Block chain-based artificial intelligence content multimedia digital watermark generation method

Abstract

According to the block chain-based artificial intelligence content multimedia digital watermark generation method provided by the invention, ACGI watermarks are embedded by using the nonce value and the user intervention record of the block chain, digital right confirmation is carried out while the digital watermarks with better robustness are embedded into output contents, and functions of affiliation authentication, traceability tracking, infringement prevention and the like are realized.

Description

Technical Field

[0001] The present invention relates to a method for generating a digital watermark for AIGC based on blockchain technology. Background Art

[0002] The pain points of AIGC, or artificial intelligence generated content multimedia, include: 1. The rights of copyright holders or personal privacy are difficult to protect 1.1 Some AI platforms will sign agreements with copyright holders regarding copyright issues in training data to ensure compliance with the copyright holders’ requirements when using training data. However, copyright holders may request to stop training on specific content, and it is difficult for copyright holders to independently verify the results of the platform’s removal of copyright holders’ data or cessation of training on copyright privacy data.

[0003] 1.2 Some AI platforms may lack transparency and accountability in the processing of personal data and copyrighted content, making it difficult for copyright holders to understand the processing and use of personal data and copyrighted content, and thus difficult to protect their rights and interests; 1.3 Generative AI service providers should abide by copyright laws and regulations, respect the copyright rights of others, and not use the copyrighted content of others without authorization. Individuals or copyright holders also want to know whether the AI ​​platform uses their data for training.

[0004] 2. Copyright issues of AIGC At present, the laws and regulations regarding AIGC technology and the copyright issues it brings are still not perfect. The content generated by AIGC technology often involves a combination of multiple elements and data, which makes the ownership of the content complicated and ambiguous. Since AIGC technology is highly innovative and creative, the content generated by AIGC technology itself may also become an object that needs copyright protection. When handling similar cases, judicial procedures have difficulty in obtaining credible materials to analyze AIGC content.

[0005] 3. AIGC content fraud problem AIGC's deep fakes can produce nearly real photos, movies, and audios that can be used to depict non-existent events or individuals. If not prevented, it will inevitably bring new risks to national security, citizens, and corporate rights. For example, if someone uses deep fake technology to create false photos or audios, it may mislead the public's understanding of certain events or people, thereby affecting social stability. Using deep fake technology to forge identities or forge documents may cause unnecessary troubles and losses to citizens and businesses.

[0006] Problems and difficulties in combining blockchain digital watermarking with generative artificial intelligence: 1. If every AIGC content is to be put on the chain, there will be too many AIGC actions and the burden of putting on the chain will be too heavy, which will easily cause transaction congestion, and the blockchain will be difficult to apply to high-frequency and efficient scenarios. Instead of testing the concurrent processing capabilities of the blockchain, it is better to modify the adaptive consensus method or transaction process.

[0007] 2. If we only use batch on-chain processing to deal with the problem, there may be multiple user behavior data included together, or it may be difficult to locate which block the on-chain content is in. This is because when the watermark is embedded, the training data has not yet been on-chain.

[0008] 3. Blockchain is an open and transparent distributed ledger in which the data stored is visible to all participants. However, generative AI models may need to process sensitive data, such as personal identity information, business secrets, etc. Ensuring the privacy and security of this on-chain data is a key challenge. Summary of the invention

[0009] In response to the above situation, this article describes a blockchain-based multimedia digital watermarking solution for AI-generated content. It optimizes the transaction process through Nonce value processing in blockchain technology, combines data marker labels in the training data set, and records the training traces and human intervention behaviors in AIGC behavior on the chain. While embedding a more robust digital watermark into the output content, it performs digital rights confirmation to achieve functions such as ownership authentication, traceability tracking, and infringement prevention.

[0010] The present invention is achieved through the following technical solutions: A method for generating content multimedia digital watermarks based on blockchain using artificial intelligence: Including alliance chain, The consortium chain includes Group A and Group B, where Group A is the consortium chain sub-chain related to AIGC work information, and Group B is the consortium sub-chain related to digital identification tags; The consortium chain pregenerates nonce values ​​for users; Users use the AI ​​platform linked to the consortium chain to train and generate content, and the AI ​​platform records the data labels and interventions used by users; When AIGC content is generated, the user's nonce value is used to generate a watermark and embed it into the generated content; When the user needs copyright protection or has stopped generating, the user's intervention record is encrypted using the nonce value and uploaded to group A of the alliance chain. The user's nonce value and the data tag used are encrypted with the regulatory key and uploaded to group B of the alliance chain.

