Creative Chinese-based watermark intelligent processing method
By adopting an intelligent processing method based on creative Chinese watermarks in the data transaction solution, the problem of difficult traceability after data leakage is solved, effective protection and traceability of data is achieved, and data value is prevented from shrinking and abuse.
Patent Information
- Application Number
- CN202411246978.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-06
- Publication Date
- 2025-01-21
AI Technical Summary
The existing data trading solutions lack effective protection and traceability mechanisms for data sets. Once the data is leaked, the data may be traded in a secondary manner, resulting in a significant reduction in the value of the data and the abuse of data, infringing on the interests of the data owners.
A method based on intelligent processing of creative Chinese watermarks is used to generate watermark fingerprints from the input watermark data through the watermark extraction process and store them in the database. The posters containing creative Chinese are corrected for spelling. China Telecom Xingchen artificial intelligence large model and log analysis are used. The layer mining method analyzes and trains watermark data to reduce watermark effect errors.
Effectively protect and trace data sets to avoid secondary transactions of data, protect the value and security of data, prevent data abuse, and ensure that the interests of data owners are not infringed.
Smart Images
Figure CN119338657A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the technical field of anti-counterfeiting processing, in particular to a method based on creative Chinese watermark intelligent processing. Background Art
[0002] Anti-counterfeiting has become a new and constantly evolving field. Anti-counterfeiting technology can generally be divided into physical anti-counterfeiting technology, chemical ink anti-counterfeiting technology, biological anti-counterfeiting technology, material anti-counterfeiting technology, computer network anti-counterfeiting technology, anti-counterfeiting printing technology, anti-counterfeiting packaging technology, etc. The images, texts and other information we usually come into contact with must eventually be output through printing anti-counterfeiting, such as trademarks, product packaging, books, certificates, documents, stamps, magnetic cards, publications, currency, valuable tickets, documents, passports and other certificates and documents, bills and tickets, thus highlighting the importance of anti-counterfeiting printing technology. Anti-counterfeiting printing technology is the most commonly used, easiest to identify and lowest cost anti-counterfeiting technology in my country. Summary of the invention
[0003] With the rapid development of computer science and network technology, data has gradually become the mainstay of production and life in the new era, and the value of data has also been reflected. However, what follows is the issue of data security. How to ensure the security and availability of data and maximize the value of data assets has become a problem that must be solved.
[0004] In the process of data circulation, it is very important to trace the data after it is illegally leaked. The existing data trading solutions lack effective protection and traceability mechanisms for data sets. Once the data is leaked, the data may be traded again, resulting in a significant reduction in the value of the data and data abuse, infringing on the interests of the data owner. Summary of the invention
[0005] A method based on creative Chinese watermark intelligent processing, comprising the following steps: Step 1: First, through watermark extraction, the input watermark data is converted into a watermark fingerprint and stored in the database; secondly, for posters containing creative Chinese, the application scenario is analyzed first, and then the spelling correction DE-CO model is used to correct the hidden misspelled words in creative Chinese; Step 2: Then, sort the multiple watermark training results of the AI posters containing watermarks in order, and use the mean square error (MSE) and root mean square error (RMSE) to calculate the watermark generation error of the AI poster watermark data set; Step three, then, use the deployed China Telecom Xingchen artificial intelligence large model combined with the poster watermark data parameters to analyze and obtain the watermark data and related data sets, and then use the log layered mining method to obtain each creative Chinese and related relationships in the AI poster as dynamic parameters to participate in the Xingchen small model training to obtain each creative Chinese data set, and finally, multiple data sets are de-recombined to complete the pre-processing of the creative Chinese in the AI poster watermark data set and reduce the watermark effect error.
