A dynamic screen watermarking technique with anti-screen capture capability
Patent Information
- Application Number
- CN202610832195.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-06-10
- Publication Date
- 2026-10-30
AI Technical Summary
[0005]本发明的目的在于提供一种具有抗屏摄能力的动态屏幕水印技术，解决现有屏幕水印变换简单、易破解、抗屏摄能力弱、无法满足专利申请要求的技术问题，实现水印的动态自适应调整、抗屏摄失真、难破解、可溯源，同时兼顾水印的不可见性和鲁棒性，满足专利申请的各项要求
[0013]本技术中逆梯度注意力模块的作用为对屏幕图像进行特征提取，捕捉载体图像中鲁棒性强且视觉质量高的像素区域，将水印信息优先嵌入到这些不易被察觉和破坏的区域，解决现有技术中水印嵌入位置不合理、易产生视觉干扰或易被破坏的问题；可逆神经网络（INN）实现水印嵌入与提取的一体化编解码，利用其前向传播和反向传播的双射性质，将水印嵌入（前向传播）和水印提取（反向传播）整合到同一网络模型中，减少信息传递损失，确保水印嵌入后不会导致屏幕内容失真，同时提升水印提取的准确性，与现有技术中编码器与解码器相互独立的设计相比，能有效避免水印信息丢失，提升抗屏摄失真能力；多模态水印融合设计中，每个二进制水印序列对应唯一的终端和时间信息，便于泄密溯源，动态密钥与水印序列深度联动，不仅控制像素变换规则，还作为水印提取的验证密钥，进一步提升水印的安全性和抗破解能力，区别于现有无验证机制的简单水印方案；环境与操作自适应调整机制中，当环境光照较强（如户外），适当增大像素变换系数（a、b、c），确保水印在强光下仍可提取，当环境光照较弱（如室内昏暗环境），减小变换系数，避免水印产生明显视觉干扰，兼顾水印的不可见性和鲁棒性。
Smart Images
Figure FT_1 
Figure FT_2 
Figure FT_3
Abstract
Description
Technical Field
[0001] This invention relates to the field of digital watermarking technology, specifically to a screen watermarking technology that is applied to screen display scenarios, has anti-screen capture attack capabilities, and dynamically adaptively adjusts. It can be widely used in scenarios such as preventing leakage of sensitive enterprise data, copyright protection, and terminal traceability, and belongs to the field of digital security and anti-counterfeiting technology. Background Technology
[0002] Existing screen dynamic watermarking technologies mostly divide the screen into a fixed m*n grid and represent watermark information 0 or 1 by adding or subtracting fixed values (such as R+3, G-2, B+1) to the color channels of pixels within the grid. Its core drawback is that the transformation calculation is too simple and belongs to a conventional method that is easy for a person skilled in the art to think of. It has been covered by existing patents and cannot meet the novelty and inventiveness requirements of patent applications.
[0003] Meanwhile, existing technologies of this kind have the following shortcomings: First, they do not consider the screen capture leakage scenario in actual applications. Watermarks are prone to moiré patterns, blurring, and perspective distortion after being captured by mobile phones or cameras, making it impossible to accurately extract the watermark. Second, the watermark operation method is fixed, making it easy for attackers to cancel pixel changes through reverse calculation, thus losing the core function of watermark anti-counterfeiting and traceability. Third, the watermark information is out of touch with the screen content and display environment, and cannot adaptively adjust the embedding strategy, resulting in watermarks that either cause obvious visual interference, affecting the screen display effect, or have weak pixel changes, making extraction difficult, and making it difficult to balance invisibility and robustness.
[0004] Therefore, there is an urgent need for a dynamic screen watermarking technology with outstanding substantive features and significant progress, which can avoid the defects of existing technologies, meet the requirements of novelty, inventiveness and utility in patent applications, and solve the problems of easy cracking, weak anti-screen capture ability and poor adaptability of existing technologies. Summary of the Invention
[0005] The purpose of this invention is to provide a dynamic screen watermarking technology with anti-screen capture capability, solving the technical problems of existing screen watermarks being simple to transform, easily cracked, weak in anti-screen capture capability, and unable to meet patent application requirements. This invention achieves dynamic adaptive adjustment of the watermark, resistance to screen capture distortion, difficulty in cracking, and traceability, while also ensuring the watermark's invisibility and robustness, meeting all patent application requirements. The core innovation of this invention lies in abandoning the simple transformation of fixed pixel channels by adding or subtracting, and adopting an integrated architecture of "reversible neural network encoding and decoding + inverse gradient attention localization + dynamic key linkage + multimodal watermark fusion." This deeply binds the watermark to screen content features, display environment parameters, and terminal device information, achieving dynamic adaptive adjustment of the watermark. This differs from existing simple pixel transformation schemes and possesses the novelty and inventiveness required for patent applications.
