-chain storage, and the original version is output and compared with the currently tampered content to achieve the original version tracing of the tampered area.

[0157] It should also be noted that in scenarios where off-chain storage is unavailable, the server can employ a self-recovery method that does not rely on off-chain storage. During the embedding phase, highly compressed low-resolution recovery information is redundantly embedded in the watermarks of each content block. This recovery information is distributed and embedded into the watermarks of other content blocks. When a content block is tampered with, the server can extract this redundant compressed recovery information from the watermarks of other untampered content blocks to reconstruct an approximate original version of the tampered content block. In this way, even if off-chain storage is temporarily inaccessible, the server can still independently complete an approximate recovery of the original version.

[0158] Ultimately, the server can generate a source tracing output that includes the content before and after the tampering. This source tracing output includes at least source metadata, the location identifier of the tampered content block, the corresponding original content or approximately restored content, and a visual annotation of the tampering location, such as highlighting the tampered area with a highlight mask, so that viewers can intuitively understand the source of the content, the location of the tampering, and the original state.

[0159] Based on the same inventive concept as the AI-generated content management method provided in this embodiment, this embodiment also provides an AI-generated content management device. This device includes at least one software functional module that can be stored in a memory or embedded in an electronic device. A processor in the electronic device executes the executable module stored in the memory. For example, the software functional module and computer program included in this device. Please refer to... Figure 4 Functionally, the device may include: The raw data module 11 is used to obtain the generated content and the corresponding source metadata. The content processing module 12 is used to extract content features from the generated content, obtain content integrity features, store source traceability metadata, and obtain the metadata storage address; The traceability information module 13 is used to generate a watermark commitment and on-chain index based on a preset watermark key, generate a content binding value based on the watermark key and content integrity features, and generate a zero-knowledge proof. The zero-knowledge proof is used to prove that the watermark commitment, on-chain index and content binding value are associated with the watermark key, and that the content binding value is associated with the content integrity features. The traceability information module 13 is also used to write content integrity features, watermark commitments, content binding values, metadata storage addresses and zero-knowledge proofs into the blockchain based on the on-chain index as a query identifier, forming traceability credentials. The watermark generation module 14 is used to embed the on-chain index and content binding value as watermark payload into the generated content to obtain watermarked generated content.

[0160] Optionally, the content processing module 12 extracts content features from the generated content to obtain content integrity features and stores source metadata. The methods for obtaining the metadata storage address include: The generated content is divided into multiple content blocks, and the perceptual hash of each content block is calculated to obtain a list of perceptual hashes for the content blocks. Construct a Merkle tree by using each perceptual hash in the content block perceptual hash list as a leaf node, and use the Merkle root of the Merkle tree as the content integrity feature. The source traceability metadata is stored in an off-chain storage system, and the metadata storage address is obtained.

[0161] Optionally, the traceability information module 13 generates a watermark commitment and on-chain index based on a preset watermark key, generates a content binding value based on the watermark key and content integrity features, and generates zero-knowledge proofs, including: Generate a random blinding factor, and generate a watermark commitment based on the watermark key and the random blinding factor; Perform a one-way hash operation on the combination of the watermark key and the preset separator string to obtain the on-chain index; The message authentication code is calculated using the watermark key on the Merkle root to obtain the content binding value; Zero-knowledge proofs are generated by using the watermark commitment, on-chain index, content binding value, and Merkle root as public inputs, and the watermark key and random blinding factor as secret inputs.

[0162] Optionally, the watermark generation module 14 embeds the on-chain index and content binding value as the watermark payload into the generated content to obtain the watermarked generated content in the following ways: Generate corresponding block-level verification information based on the perceptual hash and watermark key of each content block; Redundant encoding is performed on the on-chain index and content binding values, and these are embedded as robust watermark blocks into each content block. Each block-level verification information is embedded as a semi-fragile watermark into the corresponding content block to obtain the watermarked generated content.

[0163] Optionally, the device further includes: Content verification module 15 is used to extract the index of the verification chain and the binding value of the content to be verified from the content to be verified; Based on the watermark commitment, content integrity features, on-chain index to be verified, and content binding value to be verified in the target traceability certificate, the zero-knowledge proof is verified, and the content binding value to be verified is compared with the content binding value in the target traceability certificate to obtain the watermark authenticity verification result.

[0164] Optionally, after the watermark authenticity verification result indicates that the verification has passed, the content verification module 15 is further used to divide the content to be verified into multiple content blocks to be verified, and calculate the perceptual hash of each content block to be verified. Merkel path verification is performed based on the Merkel tree root in each perceptual hash to be verified and the target traceability certificate; The content block to be verified through the Merkel path is used as the target content block. The semi-fragile watermark of the target content block is compared with the block-level verification information of the target content block to obtain the block-level verification result. The tampered content block is determined based on the block-level verification result.