[0006] The step 1 is divided into the following steps: S1. Through watermark extraction, the input watermark data is converted into a watermark fingerprint and stored in the database. The watermark fingerprint is the hash value of a file or data, also known as a hash value. A hash function is a function that can convert data of any size into data of a fixed size. It is usually used to generate a short, unique "fingerprint" string. In Python, the built-in hashlib module is used to calculate the hash value of a file. The watermark data is processed to obtain a data set for model training, and database images corresponding to multiple application scenario definitions are searched in the database. The database images are pre-processed by cleaning, cropping, size standardization, format conversion, etc. S2. For posters containing creative Chinese, we first analyze the application scenarios, and then use the spelling correction DE-CO model to correct the hidden misspelled words in creative Chinese. The DE-CO model consists of a misspelled word recognizer and a misspelled word corrector. The data set to be checked is put into the misspelled word detector for detection. The function of the misspelled word recognizer is to identify the misspelled words in the adversarial sample and provide these misspelled words to the misspelled word corrector for correction, while ensuring that the correctly spelled words in the adversarial sample are not modified. The function of the misspelled word corrector is to correct the misspelled words based on the spelling information and context information of the misspelled words.
[0007] The size of the MSE value and the mean square error refer to the expected value of the square of the difference between the estimated value and the true value, which is proportional to the error of the experimental data measured by the model. The smaller the MSE value, the higher the accuracy. Where N is the dimension of the sample space: The estimated position coordinates of the target to be measured in the N-dimensional sample space; H is the actual position coordinates of the target to be measured in the N-dimensional sample space; The root mean square error is used to evaluate the position estimation accuracy of the AI poster area division label. Assuming that there are N labels for area division, the true position of the positioning label i is , the estimated position of the positioning tag obtained by the positioning algorithm is , then the root mean square error of the estimated position of the positioning tag is (Formula 3) If there are abnormal data with large errors in the estimated position coordinates of the tag, remove the abnormal data and then use the root mean square error to evaluate the performance of the positioning algorithm. Otherwise, the root mean square error value will be very large, and the evaluation of the positioning algorithm performance will be inaccurate.
[0008] Technical innovations of the present invention: 1. First, through the watermark extraction process, the input watermark data is intelligently processed to generate a watermark fingerprint and store it in the database. Secondly, for posters containing creative Chinese, the application scenario is first analyzed, and then a two-step spelling correction (Detector-Corrector, DE-CO) model is used to correct the possible hidden misspelled words in creative Chinese. Technical innovation 2. Then, the deployed China Telecom Xingchen artificial intelligence large model is combined with the poster watermark data parameters for analysis to obtain watermark data and related data sets. Then, according to the log hierarchical mining method, each creative Chinese and the associated relationship in the AI poster are obtained as dynamic parameters to participate in the Xingchen small model training to obtain each creative Chinese data set. Finally, multiple data sets are de-recombined to obtain, thereby completing the preprocessing of the creative Chinese in the AI poster watermark data set and reducing the watermark effect error.
[0009] Existing data trading solutions lack effective protection and traceability mechanisms for data sets. Once data is leaked, it may be traded again, resulting in a significant reduction in the value of the data and data abuse, infringing on the interests of the data owner.
[0010] For posters containing creative Chinese, the watermark extraction process is used to optimize the AI posters. Secondly, for posters containing creative Chinese, the application scenarios are first analyzed, and then the two-step spelling correction (Detector-Corrector, DE-CO) model is used to correct the possible hidden misspelled words of creative Chinese. This effectively protects and traces the data set. Finally, in order to avoid the optimal solution being a local optimal solution rather than a global optimal solution, the mean square error (MSE) and root mean square error (RMSE) are used to calculate the watermark generation error of the AI poster watermark data set. BRIEF DESCRIPTION OF THE DRAWINGS
[0011] Figure 1 This is a schematic diagram of the data watermark extraction process; Figure 2 Schematic diagram of the two-step spelling correction model; Figure 3 Schematic representation of the poster's training effects. DETAILED DESCRIPTION
[0012] A method based on creative Chinese watermark intelligent processing, comprising the following steps: Step 1: First, through watermark extraction, the input watermark data is converted into a watermark fingerprint and stored in the database; secondly, for posters containing creative Chinese, the application scenario is analyzed first, and then the spelling correction DE-CO model is used to correct the hidden misspelled words in creative Chinese; Step 2: Then, sort the multiple watermark training results of the AI posters containing watermarks in order, and use the mean square error (MSE) and root mean square error (RMSE) to calculate the watermark generation error of the AI poster watermark data set; Step three, then, use the deployed China Telecom Xingchen artificial intelligence large model combined with the poster watermark data parameters to analyze and obtain the watermark data and related data sets, and then use the log layered mining method to obtain each creative Chinese and related relationships in the AI poster as dynamic parameters to participate in the Xingchen small model training to obtain each creative Chinese data set, and finally, multiple data sets are de-recombined to complete the pre-processing of the creative Chinese in the AI poster watermark data set and reduce the watermark effect error.