[0006] This technology uses four core modules to collaboratively complete the entire closed-loop watermarking process. The specific implementation steps are as follows:
[0007] Step S1: The watermark generation module generates multimodal watermark information, which integrates three types of information: terminal device fingerprint, real-time timestamp, and encrypted QR code. The multimodal information is converted into a binary watermark sequence of variable length through a reversible hash algorithm. This sequence is different from the existing single 0 / 1 sequence, which can ensure the uniqueness and encryption of the watermark information, and provide a basis for subsequent traceability and anti-counterfeiting.
[0008] Step S2: The dynamic adjustment module generates a dynamic key. This dynamic key is generated by a hash algorithm using the terminal MAC address, the current timestamp, and the screen resolution. It is updated in real time at a preset period of 100ms, and the pixel transformation rules are adjusted synchronously after each update.
[0009] In step S3, the adaptive embedding module uses an image segmentation algorithm to dynamically divide the grid in real time according to the complexity of the screen display content. Specifically, when the screen content is complex (such as high-definition images or multiple windows overlay), a 16×16 fine grid is divided, and when the screen content is simple (such as solid color backgrounds or single text), an 8×8 coarse grid is divided. At the same time, the optimal pixel area for watermark embedding is located through an inverse gradient attention mechanism. This area is a robust area with low visual sensitivity in the screen display content, including image edges, text gaps, and color transition areas.
[0010] Step S4: The adaptive embedding module, based on a reversible neural network encoding / decoding architecture, determines the pixel transformation formula for the corresponding grid according to the binary watermark sequence (0 or 1) and the dynamic key, embedding the watermark into the optimal pixel region. Specifically, the pixel transformation formula is as follows: First, the dynamic key is converted into three sets of random coefficients a, b, and c, corresponding to the RGB color channels. Combined with the average pixel color values R_avg, G_avg, and B_avg of the current grid region, when the watermark information is 1, the transformation formula is R' = R_avg + a×(R - R_avg), G' = G_avg + b×(G - G_avg), B' = B_avg + c×(B - B_avg). When the watermark information is 0, the transformation formula is R' = R_avg - a×(R - R_avg), G' = G_avg - b×(G - G_avg), B' = B_avg - c×(B - B_avg). The value ranges from 0.01 to 0.05 and can be dynamically adjusted according to the ambient light intensity.
[0011] Step S5: The dynamic adjustment module collects screen display parameters (brightness, contrast, color saturation), ambient light intensity, and terminal operation behavior (including screenshot operation and screen capture action detection) in real time, and dynamically adjusts the grid division method and pixel transformation intensity. When screen capture action is detected, the watermark enhancement mechanism is triggered, which specifically increases the grid pixel transformation intensity, encrypts the watermark sequence, increases the number of times the watermark is repeatedly embedded, and embeds the screen capture time and location information into the watermark to facilitate subsequent traceability. When normal operation (no screen capture, no screenshot) is detected, the normal watermark embedding intensity is maintained.
[0012] Step S6: The anti-distortion extraction module acquires the screen image containing the watermark (including the screen capture image). First, it inputs the terminal device fingerprint and the acquisition timestamp to generate the corresponding dynamic key and completes the key verification. Then, it uses a training model that simulates screen capture distortion to correct moiré patterns, blur, and perspective distortion in the acquired image. This correction process uses the Learnable Perceptual Patch Similarity (LPIPS) loss function to optimize the extraction algorithm. Subsequently, it uses a reversible neural network to decode and extract the binary watermark sequence in the grid. Finally, it parses the extracted binary sequence into multimodal traceability information such as the terminal device fingerprint, real-time timestamp, and encrypted QR code. The encrypted QR code contains sensitive data identifiers and responsible person information. If the key does not match, the valid watermark cannot be extracted.