[0011] This also includes the tracing of digital watermarks, specifically: Extract the watermark from the generated AIGC content, and call the private key to decrypt the nonce value to determine whether the AIGC content has been tampered with.

[0012] This also includes data identification label monitoring, specifically synchronizing the on-chain content of Group B to the centralized database, and regularly performing integrity monitoring tasks based on the signature field content of the regulatory key. If it is determined that the data has been tampered with, a regulatory warning will be issued.

[0013] Among them, the content of group A uploaded to the alliance chain includes nonce value, basic information of the work, user intervention behavior records, AI platform name, user public key, and user basic information; among them, nonce value, basic information of the work, user intervention behavior records, user public key and user basic information must be encrypted by the public key before being uploaded to the chain.

[0014] Among them, the content of group B uploaded to the alliance chain includes the digital label identifier, nonce value, AI platform name, user public key, timestamp, encryption identifier, and the signature result of digitally encrypting the above content using the regulatory key used by the user to generate AIGC, and is uploaded to group B after desensitization.

[0015] Among them, the watermark generation of AIGC content includes the name of the AI ​​platform, alliance chain information, user identifier, encrypted nonce value, timestamp and digital fingerprint.

[0016] Compared with the prior art, the present invention has the following beneficial effects: 1. Use the pre-generated nonce value as the random number of the embedding algorithm and encrypt it into the watermark, without the need to embed the entire training data into the watermark, and without the need to upload information such as generated training data and user intervention records to the chain in real time, thus avoiding transaction congestion and transaction overhead.

[0017] 2. When users generate content frequently, the watermark embedding process does not need to include the link of data chaining, that is, chaining and watermark embedding are decoupled, but the identifier that can be associated with the specific content on the chain is placed in the watermark. While having the characteristics of blockchain, it greatly improves the efficiency and feasibility of blockchain watermark embedding.

[0018] 3. The nonce value in the watermark can quickly locate the block and extract the specific transaction content based on the nonce value. The transactions in the block contain all the training behaviors of the specified user for the relevant works.

[0019] 4. After the nonce value is decrypted, the watermark embedding algorithm can locate the positions of other watermarks in the work according to the nonce value, which can help extract more tamper-proof and complete watermark information.

[0020] 5. The data label identifiers in the training data set can be used for compliance monitoring and other purposes, but it is not convenient to save them together with the watermark on-chain and other content. Under the transaction submitted based on the same nonce value, the above data can be divided into two transactions and submitted to different sub-chain environments respectively. The nonce value can associate the two transactions. The content on the chain with the data label identifier does not need to repeat the information such as user intervention behavior on the chain.

[0021] 5. Because blockchain nonce has strong randomness, digital watermarks based on nonce values ​​can have stronger security and robustness, and as long as a correct nonce encryption string is extracted from the content, complete and reliable behavioral data can be traced back on the blockchain. DETAILED DESCRIPTION

[0022] The following will be described clearly and completely in conjunction with the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0023] It should be understood that when used in this specification and the appended claims, the terms "include" and "comprises" indicate the presence of described features, integers, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or combinations thereof.

[0024] It should also be understood that the terms used in this specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. As used in the specification of the present invention and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an" and "the" are intended to include plural forms.

[0025] It should be further understood that the term "and / or" used in the present description and the appended claims refers to any and all possible combinations of one or more of the associated listed items, and includes these combinations.

[0026] 1. Watermark embedding and data chain uploading process in AIGC process.

[0027] 1. The AI ​​platform joins the alliance chain; This solution requires the AI ​​platform to join the alliance chain as a participant. Example: On the basis of the alliance chain, a subchain environment is created, and the AIGC operator joins the subchain as a node. Multiple AIGC operators can join. Assume that all AI platforms in this solution join Group A (alliance chain subchain environment).

