A robust watermarking method for preserving statistical properties for industrial video databases

Abstract

This invention relates to a robust watermarking method for industrial video databases that preserves statistical properties, belonging to the fields of digital watermarking, industrial video data security, and industrial big data processing. Addressing the shortcomings of existing statistical property-preserving database watermarking schemes, such as inability to adapt to multi-attribute columns, low sampling efficiency, and lack of parameter adaptability, this invention proposes a robust watermarking method that supports multi-attribute numerical databases, possesses adaptive sampling capabilities, and maintains unchanged statistical properties. Specifically, firstly, a covariance matrix is ​​constructed and solved based on the invariance of statistical properties (mean, variance, covariance). Then, the potential watermarks of the tuple data involved in embedding in the original database are extracted. Next, the potential watermarks are compared with the target watermark of their respective groups. For inconsistent non-good data, a perturbation is constructed using a residual column framework until the data becomes good. Finally, a majority voting mechanism is used to extract the watermark.

Description

Technical Field

[0001] This invention relates to a robust watermarking method for preserving statistical properties of industrial video databases, belonging to the fields of digital watermarking, industrial video data security, and industrial big data processing technology. Background Technology

[0002] With the rapid iteration of industrial intelligent manufacturing, industrial internet, and intelligent safety monitoring technologies, industrial scenarios have fully entered the stage of visualized, digitalized, and intelligent supervision. In various industrial production workshops, assembly lines, intelligent warehouses, equipment operation and maintenance, and safety production scenarios, a large amount of high-definition surveillance video, equipment inspection video, and production inspection video data are continuously collected, stored, and analyzed, forming a massive, multi-dimensional, and highly time-series industrial video database. As the core data carrier for industrial visual analysis, production anomaly detection, equipment fault tracing, safety production evidence collection, and production process review, the industrial video database differs from ordinary civilian video databases. It features a massive data volume, strong frame time sequence continuity, significant regularity in image texture, stable data statistical characteristics, and extremely high business relevance. The video data it carries contains a large amount of core confidential and commercially sensitive information, such as production process parameters, equipment operating status, production operation procedures, and safety compliance details. It is a core data asset in the industrial digital system, possessing extremely high confidentiality and traceability value. Currently, the large-scale application of industrial video databases, along with cloud sharing, cross-terminal transmission, algorithm iteration training, and compliant data flow, has become a core component of the implementation of industrial vision intelligence. However, throughout the entire process of data sharing, transmission and distribution, secondary processing, and AI model training, industrial video data is highly susceptible to security risks such as illegal copying, malicious tampering, unauthorized dissemination, and data theft. This can not only lead to the leakage of core enterprise production data but also cause serious problems such as distorted training of industrial anomaly detection models, ineffective production traceability and evidence collection, and the collapse of compliance and regulatory systems. Therefore, building efficient and adaptable copyright protection and content traceability watermarking technologies for industrial video databases has become a key technical requirement for ensuring the security of industrial video data, standardizing data flow, and supporting the implementation of industrial vision intelligence applications. Existing traditional video watermarking technologies are mostly designed for civilian entertainment videos and general surveillance videos, and generally suffer from poor adaptability, insufficient robustness, and susceptibility to disrupting video statistical characteristics, making them difficult to meet the specific application requirements of industrial video databases. Firstly, most traditional watermarking algorithms embed watermark information based on spatial domain or simple frequency domain, failing to consider the core characteristics of industrial videos: strong regularity, stable statistical features, and high inter-frame correlation. The watermark embedding process easily disturbs the inherent grayscale distribution, texture features, and temporal statistical regularities of industrial videos, disrupting the overall statistical characteristics of the video database. This leads to a significant decrease in the accuracy of subsequent visual algorithms based on video data, such as industrial anomaly detection, equipment status analysis, and intelligent quality inspection, severely impacting the accuracy of intelligent analysis results in industrial production. Secondly, industrial video data in practical applications frequently undergoes routine operations such as compression encoding, frame cropping, noise superposition, frame rate adjustment, local image enhancement, and batch preprocessing. Traditional watermarking has weak anti-interference capabilities, and after these routine processes, problems such as watermark loss, extraction failure, and watermark distortion easily occur, making its robustness unsuitable for the high-frequency batch processing business scenarios of industrial video databases. Third, existing watermarking technologies mostly focus on embedding watermarks and preventing counterfeiting in single videos, failing to adapt to the massive, batch, structured, and time-series data characteristics of industrial video databases. They cannot maintain the consistency of the overall statistical characteristics of the database, and batch watermark embedding easily leads to issues such as feature shifts across multiple videos and disordered data distribution within the database. This damages the integrity and statistical standardization of the industrial video database dataset, significantly impacting the training stability and generalization ability of industrial vision models. Furthermore, some highly robust watermarking algorithms excessively modify video pixel features to enhance anti-attack capabilities, resulting in video distortion and blurred key production details, failing to meet the stringent requirements of high-precision visual traceability and compliant evidence collection in industrial scenarios. In summary, the industry currently lacks a robust watermarking solution that can adapt to industrial video database scenarios, effectively maintain the overall statistical characteristics of the video, and possess strong anti-interference and high stability. This makes it difficult to simultaneously address the multiple requirements of industrial video data security and copyright protection, complete preservation of database statistical characteristics, and ensuring the accuracy of industrial intelligent analysis. Therefore, this invention proposes a robust watermarking method for industrial video databases that preserves statistical characteristics, aiming to overcome many technical shortcomings of existing technologies. Summary of the Invention