[0013] The step 1 is divided into the following steps: S1. Through watermark extraction, the input watermark data is converted into a watermark fingerprint and stored in the database. The watermark fingerprint is the hash value of a file or data, also known as a hash value. A hash function is a function that can convert data of any size into data of a fixed size. It is usually used to generate a short, unique "fingerprint" string. In Python, the built-in hashlib module is used to calculate the hash value of a file. The watermark data is processed to obtain a data set for model training, and database images corresponding to multiple application scenario definitions are searched in the database. The database images are pre-processed by cleaning, cropping, size standardization, format conversion, etc. S2. For posters containing creative Chinese, we first analyze the application scenarios, and then use the spelling correction DE-CO model to correct the hidden misspelled words in creative Chinese. The DE-CO model consists of a misspelled word recognizer and a misspelled word corrector. The data set to be checked is put into the misspelled word detector for detection. The function of the misspelled word recognizer is to identify the misspelled words in the adversarial sample and provide these misspelled words to the misspelled word corrector for correction, while ensuring that the correctly spelled words in the adversarial sample are not modified. The function of the misspelled word corrector is to correct the misspelled words based on the spelling information and context information of the misspelled words.
[0014] The size of the MSE value and the mean square error refer to the expected value of the square of the difference between the estimated value and the true value, which is proportional to the error of the experimental data measured by the model. The smaller the MSE value, the higher the accuracy. Where N is the dimension of the sample space: The estimated position coordinates of the target to be measured in the N-dimensional sample space; H is the actual position coordinates of the target to be measured in the N-dimensional sample space; The root mean square error is used to evaluate the position estimation accuracy of the AI poster area division label. Assuming that there are N labels for area division, the true position of the positioning label i is , the estimated position of the positioning tag obtained by the positioning algorithm is , then the root mean square error of the estimated position of the positioning tag is (Formula 3) If there are abnormal data with large errors in the estimated position coordinates of the tag, remove the abnormal data and then use the root mean square error to evaluate the performance of the positioning algorithm. Otherwise, the root mean square error value will be very large, and the evaluation of the positioning algorithm performance will be inaccurate.
[0015] The specific implementation cases are as follows: Step 1: First, through the watermark extraction process, the input watermark data is intelligently processed to generate a watermark fingerprint and store it in the database. Secondly, for posters containing creative Chinese, the application scenario is first analyzed, and then the two-step spelling correction (Detector-Corrector, DE-CO) model is used to correct the possible hidden misspelled words in creative Chinese.
[0016] Existing technology: There are two main ways to watermark AI-generated content: The first is to add it during the content creation process, which requires access to the model itself, but is more robust because it is part of the generation process. The second method is to apply it after the content is generated, which can be used for content generated by closed-source and proprietary models, but may not be suitable for all types of content (such as text).
[0017] Data poisoning works by subtly altering images shared online to prevent AI algorithms from processing them correctly. While humans can view these images normally, AI algorithms do not have access to similar content and are therefore unable to create new images. Such techniques include Glaze and Photoguard. Other tools “poison” images to undermine the inherent assumptions in the training of AI algorithms, making it impossible for AI systems to learn what people look like based on images shared online - making it more difficult for these systems to generate fake images of people. Such tools include Nightshade and Fawkes. The authenticity and reliability of content can also be maintained through the use of “signature” techniques, which link content to metadata about its source, such as the work of Truepic, which embeds metadata that follows the C2PA standard. Image signatures help understand the source of an image.