[0028] 2. When users use the AI ​​platform, the AI ​​model is trained to generate content based on user needs; Example: Users provide text requirements containing prompt words on the AI ​​image platform to obtain AIGC results. Users can optimize training or generate other AIGC content based on these results.

[0029] 3. Because the user has trained AIGC, the blockchain service will pre-generate a nonce value for the user (refer to the blockchain-related process for details); 4. Obtain training identification and user intervention behavior, cache or save to other locations; Details: Extract the data identifiers or data labels from the data set used in the AI ​​model training process. The intervention records clicked or entered by the user in the training behavior are recorded in the cache in the form of strings together with the data identifier labels. Since the data identifier labels and intervention behavior records are temporary data and need to be updated and read frequently, they do not need to be stored in a persistent storage medium. There is no restriction on the cache technology here. Whether it is Redis, Memcached or other cache technologies, they can provide high-performance data access capabilities and support setting data expiration time so that the data automatically expires after a certain period of time.

[0030] There is no restriction on the cached field key. It is only necessary to cache the user's data tag identifier and intervention behavior record. Here, we plan to use the cache tool redis. Redis generates a hash table for users who are using the AI ​​platform. Each user has a corresponding unique hash table item. In the hash table, create two fields (such as "data_tag" and "intervention_behavior"). When the data identification tag is extracted, the corresponding content is stored in the "data_tag" field. When the user generates an intervention behavior, the corresponding record is stored in the "intervention_behavior" field. When more data identification tags or intervention behavior records are obtained, use Redis's APPEND command to append the content to the value of the corresponding field.

[0031] 5. After the content is generated, watermark is embedded into the content, and the nonce value is used as the random number of the watermark algorithm; There is no restriction on the specific digital watermarking algorithm here. For different application requirements and multimedia types, each digital watermarking scheme has its advantages and limitations, which need to be evaluated and selected according to the specific situation. The choice of digital watermarking scheme is mostly determined by the operator of the AI ​​platform, but in order to improve the robustness and security of the watermark, the content producer will use some watermarking strategies, such as pseudo-random embedding, error correction code, multi-watermark technology, mixed domain watermark, frequency domain watermark (DWT, DCT) and other methods. Most of these methods require the use of random numbers. In practical applications, select appropriate random number generation algorithms and seeds to ensure that the generated random numbers have sufficient quality and randomness.

[0032] After AIGC content is generated, the watermark algorithm requests the blockchain service to obtain the nonce random number used by the current user, and the random number is calculated by the watermark algorithm and embedded in the production content. The embedded watermark is a text watermark, which contains the necessary information such as the name of the AI ​​platform, blockchain platform information, owner identifier (owner user ID, etc.), encrypted nonce random number, timestamp and digital fingerprint. The encrypted nonce random number is the ECC public key encryption of the nonce value random number required by the multimedia content using the platform's public key. The digital fingerprint is a hash operation on the watermark data content except the digital fingerprint information, which can be used to verify the integrity of the digital watermark.

[0033] 6. If the user needs copyright protection or stops generating, the nonce value will be used to upload all the user's training behaviors and identifications to the chain.

[0034] While users are continuously generating content, the Nonce value remains unchanged, and new data identification tags and user intervention records are appended to the cache field. When the user session times out, or when the user clicks on the "Copyright Protection" function on the operation platform, or other on-chain triggering mechanisms, the blockchain service will upload the current nonce value and all digital identification tags and user intervention behavior records in the cache since the nonce value was generated. (For details, refer to the blockchain-related process) 2. Blockchain-related processes The nonce value in the blockchain is a random number used to verify the uniqueness of transactions and prevent replay attacks. Here, we take the FISCO BCOS consortium chain as an example. When a node wants to execute a transaction, it needs to first obtain the nonce value of the transaction and attach it to the transaction. When the node executes the transaction and broadcasts it to other nodes, other nodes will verify whether the nonce value of the transaction is consistent with the one they have saved to ensure that the transaction is unique and prevent replay attacks.

[0035] 1. Pre-generation of transaction nonce value The generation of nonce values ​​in blockchain services uses a non-deterministic random number generator (N-DRNG) to ensure that each nonce value is unique in the entire network. The nonce value generated by N-DRNG is not simply incremented, but each generated nonce value is random and unpredictable. The simply incremented nonce value is not suitable for this scenario, and attackers can easily infer the nonce values ​​used by other watermarks based on the obtained nonce value.