[0165] Optionally, the source metadata of the content to be verified is also associated with the original content or recovery data of each content block to be verified. After the tampered content block is determined, the content verification module 15 is also used to read the source metadata of the content to be verified according to the metadata storage address of the content to be verified. Based on the source metadata of the content to be verified, read the original content or recover the data corresponding to the tampered content block; Based on the original content or restored data corresponding to the tampered content block, generate a source tracing output that includes the content before and after the tampering.

[0166] In addition, the functional modules in the various embodiments of this application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0167] It should also be understood that if the above embodiments are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application.

[0168] Therefore, this embodiment also provides a storage medium, which is a computer-readable storage medium. This storage medium stores a computer program, which, when executed by a processor, implements the artificial intelligence-generated content management method provided in this embodiment. The storage medium can be any medium capable of storing program code, such as a USB flash drive, external hard drive, read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk.

[0169] Please refer to Figure 5 The electronic device provided in this embodiment may include a processor 22 and a memory 21. The memory 21 stores a computer program, and the processor reads and executes the computer program in the memory 21 corresponding to the above-described embodiments to implement the artificial intelligence-generated content management method provided in this embodiment.

[0170] See also Figure 5 The electronic device also includes a communication unit 23. The memory 21, processor 22 and communication unit 23 are electrically connected to each other directly or indirectly through system bus 24 to realize data transmission or interaction.

[0171] The memory 21 can be an information recording device based on any electronic, magnetic, optical, or other physical principles, used to record execution instructions, data, etc. In some embodiments, the memory 21 can be, but is not limited to, volatile memory, non-volatile memory, memory drive, etc.

[0172] In some embodiments, the volatile memory may be random access memory (RAM); in some embodiments, the non-volatile memory may be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), flash memory, etc.; in some embodiments, the storage drive may be a disk drive, solid-state drive, any type of storage disk (such as optical disc, DVD, etc.), or similar storage media, or a combination thereof.

[0173] The communication unit 23 is used to send and receive data over a network. In some embodiments, the network may include a wired network, a wireless network, a fiber optic network, a telecommunications network, an intranet, the Internet, a local area network (LAN), a wide area network (WAN), a wireless local area network (WLAN), a metropolitan area network (MAN), a public switched telephone network (PSTN), a Bluetooth network, a ZigBee network, or a near field communication (NFC) network, or any combination thereof. In some embodiments, the network may include one or more network access points. For example, the network may include wired or wireless network access points, such as base stations and / or network switching nodes, through which one or more components of the service request processing system can connect to the network to exchange data and / or information.

[0174] The processor 22 may be an integrated circuit chip with signal processing capabilities, and may include one or more processing cores (e.g., a single-core processor or a multi-core processor). By way of example only, the processor described above may include a Central Processing Unit (CPU), an Application Specific Integrated Circuit (ASIC), an Application Specific Instruction-set Processor (ASIP), a Graphics Processing Unit (GPU), a Physics Processing Unit (PPU), a Digital Signal Processor (DSP), a Field Programmable Gate Array (FPGA), a Programmable Logic Device (PLD), a controller, a microcontroller unit, a Reduced Instruction Set Computing (RISC) computer, or a microprocessor, or any combination thereof.

[0175] Understandable. Figure 5 The structure shown is for illustrative purposes only. Electronic devices may also have more advanced features. Figure 5 Showing more or fewer components, or having with Figure 5 The different configurations shown. Figure 5 The components shown can be implemented using hardware, software, or a combination thereof.

[0176] It should be understood that the apparatus and methods disclosed in the above embodiments can also be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the flowcharts and block diagrams in the accompanying drawings show the architecture, functions, and operations of possible implementations of apparatus, methods, and computer program products according to various embodiments of this application. In this regard, each block in the flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, or they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagram and / or flowchart, and combinations of blocks in the block diagram and / or flowchart, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0177] The above descriptions are merely various embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for managing artificial intelligence-generated content, characterized in that, The method includes: Obtain the generated content and the corresponding source metadata; The generated content is subjected to content feature extraction to obtain content integrity features, and the source tracing metadata is stored to obtain the metadata storage address; A watermark commitment and on-chain index are generated based on a preset watermark key. A content binding value is generated based on the watermark key and the content integrity feature. A zero-knowledge proof is generated, wherein the zero-knowledge proof is used to prove that the watermark commitment, the on-chain index and the content binding value are associated with the watermark key, and that the content binding value is associated with the content integrity feature. Based on the on-chain index as the query identifier, the content integrity feature, the watermark commitment, the content binding value, the metadata storage address, and the zero-knowledge proof are written into the blockchain to form a traceability certificate. The on-chain index and the content binding value are embedded as watermark payloads into the generated content to obtain watermarked generated content.