[0003] This invention proposes a robust watermarking method for preserving statistical properties in industrial video databases, the method comprising: (1) Method for solving parameters of multi-column residuals; (2) Adaptive rejection sampling watermark embedding method; (3) Double hash tuple grouping and majority vote watermark extraction method. The details are as follows: like Figure 1The overall technical roadmap first constructs and solves the covariance matrix based on the invariance of statistical characteristics (mean, variance, covariance). Then, it extracts the potential watermarks from the tuple data involved in the embedding in the original database. Next, it compares the potential watermarks with the target watermarks in the group. For inconsistent non-good data, it constructs perturbations using the residual column framework until the data becomes good. Finally, it uses a majority voting mechanism to extract the watermarks. (1) Method for solving the parameters of multi-column residual columns. The residual column R is a linear combination of all N numerical columns plus independent Gaussian noise. , where X1, ..., X N For numeric attribute columns, k1, ..., k N k represents the coefficients of the linear combination. N+1 k is the Gaussian noise scaling factor. N+2 The mean compensation constant is used; substituting the residual column into the three statistical constraints (mean E[R]=0, variance Var(R)+2Cov(X1, R) =0, covariance Cov(R, Xj ) = 0 for j=1) yields N+1 equations, and the residual standard deviation is... The user specifies the filling of the unique degree of freedom, thereby completely determining the N+2 parameters. (2) Adaptive rejection sampling watermark embedding method. For a floating-point number x, convert its integer part and fractional part to binary and concatenate them to obtain the binary string x. b and total number of digits L x Using key k s PK with primary key i The k positions are determined by hashing, where k is the watermark density, and the s-th position in the binary string is... j The bit value (0 or 1) of the j-th bit is the potential watermark; only the position corresponding to the decimal part is selected to control the range of disturbance reach. From the initial residual standard deviation Starting with each non-good data point, a rejection sampling loop is performed—each round independently sampling from a standard normal distribution. Construction residuals ,Will After rounding to the original data precision, check whether the potential watermark matches the target watermark; if they match, the data is considered a good number and the loop is exited; if they do not match, the iteration count is incremented by one. When the number of consecutive sampling failures reaches the adjustment interval K, the current residual standard deviation is amplified by a growth factor γ > 1. Once a data point is successfully designated as a good number, it will... Reset to initial value. When the number of sampling iterations exceeds the preset upper limit M, a three-level backoff is initiated sequentially—Level 1 backoff: Set as Perform another round of sampling with the maximum perturbation amplitude; Second-level backoff: If still unsuccessful, use a key variant. Recalculate the bit position and replace the embedded bit; third-level rollback: if it is still unsuccessful, retain the original data value and record it as a modification failure. (3) Double hash tuple grouping and majority vote watermark extraction method. Watermarks are divided into sections based on watermark density k. There are groups, of which L w To determine the length of the watermark W to be embedded; tuple grouping uses double hashing. Introducing an embedding interval p makes only The tuples participate in the grouping. During extraction, the same key and primary key as the embedding are used for grouping. Within each group, potential watermark bits are extracted from all tuples, and then a bit-by-bit majority vote is performed—the number of occurrences of 0 and 1 in each bit position of all tuples in the group is counted, and the one with more votes is determined as the watermark value of that bit in that group; the watermarks of all groups are concatenated according to the group number to obtain the complete extracted watermark. Attached Figure Description

[0004] Figure 1 This is the overall technical approach of the present invention. Detailed Implementation

[0005] The steps of this invention are as follows: Step 1: Solving for parameters of multi-column residual columns An N×N coefficient matrix is ​​constructed based on the row reorganization of the covariance matrix. The closed-form solution of the N linear combination coefficients of the residual column, the Gaussian noise coefficient, and the mean compensation coefficient is obtained by matrix inversion. Step 2: Adaptive Rejection Sampling Watermark Embedding The tuple data involved in the embedding is converted to natural binary. The watermark extraction position is determined by hash operation in combination with the block key and the primary key. The potential watermark corresponding to the tuple is extracted. The potential watermark is compared with the target watermark of the block. If the two are consistent, it is determined to be a good number and no data modification is made. If the two are inconsistent, it is a bad number. The bad number data is perturbed by constructing a residual column framework. When continuous sampling fails, the residual standard deviation is dynamically increased to accelerate convergence. After successful sampling, it is reset to the initial value. Step 3: Extracting the majority vote watermark Using the same block key, embedding interval, and number of watermark blocks as the watermark embedding stage, the watermark value of each bit is extracted through a majority voting mechanism within the block.

Claims

1. A robust watermarking method for preserving statistical properties in industrial video databases, characterized in that, The method includes: (1) Method for solving parameters of multi-column residuals; (2) Adaptive rejection sampling watermark embedding method; (3) Double hash tuple grouping and majority vote watermark extraction method.

2. The method for solving multi-column residual column parameters according to claim 1, characterized in that: An N×N coefficient matrix is ​​constructed based on the row reorganization of the covariance matrix. The closed-form solution of the N linear combination coefficients of the residual column, the Gaussian noise coefficient, and the mean compensation coefficient is obtained by matrix inversion.

3. The adaptive rejection sampling watermark embedding method according to claim 1, characterized in that: The tuple data involved in the embedding is converted to natural binary. The watermark extraction position is determined by hash operation in combination with the block key and the primary key. The potential watermark corresponding to the tuple is extracted. The potential watermark is compared with the target watermark of the block. If the two are consistent, it is determined to be a good number and no data modification is made. If the two are inconsistent, it is a bad number. The bad number data is perturbed by constructing a residual column framework. When continuous sampling fails, the residual standard deviation is dynamically increased to accelerate convergence. After successful sampling, it is reset to the initial value.

4. The method for dual hash tuple grouping and majority vote watermark extraction according to claim 1, characterized in that: Using the same block key, embedding interval, and number of watermark blocks as the watermark embedding stage, the watermark value of each bit is determined by majority vote within the block during extraction.

Images