[0018] Generating fingerprints usually refers to the hash value of a file or data, also known as a hash value. A hash function is a function that can convert data of any size into data of a fixed size, and is usually used to generate a short, unique "fingerprint" string. In Python, you can use the built-in hashlib module to calculate the hash value of a file.
[0019] import hashlib def compute_file_hash(file_path, hash_algorithm='sha256'): hash_func = getattr(hashlib, hash_algorithm) sha256 = hash_func() with open(file_path, 'rb') as file: while True: data = file.read(65536) if not data: break sha256.update(data) return sha256.hexdigest() # Use the function to calculate the hash value of the file file_path = 'example.txt' print(compute_file_hash(file_path)) Intelligent processing of watermark data: 1. Obtain the data set for model training and search the database for database images corresponding to multiple scene definitions; 2. Preprocessing the multiple database images by cleaning, cropping, size standardization, format conversion, etc.; 3. The two-step spelling correction (Detector-Corrector, DE-CO) model is used to correct possible hidden misspelled words in creative Chinese, which helps to improve the performance and generation effect of the generative model.
[0020] 4. Based on the subject features, a generative model is used to generate multiple flat creative pictures containing the subject.
[0021] S2. Secondly, for posters containing creative Chinese, after confirming the application scenario to which they belong, a two-step spelling correction (Detector-Corrector, DE-CO) model is used for the content of the creative Chinese related data set to correct possible hidden misspelled words in creative Chinese.
[0022] The two-step spelling correction (Detector-Corrector, DE-CO) model structure is as follows Figure 2 The DE-CO model consists of a misspelled word recognizer and a misspelled word corrector. The data set to be checked is placed in the misspelled word detector for detection. The function of the misspelled word recognizer is to identify the misspelled words in the adversarial sample and provide these misspelled words to the misspelled word corrector for correction, while ensuring that the correct spelled words in the adversarial sample are not modified. The function of the misspelled word corrector is to correct the misspelled words based on the spelling information and context information of the misspelled words. The workflow of the DE-CO model and the defense against adversarial samples through the DE-CO model are shown below.
[0023] The deployed China Telecom Xingchen artificial intelligence large model is used to analyze the poster watermark data parameters to obtain watermark data and related data sets. Then, the log hierarchical mining method is used to obtain each creative Chinese character and its associated relationship in the AI poster as a dynamic parameter to participate in the Xingchen small model training to obtain the data set of each creative Chinese character. Finally, multiple data sets are de-recombined to complete the pre-processing of the creative Chinese characters in the AI poster watermark data set and reduce the error of the watermark effect.
[0024] S1. China Telecom Xingchen artificial intelligence model's first operation: In the training process of large models, unsupervised learning methods are usually used. This means that the model does not need manually labeled data during training, but directly learns language rules and knowledge from the original text. Since the model training results depend on dynamic parameters such as input conditions, when the input parameters and conditions are inaccurate, mainstream prediction models are often required for prediction.
[0025] First, four dynamic parameters are input: the current traffic light IP, vehicle or drone IP, destination IP, and congestion trend probability.
[0026] Secondly, since the big model itself does not require manually labeled data, the Xingchen big model uses its own vector database to obtain the relationship between the current intersection and other intersections and mark the data.
[0027] The second calculation of China Telecom Xingchen AI model: The large model operates to obtain the labeled data set, and the newly added policy keyword parameters are put into the small model again for secondary training within the scope of the large data operation set and the policy keywords are stored in the vector database to improve the training accuracy.
[0028] S2. Log hierarchical mining method.
[0029] Specific description of log layered mining method: The first layer is to build the association between creative Chinese characters in AI posters, which can be one-to-many or many-to-many. The number of occurrences of creative Chinese characters in the log data and the business weights in the AI posters are analyzed and marked, and then other creative Chinese characters associated with the mark are extracted from the log to form a data set.
[0030] The second layer builds records of the co-occurrence between other creative Chinese data and other AI posters. These data record the correlation between AI posters, between creative Chinese, and between AI posters and creative Chinese in different time dimensions.