[0036] 2. Pregenerate nonce value and embed it into the transaction and submit the transaction a Before the transaction request is sent, the pre-generated nonce value is attached to the transaction request. When the node receives the transaction request, it will use the nonce value in the request body to add it to the transaction content. After verification, the request content will use the same nonce value to embed two transactions (a, b) and submit them to different groups. The content of transaction a is the basic information of the work, the record of user intervention behavior, the name of the AI ​​platform, the user's public key, and the basic information of the user. Except for the name of the AI ​​platform, the other content is encrypted by the platform public key and uploaded to group A (alliance chain sub-chain environment) 3. Pregenerate nonce value and embed it into the transaction and submit transaction b The content of transaction b is all the digital identification tags, nonce values, AI platform names, user public keys, timestamps, whether the identification is encrypted, and the results of digitally signing the above content using the regulatory key in the cache since the nonce value of the user was generated. After being desensitized, it is uploaded to group B (alliance chain sub-chain environment). Since for data providers or AI model operators, the data labels in the data set themselves are usually not private content. Data labels are descriptions or annotations of data, usually to help understand and classify data. Data providers and AI model operators usually access and use label information in the data set for model training, evaluation and improvement. These labels are usually public and generally do not contain specific personal identities or sensitive information. The information uploaded to the chain will also be desensitized once. The nonce value of the on-chain content of transaction b will not be associated with the user's work information. There is no user public key information in the work watermark. In the case that the platform user's privacy information is not leaked, attackers cannot associate the work information or the nonce value stored on the chain with each other. At the same time, in order to facilitate the fuzzy matching of digital identification tags, this transaction b will be uploaded to the chain in plain text.

[0037] 4. Two transactions with the same nonce value on the chain of groups A and B The main content uploaded to the chain by Group A is the work information that needs to be uploaded to the chain by the AI ​​platform with nodes. In the application, one or more AI platforms can access the sub-chain environment to upload data to the chain. Since the content uploaded to the chain can be queried by all participating nodes, encryption is used to avoid plain text storage to ensure the privacy and security of the platform data.

[0038] The main content of group B on the chain is the digital identification tag, which is mainly used for trusted monitoring of digital identification. The regulatory node can access the group environment. The sensitive content in the digital identification tag will be encrypted and stored on the chain by a pair of public and private keys for sensitive data encryption and decryption. The key is kept and used by the regulatory node. When sensitive content appears, the desensitization program alerts the regulatory node and encrypts the digital identification tag and puts it on the chain together with other transaction content. The regulatory node can filter out sensitive content in batches by "whether to encrypt the identification" and determine the platform party and behavioral user that generated the sensitive content. The nonce value can help quickly locate the specific transaction content on the chain.

[0039] 3. Digital Watermark Tracing Process In the case of copyright disputes or other scenarios where watermark verification is required for generated content, after the watermarked content is extracted (the method of extracting the watermark is not limited here), if the platform information in the content is a node participating organization in the alliance chain, the digital watermark can be traced in the services provided by the alliance chain. The tracing process is as follows: 1. Analyze the generated content to extract the watermark The watermarking methods used by different media content and AI platforms are not consistent. Watermark extraction can be achieved through analysis of the original embedding algorithm or other methods.

[0040] 2. Based on the information from the platform, call the platform's private key to decrypt the encrypted content and obtain the nonce value After extracting the watermark, hash the contents other than the digital fingerprint. If the hash result does not match the digital fingerprint, the watermark content has been tampered with. However, whether the match succeeds or fails, it is impossible to truly determine whether the watermark content has been modified. The attacker can easily generate a new digital fingerprint to replace the original digital fingerprint.

[0041] Next, you need to apply for a private key to decrypt the nonce value in the watermark information from the AI ​​platform. If the decryption fails, the data will be tampered with. If multiple results are generated during watermark extraction, apply to the corresponding AI platform for decryption of the nonce value. After the decryption is successful, use the nonce to quickly locate the specific transaction, obtain the platform name information in the transaction content, and apply to the platform for a private key to decrypt other encrypted content in the transaction.