2. The method for managing AI-generated content according to claim 1, characterized in that, The generated content is subjected to content feature extraction to obtain content integrity features, and the source tracing metadata is stored to obtain the metadata storage address, including: The generated content is divided into multiple content blocks, and the perceptual hash of each content block is calculated to obtain a list of perceptual hashes for the content blocks. Construct a Merkle tree by using each perceptual hash in the content block perceptual hash list as a leaf node, and use the Merkle root of the Merkle tree as the content integrity feature. The source traceability metadata is stored in an off-chain storage system to obtain the metadata storage address.

3. The method for managing AI-generated content according to claim 2, characterized in that, A watermark commitment and on-chain index are generated based on a preset watermark key. A content binding value is generated based on the watermark key and the content integrity feature, and a zero-knowledge proof is generated, including: Generate a random blinding factor, and generate the watermark commitment based on the watermark key and the random blinding factor; The combination of the watermark key and the preset separator string is subjected to a one-way hash operation to obtain the on-chain index; The content binding value is obtained by performing message authentication code calculation on the Merkle root using the watermark key. The zero-knowledge proof is generated by using the watermark commitment, the on-chain index, the content binding value, and the Merkle root as public inputs, and the watermark key and the random blinding factor as secret inputs.

4. The method for managing AI-generated content according to claim 3, characterized in that, The on-chain index and the content binding value are embedded as watermark payloads into the generated content to obtain watermarked generated content, including: Based on the perceptual hash of each content block and the watermark key, corresponding block-level verification information is generated; The on-chain index and the content binding value are redundantly encoded and embedded as robust watermark blocks into each content block; Each block-level verification information is embedded as a semi-fragile watermark into the corresponding content block to obtain the watermarked generated content.

5. The method for managing AI-generated content according to claim 4, characterized in that, The method further includes: Extract the on-chain index and the bound value of the content to be verified from the content to be verified. Use the on-chain index to be verified as a query identifier to read the corresponding target traceability certificate from the blockchain; Based on the watermark commitment, content integrity features, on-chain index to be verified, and content binding value to be verified in the target traceability certificate, the zero-knowledge proof is verified, and the content binding value to be verified is compared with the content binding value in the target traceability certificate to obtain the watermark authenticity verification result.

6. The method for managing AI-generated content according to claim 5, characterized in that, After the watermark authenticity verification result indicates that the verification has passed, the method further includes: The content to be verified is divided into multiple content blocks to be verified, and the perceptual hash of each content block to be verified is calculated. Merkel path verification is performed based on the Merkel tree root in each of the perceptual hashes to be verified and the target traceability certificate. The content block to be verified through the Merkel path is taken as the target content block, and the semi-fragile watermark of the target content block is compared with the block-level verification information of the target content block to be verified to obtain the block-level verification result. The tampered content block is determined based on the block-level verification result.

7. The method for managing AI-generated content according to claim 6, characterized in that, The source metadata of the content to be verified is also associated with the original content or recovery data of each content block to be verified. After determining the tampered content block, the method further includes: The source metadata of the content to be verified is read according to the metadata storage address of the content to be verified; Based on the source metadata of the content to be verified, read the original content or restored data corresponding to the tampered content block; Based on the original content or restored data corresponding to the tampered content block, a traceability output result containing the content before and after the tampering is generated.

8. A management device for artificial intelligence-generated content, characterized in that, The device includes: The raw data module is used to acquire the generated content and the source metadata corresponding to the generated content; The content processing module is used to extract content features from the generated content to obtain content integrity features, store the source traceability metadata, and obtain the metadata storage address; The traceability information module is used to generate a watermark commitment and an on-chain index based on a preset watermark key, generate a content binding value based on the watermark key and the content integrity feature, and generate a zero-knowledge proof, wherein the zero-knowledge proof is used to prove that the watermark commitment, the on-chain index and the content binding value are associated with the watermark key, and that the content binding value is associated with the content integrity feature; The traceability information module is also used to write the content integrity feature, the watermark commitment, the content binding value, the metadata storage address and the zero-knowledge proof into the blockchain based on the on-chain index as the query identifier, thereby forming a traceability certificate; The watermark generation module is used to embed the on-chain index and the content binding value as watermark payload into the generated content to obtain watermarked generated content.

9. A storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the management method for artificial intelligence-generated content as described in any one of claims 1-7.

10. An electronic device, characterized in that, The electronic device includes a processor and a memory, the memory storing a computer program, which, when executed by the processor, implements the management method for artificial intelligence-generated content as described in any one of claims 1-7.

Images