[0031] Embodiment Example 1 The specific embodiments of the invention mainly include the following specific contents (including description and code examples, using Java language as an example): Data collection and processing: The data collection and processing module collects historical reference data of the same type as the poster, cleans the collected data and extracts features to implement data preprocessing logic, such as normalization and feature extraction. The data processing example is as follows: @component public class DataAcquisitionAndProcessor { public List <processeddata>preprocessData(List <datapoint>rawData){ / / Implement data preprocessing logic, such as normalization, feature extraction, etc. ReturnrawData.stream() .map(dataPoint->new ProcessedData(dataPoint.getName(), normalize(dataPoint.getValue()))) .collect(collectors.tolist()); } private double normalize(int value){ / / Simple normalization example return value / 100.0; } { Model training based on deep learning: Perform model training based on the feature information data processed in step 1 to generate a business orchestration recommendation model. At the same time, deploy the model so that it can provide computing power to the outside world. The following is an example of deep training model processing: @component public class DeepLearningModel{ private final RestTemplate restTemplate = new RestTemplate(); private final String modelEndpoint = "http: / / deep-learning-service / predict"; public PredictionResult predict(List <processeddata>inputData){ HttpHeaders headers=new HttpHeaders(); headers.setcontentType(MediaType.APPLICATION JSON); HttpEntity<List <processeddata>>requestEntity = new HttpEntity<>(inputData, headers); ResponseEntity <predictionresult>response=restTemplate.exchange(modelEndpoint, HttpMethod.PosT, requestEntity, PredictionResult.class); return response.getBody(); } { Deep learning model integrated with LiteFlow framework: Build a project based on SpringBoot, complete the introduction of LiteFlow component dependencies, and improve the configuration file content; the project configuration and LiteFlow component configuration are as follows: #Configure this item in the SpringBoot configuration file liteflow.rule-source=classpath:flow.el # LiteFlow process-specific configuration chain: THEN (optimizeProcess); Define LiteFlow process chain and process rules: Define LiteFlow components to process business logic and deep learning model integration, and use this component to call the model's automatic reasoning and calculation capabilities to generate watermarks for the AI poster watermark data set of mean square error (MSE) and root mean square error (RMSE). The following is an example of calculating the positioning error: @Component("optimizeProcess") public class OptimizeProcessComponent extends NodeComponent { @Autowired private DataAcquisitionAndProcessor dataAcquisitionAndProcessor; @Autowired private DeepLearningModel deepLearningModel; @override public void process()throws Exception{ / / Data preprocessing List <processeddata>processedDatamdata=AcquisitionAndProcessor.preprocessData(rawData); / / Call the deep learning model for prediction PredictionResult prediction =deepLearningModel.predict(processedpata); / / Adjust business processes based on prediction results if(prediction.requiresoptimization()) {adjustWorkflow(prediction); } / / Store the prediction results in the context for use by subsequent components this.getslot().setRequestData("predictionResult", prediction); } private void adjustWorkflow(PredictionResult prediction){ / / Implement business process optimization logic, such as adjusting task priorities, reallocating resources, etc. } } Create a controller to trigger the process: provide an API interface for other services to call. The example is as follows @Restcontroller public class Processcontroller{1 @Autowired Private FlowExecutor flowExecutor; @GetMapping(" / optimizeProcess") public string optimizeProcess(){ flowExecutor.execute2Resp("chain","arg"); return "Process optimized"; } } Complete watermark processing.
[0032] Glossary Intelligent data watermarking is a technology that hides watermark information into structured data through certain rules and algorithms. The value of the data remains almost unchanged after hiding. It is mainly used for copyright protection and tracking and tracing after leakage. This technology is used to embed the watermark information of the right holder in the data to realize the ownership declaration, and after the data is leaked, the watermark in the leaked data file can be extracted to hold the right holder accountable and protect the security of the data.
[0033] Graphic creative images usually refer to creative and artistic images designed and produced on a two-dimensional plane. Such images are usually used in design fields such as advertising, promotion, illustrations, posters, and business cards.
[0034] Dataset: A dataset is a collection of organized data, which is usually used in machine learning, deep learning, data analysis and other fields. A dataset can contain various types of data, such as text, images, videos, audio, etc., as well as the labels or annotation information corresponding to these data.