[0042] 3. Re-judgment of watermark integrity There are many ways for attackers to tamper with watermarks, so the information in the watermark, such as the platform name, owner identifier, decrypted nonce value and timestamp, needs to be completely consistent with the content on the chain, otherwise it will be judged that the watermark information has been tampered with.

[0043] 4. Data Identification Label Monitoring Process In response to AIGC content fraud and the use of copyright data, processing and statistics of data identification labels in the training data set can enable more intuitive compliance monitoring of AIGC, while decentralized records will ensure that the records of data identification labels are more transparent and reliable, and the data sharing mechanism of blockchain can also improve regulatory efficiency. The chain content of Group B (alliance chain sub-chain environment) is related to data label identification, and the chain content will be synchronized to the centralized database using block data export technology, and there is no restriction on block data export technology here. After exporting to the centralized database, it is necessary to conduct complete detection tasks on peers regularly in batches according to the signature field content of the regulatory key. The data identification label monitoring process is divided into two aspects: one is data integrity monitoring and monitoring early warning. The second is identification usage query and monitoring early warning.

[0044] Data integrity monitoring and early warning 1. Regularly batch-check the public key signature of the field content. If the signature fails, it is determined that the data has been tampered with. In the previous section "Pre-generate nonce value and embed it in the transaction and submit transaction b", it is mentioned that data identification tags and other content are uploaded to group B through transaction b, which includes the signature encryption of the data by the regulatory key. Digital signature technology is a security authentication technology based on asymmetric encryption technology and digital summary technology. It involves the use of ECC public and private keys. The sender uses the private key to encrypt the data, and the receiver uses the public key to decrypt it. Digital signature technology can ensure the integrity and authenticity of the data, and prevent the data from being tampered with or forged. Here, the public key of the regulator can be used to verify the signature field content.

[0045] 2. After the signature verification is successful, the hash summary is compared with the peer data; Digital signature verification can confirm whether the encrypted signature has been tampered with and whether it comes from the private key signature of the regulator. After decryption, a hash summary will be obtained. The hash summary of other contents in the same line can be compared with the decrypted summary. If the summary matches, the data is complete. If the summary fails to match, it is determined that the data has been tampered with.

[0046] 3. If there is any suspicion of tampering, an early warning will be issued and the on-chain data will be overwritten If it is determined that the data has been tampered with, an early warning will be issued to the regulator. The warning content includes but is not limited to the tampered data, data on the blockchain, etc. The centralized data log is checked, and the transaction content in the range is extracted based on the nonce value of the complete data before and after the tampered data, and the centralized database content is synchronously replaced.

[0047] Advantages and innovations of the method used in the present invention: 1. Training behavior on-chain The decentralized records of user intervention behaviors and data identification labels are more credible, and the transparent AIGC training behaviors and data identification can also provide credible judicial support for “first-come, first-served” copyrights.

[0048] 2. Data privacy protection a. When the watermark data of the work is uploaded to the chain, the private data therein will be encrypted. The private data uploaded to the chain by different AI platforms will be encrypted using the public keys of different AI platforms. Only the private key of the AI ​​platform can decrypt it, avoiding plaintext storage and ensuring the privacy of the uploaded data.

[0049] b. The nonce value in the watermark is also encrypted with the public key of the AI ​​platform. It is very difficult for an attacker to crack the nonce value in the watermark without the private key of the AI ​​platform. By decrypting the encrypted nonce, it is possible to quickly determine whether the content of the watermark embedding location has been tampered with.

[0050] c. The identification tags of data are very likely to contain private information. This solution will perform another desensitization process on the identification tags before uploading them to the chain. In order to ensure the integrity of the entire chain of behavioral traces, sensitive data will be uploaded to the chain after being encrypted with the public key. The private key of the non-regulatory party cannot decrypt it.

[0051] 3. Improve regulatory efficiency a. If the copyright or fraud of AIGC content needs to be verified, the nonce value in the watermark can be quickly traced back to the credible content of on-chain transaction a (information related to copyright) and transaction b (information related to digital identification tags); b. Once sensitive or non-compliant content appears in the data identification tag, it can directly warn the regulatory node, and the regulatory node can quickly locate the related users, platforms, work information, etc. based on the on-chain data in the shared environment.