[0013] In this technology, the inverse gradient attention module extracts features from the screen image, capturing robust and visually high-quality pixel regions. Watermark information is preferentially embedded into these less noticeable and easily damaged regions, solving the problems of unreasonable watermark embedding positions, visual interference, or susceptibility to damage in existing technologies. The invertible neural network (INN) achieves integrated encoding and decoding of watermark embedding and extraction. Utilizing its bijective properties of forward and backward propagation, it integrates watermark embedding (forward propagation) and watermark extraction (backward propagation) into the same network model, reducing information transmission loss and ensuring that watermark embedding does not cause screen content distortion. It also improves the accuracy of watermark extraction, unlike the independent encoder-decoder design in existing technologies. Compared to other methods, this approach effectively avoids watermark information loss and enhances resistance to screen capture distortion. In the multimodal watermark fusion design, each binary watermark sequence corresponds to a unique terminal and time information, facilitating traceability of leaks. The dynamic key is deeply linked with the watermark sequence, not only controlling pixel transformation rules but also serving as a verification key for watermark extraction, further enhancing the security and anti-cracking capabilities of the watermark, unlike existing simple watermark schemes without verification mechanisms. In the environment and operation adaptive adjustment mechanism, when the ambient light is strong (such as outdoors), the pixel transformation coefficients (a, b, c) are appropriately increased to ensure that the watermark can still be extracted under strong light. When the ambient light is weak (such as in a dimly lit indoor environment), the transformation coefficients are reduced to avoid significant visual interference from the watermark, balancing the invisibility and robustness of the watermark. Attached Figure Description
[0014] Figure 1 This is a diagram of the overall system architecture.
[0015] Figure 2 Flowchart for embedding watermark information on the screen;
[0016] Figure 3 A flowchart for extracting watermarks from captured screen images;
[0017] Figure 4 This is the core algorithm formula for dynamically fusing watermark data with screen display content. Detailed Implementation
[0018] This specific implementation method takes the scenario of displaying sensitive design drawings on an enterprise office computer as an example, and describes in detail the specific implementation process of the dynamic screen watermarking technology with anti-screen capture capability of the present invention according to the logical order of steps S1 to S6.
[0019] First, in step S1, the watermark generation module collects the MAC address of the company's office computer as the terminal device fingerprint, obtains the current timestamp accurate to milliseconds, and generates an encrypted QR code containing the employee's ID number, design drawing number, and confidentiality level. Then, the terminal device fingerprint, timestamp, and encrypted QR code are fused together using the SHA-256 reversible hash algorithm to generate a 64-bit binary watermark sequence, ensuring the uniqueness and encryption of the watermark information and providing a foundation for subsequent traceability.
[0020] In step S2, the dynamic adjustment module takes the computer's MAC address, current timestamp, and screen resolution of 1920×1080 as inputs and generates a dynamic key using the MD5 hash algorithm. This key is automatically updated every 100ms and is then synchronously converted into three sets of random coefficients a=0.03, b=0.02, and c=0.04 to control the amplitude of subsequent pixel transformations.
[0021] In step S3, the adaptive embedding module analyzes the complexity of the design drawing content displayed on the screen using an image segmentation algorithm. Since the drawing contains a large number of lines, color transitions and detailed information, it is determined to be a complex content area. Therefore, it divides the drawing into a 16×16 fine grid. At the same time, it uses an inverse gradient attention mechanism to extract robust and low visual sensitivity areas such as the edges of the drawing, the gaps between text and the color transition areas. These areas are then identified as the optimal pixel areas for watermark embedding.
[0022] In step S4, the adaptive embedding module, based on a reversible neural network encoding and decoding architecture, combines the binary watermark sequence generated in step S1 with the dynamic key generated in step S2. For each grid within an optimal pixel region, it combines the average pixel color values R_avg, G_avg, and B_avg of that region and performs pixel transformation according to a preset formula: when the watermark information is 1, it executes R' = R_avg + 0.03×(R - R_avg), G' = G_avg + 0.02×(G - G_avg), and B' = B_avg + 0.04×(B - B_avg); when the watermark information is 0, it executes the inverse transformation formula. The watermark embedding is completed through the forward propagation of the reversible neural network, ensuring that the pixel transformation is visually imperceptible while being identifiable at the pixel value level.
[0023] In step S5, the dynamic adjustment module collects the indoor ambient light intensity in real time as 500 lux, and keeps the current pixel transformation coefficient unchanged in combination with the ambient brightness. At the same time, it detects whether there is a screen capture or screenshot operation through the screen sensor. After detection, no abnormal operation is triggered, so the normal watermark embedding intensity is maintained and no enhancement adjustment is made to ensure that the screen display effect is not affected.
[0024] In step S6, after discovering that the design drawings have been leaked by screen capture, the anti-distortion extraction module acquires the screen capture image, inputs the computer MAC address and screen capture timestamp to generate a corresponding dynamic key for verification, and then uses a pre-trained simulated screen capture distortion model to correct moiré patterns, blurring and perspective distortion in the image. The correction process uses a learnable perceptual patch similarity (LPIPS) loss function optimization algorithm to improve extraction accuracy. Then, the binary watermark sequence is extracted by reverse decoding through a reversible neural network. Finally, the sequence is parsed into traceability information such as employee ID, design drawing number, and screen capture time, so as to achieve accurate location and traceability of the leakage behavior.