[0035] Corporate IP refers to an image identity that a company or organization creates in the market through a unique image and brand style to differentiate itself from other competitors and gain recognition and trust from the target audience. It is usually composed of a series of tangible and intangible elements, including brand name, logo, products, services, and brand story.
[0036] China Telecom Artificial Intelligence Technology Co., Ltd. and China Telecom (601728) Artificial Intelligence Research Institute (TeleAI) released a product called Xingchen Big Model Software Factory (hereinafter referred to as Xingchen Software Factory), and said that this is the first breakthrough AI development tool in China.< / processeddata> < / predictionresult> < / processeddata> < / processeddata> < / datapoint> < / processeddata>
Claims
1. A method based on creative Chinese watermark intelligent processing, characterized in that The steps include: Step 1: First, through watermark extraction, the input watermark data is converted into a watermark fingerprint and stored in the database; secondly, for posters containing creative Chinese, the application scenario is analyzed first, and then the spelling correction DE-CO model is used to correct the hidden misspelled words in creative Chinese; Step 2: Then, sort the multiple watermark training results of the AI posters containing watermarks in order, and use the mean square error (MSE) and root mean square error (RMSE) to calculate the watermark generation error of the AI poster watermark data set; Step three, then, use the deployed China Telecom Xingchen artificial intelligence large model combined with the poster watermark data parameters to analyze and obtain the watermark data and related data sets, and then use the log layered mining method to obtain each creative Chinese and related relationships in the AI poster as dynamic parameters to participate in the Xingchen small model training to obtain each creative Chinese data set, and finally, multiple data sets are de-recombined to complete the pre-processing of the creative Chinese in the AI poster watermark data set and reduce the watermark effect error.
2. According to claim 1, a method for intelligent processing of creative Chinese watermarks is characterized in that The step 1 is divided into the following steps: S1. Through watermark extraction, the input watermark data is converted into a watermark fingerprint and stored in the database. The watermark fingerprint is the hash value of a file or data, also known as a hash value. A hash function is a function that can convert data of any size into data of a fixed size. It is usually used to generate a short, unique "fingerprint" string. In Python, the built-in hashlib module is used to calculate the hash value of a file. The watermark data is processed to obtain a data set for model training, and database images corresponding to multiple application scenario definitions are searched in the database. The database images are pre-processed by cleaning, cropping, size standardization, format conversion, etc. S2. For posters containing creative Chinese, we first analyze the application scenario to which they belong, and then use the spelling correction DE-CO model to correct the hidden misspelled words in creative Chinese. The DE-CO model consists of a misspelled word recognizer and a misspelled word corrector. The data set to be checked is put into the misspelled word detector for detection. The function of the misspelled word recognizer is to identify the misspelled words in the adversarial sample and provide these misspelled words to the misspelled word corrector for correction, while ensuring that the correctly spelled words in the adversarial sample are not modified. The function of the misspelled word corrector is to correct the misspelled words according to the spelling information and context information of the misspelled words.
3. According to claim 1, a method for intelligent processing of creative Chinese watermarks is characterized in that Said The size of the MSE value is proportional to the mean square error, which refers to the expected value of the square of the difference between the estimated value and the true value. The error of the experimental data measured by the model is proportional. The smaller the MSE value, the higher the accuracy. Where N is the dimension of the sample space: The estimated position coordinates of the target to be measured in the N-dimensional sample space; H is the actual position coordinates of the target to be measured in the N-dimensional sample space; The root mean square error is used to evaluate the position estimation accuracy of the AI poster area division label. Assuming that there are N labels for area division, the true position of the positioning label i is , the estimated position of the positioning tag obtained by the positioning algorithm is , then the root mean square error of the estimated position of the positioning tag is ， If there are abnormal data with large errors in the estimated position coordinates of the tag, remove the abnormal data and then use the root mean square error to evaluate the performance of the positioning algorithm. Otherwise, the root mean square error value will be very large, and the evaluation of the positioning algorithm performance will be inaccurate.