[0052] Transaction process innovation based on pre-generated nonce values: 1. Use the pre-generated nonce value as the random number of the embedding algorithm and encrypt it into the watermark, without the need to embed the entire training data into the watermark, and without the need to upload information such as generated training data and user intervention records to the chain in real time, thus avoiding transaction congestion and transaction overhead.

[0053] 2. When users generate content frequently, the watermark embedding process does not need to include the link of data chaining, that is, chaining and watermark embedding are decoupled, but the identifier that can be associated with the specific content on the chain is placed in the watermark. While having the characteristics of blockchain, it greatly improves the efficiency and feasibility of blockchain watermark embedding.

[0054] 3. The nonce value in the watermark can quickly locate the block and extract the specific transaction content based on the nonce value. The transactions in the block contain all the training behaviors of the specified user for the relevant works.

[0055] 4. After the nonce value is decrypted, the watermark embedding algorithm can locate the positions of other watermarks in the work according to the nonce value, which can help extract more tamper-proof and complete watermark information.

[0056] 5. The data label identifiers in the training data set can be used for compliance monitoring and other purposes, but it is not convenient to save them together with the watermark on-chain and other content. Under the transaction submitted based on the same nonce value, the above data can be divided into two transactions and submitted to different sub-chain environments respectively. The nonce value can associate the two transactions. The content on the chain with the data label identifier does not need to repeat the information such as user intervention behavior on the chain.

[0057] 5. Because blockchain nonce has strong randomness, digital watermarks based on nonce values ​​can have stronger security and robustness, and as long as a correct nonce encryption string is extracted from the content, complete and reliable behavioral data can be traced back on the blockchain.

Claims

1. A method for generating content multimedia digital watermarks based on blockchain artificial intelligence, characterized by: Including alliance chain, The consortium chain includes Group A and Group B, where Group A is the consortium chain sub-chain related to AIGC work information, and Group B is the consortium sub-chain related to digital identification tags; The consortium chain pregenerates nonce values ​​for users; Users use the AI ​​platform linked to the consortium chain to train and generate content, and the AI ​​platform records the data labels and interventions used by users; When AIGC content is generated, the user's nonce value is used to generate a watermark and embed it into the generated content; When the user needs copyright protection or has stopped generating, the user's intervention record is encrypted using the nonce value and uploaded to group A of the alliance chain. The user's nonce value and the data tag used are encrypted with the regulatory key and uploaded to group B of the alliance chain.

2. According to claim 1, a method for generating content multimedia digital watermarks based on blockchain artificial intelligence is characterized by: It also includes the traceability of digital watermarks, specifically: Extract the watermark from the generated AIGC content, and call the private key to decrypt the nonce value to determine whether the AIGC content has been tampered with.

3. According to claim 1, a method for generating content multimedia digital watermarks based on blockchain artificial intelligence is characterized by: It also includes data identification tag monitoring, specifically synchronizing the on-chain content of group B to the centralized database, and regularly performing integrity monitoring tasks based on the signature field content of the regulatory key. If it is determined that the data has been tampered with, a regulatory warning will be issued.

4. According to claim 1, a method for generating content multimedia digital watermarks based on blockchain artificial intelligence is characterized by: The content of Group A uploaded to the alliance chain includes nonce value, basic information of the work, user intervention behavior records, AI platform name, user public key, and user basic information; among them, nonce value, basic information of the work, user intervention behavior records, user public key and user basic information must be encrypted by the public key before being uploaded to the chain.

5. According to claim 1, a method for generating content multimedia digital watermarks based on blockchain artificial intelligence, characterized in that: The content of Group B uploaded to the alliance chain includes the digital label used by users to generate AIGC, nonce value, AI platform name, user public key, timestamp, encryption flag, and the signature result of digitally encrypting the above content using the regulatory key, which is uploaded to Group B after being desensitized.

6. According to claim 1, a method for generating content multimedia digital watermarks based on blockchain artificial intelligence, characterized in that: The watermark generation of AIGC content includes the AI ​​platform name, alliance chain information, user identifier, encrypted nonce value, timestamp and digital fingerprint.