Claims
1. A dynamic screen watermarking technology with anti-screen capture capability, characterized in that, The system comprises four modules: a watermark generation module, an adaptive embedding module, a dynamic adjustment module, and an anti-distortion extraction module. These four modules work together to achieve a closed-loop process for watermark generation, embedding, dynamic updating, and extraction. Specifically, the process includes the following steps: S1: The watermark generation module generates multimodal watermark information, which integrates the terminal device fingerprint, real-time timestamp, and encrypted QR code, and converts them into a binary watermark sequence through a reversible hash algorithm; S2: The dynamic adjustment module generates a dynamic key, which is generated by a hash algorithm based on the terminal MAC address, current timestamp, and screen resolution, and is updated in real time according to a preset period. S3: The adaptive embedding module uses an image segmentation algorithm to divide a dynamic grid in real time according to the complexity of the screen display content, and uses an inverse gradient attention mechanism to locate the optimal pixel region for watermark embedding. S4: The adaptive embedding module is based on a reversible neural network encoding and decoding architecture. It performs watermark embedding at the spatial pixel level. Based on the binary watermark sequence (0 or 1) and the dynamic key, it determines the pixel transformation formula of the corresponding grid and embeds the watermark into the optimal pixel region. The dynamic key is uniquely bound to the pixel transformation rule. The same dynamic key corresponds to a unique set of pixel transformation formulas. Without the correct key, the watermark information cannot be parsed. S5: The dynamic adjustment module collects screen display parameters and terminal operation behavior in real time, and dynamically adjusts the grid division method and pixel transformation intensity. S6: The anti-distortion extraction module acquires screen images (including screen capture images) containing watermarks. After key verification and distortion correction, the watermark sequence is extracted by reverse decoding through a reversible neural network, and multimodal traceability information is obtained through parsing.
2. The dynamic screen watermarking technology with anti-screen capture capability according to claim 1, characterized in that, The dynamic grid division method in step S3 is as follows: when the screen content is complex, a fine grid is divided; when the screen content is simple, a coarse grid is divided. The optimal pixel area is the area with strong robustness and low visual sensitivity in the screen display content, including image edges, text gaps, and color transition areas.
3. The dynamic screen watermarking technology with anti-screen capture capability according to claim 1, characterized in that, The pixel transformation formula in step S4 is as follows: The dynamic key is converted into three sets of random coefficients a, b, and c, corresponding to the three RGB color channels, and combined with the average pixel color values R_avg, G_avg, and B_avg of the current grid area. When the watermark information is 1, the transformation formula is: R' = R_avg + a×(R - R_avg), G' = G_avg + b×(G- G_avg), B' = B_avg + c×(B - B_avg); When the watermark information is 0, the transformation formula is: R' = R_avg - a×(R - R_avg), G' = G_avg - b×(G - G_avg), B' = B_avg - c×(B - B_avg); The values of a, b, and c range from 0.01 to 0.05 and can be dynamically adjusted according to the ambient light intensity.
4. The dynamic screen watermarking technology with anti-screen capture capability according to claim 1, characterized in that, The dynamic key update cycle in step S2 is 100ms, and the pixel transformation rules are adjusted synchronously after each update; the terminal operation behavior in step S5 includes screenshot operation and screen capture action detection. When screen capture action is detected, the watermark enhancement mechanism is triggered.
5. The dynamic screen watermarking technology with anti-screen capture capability according to claim 1, characterized in that, The distortion correction in step S6 is performed by using a training model that simulates screen capture distortion to correct moiré patterns, blur, and perspective distortion in the acquired image, and by using the Learnable Perceptual Patch Similarity (LPIPS) loss function to optimize the extraction algorithm.
6. The dynamic screen watermarking technology with anti-screen capture capability according to claim 1, characterized in that, The encrypted QR code in the multimodal watermark information contains sensitive data identifiers and responsible person information; the key verification requires inputting the terminal device fingerprint and real-time timestamp to generate a corresponding dynamic key, and if the keys do not match, a valid watermark cannot be extracted.
7. The dynamic screen watermarking technology with anti-screen capture capability according to claim 1, characterized in that, The specific implementation of the unique binding of the dynamic key and the pixel transformation rule is as follows: the dynamic key generates three sets of random coefficient seeds through hash transformation. These seeds remain stable within a preset period and are used to synchronously control the values of the pixel transformation formula coefficients a, b, and c in step S4, as well as the granularity of the dynamic grid in step S3. This makes the watermark embedding rule change periodically with the dynamic key, with the embedding rule being consistent within the same key period and different in different key periods.