Video data set ownership verification method and device for Top-K output
By sampling, modifying, and smoothing the original video dataset, and combining the Top-K output of the suspect model for score transformation and hypothesis testing, the problem of verifying ownership of video datasets is solved, and effective dataset protection is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- ZHEJIANG UNIV
- Filing Date
- 2025-12-15
- Publication Date
- 2026-04-28
AI Technical Summary
Existing technologies struggle to effectively protect ownership of video datasets, especially in Top-K output settings, failing to adequately address the characteristics of video data, and unauthorized commercial use may raise legal and ethical issues.
By randomly sampling, modifying, and smoothing the original video dataset, benchmark and comparison datasets are formed. The Top-K output of the suspect model is used for score transformation and hypothesis testing to verify the ownership of the video dataset.
Effectively verify the ownership of video datasets, ensure that the visual effects of modified videos are not altered, and confirm the authenticity of the datasets through statistical methods to prevent unauthorized use.
Smart Images
Figure CN121940592A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the technical field of data security, and in particular to a method and apparatus for verifying ownership of video datasets for Top-K output. Background Technology
[0002] Video datasets are increasingly used in real-world applications. For example, medical monitoring videos can effectively identify abnormal situations and provide doctors with timely instructions. To promote related research, research institutions release a series of high-quality datasets. However, these video datasets are released under strict open-source licenses designed to protect the intellectual property rights of data owners. Unauthorized commercial use can easily violate these license terms and may lead to serious legal and ethical issues. Therefore, protecting the ownership of video datasets is crucial.
[0003] On the one hand, existing research on dataset copyright protection mainly focuses on the image domain, which is not effective for video data. On the other hand, existing research often requires access to the model's output probabilities, which is unsuitable for settings with only Top-K outputs. Therefore, how to design an effective dataset ownership verification scheme that utilizes limited output information and is tailored to the characteristics of video datasets is an urgent research problem. Summary of the Invention
[0004] The purpose of this invention is to address the shortcomings of existing dataset ownership verification methods, which have excessively high requirements for model output and fail to consider the characteristics of video data, by providing a video dataset ownership verification method and apparatus oriented towards Top-K output.
[0005] The objective of this invention is achieved through the following technical solution: In a first aspect, the present invention provides a method for verifying ownership of video datasets for Top-K output, the method comprising: Step 1: Randomly select several original videos from the original video dataset, and then modify a portion of frames from each selected video based on a pre-trained model. This includes the following sub-steps: 1.1 Sample the original video dataset and randomly select several original videos from it; 1.2 For each selected original video, several original video frames are randomly selected. For each selected frame, optimization is performed based on the pre-trained model to obtain the corresponding modified frame. 1.3 Combine these modified frames with the unselected original frames to obtain a modified video. Use a filter to smooth the modified video to obtain the final modified video. 1.4 Combine the modified videos obtained in step 1.3 with their corresponding original videos to form a benchmark dataset; Step 2: Replace the relevant original videos in the original video dataset with these modified videos to obtain a published dataset, which consists of the replaced modified videos and other original videos. Step 3: Randomly sample the original video dataset again, ensuring that the sampled videos do not overlap with the original videos selected in Step 1. Then, obtain the modified videos using the same processing methods as Steps 1.2 and 1.3. Combine these modified videos with their corresponding original videos to form a comparison dataset. Step four involves inputting the original videos and corresponding modified videos from the benchmark dataset into the suspect model, and inputting the original videos and corresponding modified videos from the comparison dataset into the suspect model, performing queries respectively, and verifying the ownership of the video dataset based on the Top-K output of the suspect model. This includes the following sub-steps: 4.1 For each original video and its corresponding modified video in the benchmark dataset, input them into the suspect model to obtain and record the Top-K query results; 4.2 Combining the real tags of the original videos, the Top-K query results are transformed to obtain the ratings of each original video and its corresponding modified video. The difference between the ratings of the original video and its corresponding modified video is then calculated to obtain the rating difference sequence. ; 4.3 For each original video and its corresponding modified video in the comparison dataset, obtain and record the Top-K query results in the same manner as in step 4.1. Obtain the scores of the original video and its corresponding modified video in the comparison dataset in the same manner as in step 4.2, and calculate the difference between the scores of the original video and its corresponding modified video to obtain the difference sequence. ; 4.4 Based on difference sequences and Hypothesis testing was conducted to verify the ownership of the video dataset.
[0006] Furthermore, the optimization problem formed in step 1.2 is as follows: In the formula Represents the original frame. Represents a modified frame. For distance measurement function, This is an amplitude limiting function, i.e., maximizing the original frame. and modify frames The distance is the optimization target, and the maximum modification range for each pixel does not exceed [the specified value]. These are constraints.
[0007] Furthermore, in step 1.3, a Gaussian filter is used to smooth the modified video, that is, a weighted average of the neighboring pixels is applied.
[0008] Furthermore, in step 4.1, the Top-K query result is the top K tags with the highest confidence ranking.
[0009] Furthermore, in step 4.2, the process of transforming the Top-K query results to obtain the corresponding scores is as follows: If the video's corresponding real tag is among the top K tags, then the video's rating is the reciprocal of its tag confidence ranking. In other words, if the real tag's confidence ranking is [missing value], then the video's rating is [missing value]. If there are 1, then the corresponding score is 1. Otherwise, the video will receive a score of 0.
[0010] Furthermore, in step 4.4, the null hypothesis of the hypothesis test is: The mean is not lower than The mean, the alternative hypothesis is The mean is significantly smaller than The mean of the values was determined using a one-sided independent samples t-test.
[0011] Secondly, the present invention provides a device for verifying ownership of video datasets for Top-K output, comprising: Video optimization module: Samples several original videos from the original video dataset, then optimizes some frames of the original videos based on a pre-trained model to obtain modified videos. Based on the modified videos, a benchmark dataset and a release dataset are obtained. Then, non-overlapping sampling is performed again, and a comparison dataset is obtained through optimization. Query and transformation module: By inputting the modified videos and corresponding original videos in the benchmark dataset into the suspect model, the Top-K output is obtained, and then it is converted into the corresponding scores to obtain a score sequence. Then, the modified videos and corresponding original videos in the comparison dataset are queried and transformed in the same way to obtain another score sequence. Hypothesis testing module: Based on two rating sequences, perform hypothesis testing to verify ownership of the video dataset.
[0012] Thirdly, the present invention provides a video dataset ownership verification device for Top-K output, including a memory and one or more processors, wherein the memory stores executable code, and when the processor executes the executable code, it implements the video dataset ownership verification method for Top-K output as described in the first aspect.
[0013] Fourthly, the present invention provides a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the video dataset ownership verification method for Top-K output as described in the first aspect.
[0014] The beneficial effects of this invention are as follows: 1. By optimizing and smoothing operations, the visual effects of the modified video are ensured while amplifying the differences in behavior between the original and modified videos; 2. By querying the Top-K outputs of the suspected video and the corresponding original video, and performing score conversion, the output labels are effectively transformed into numerical values that can be used for statistical calculations; 3. Hypothesis testing based on the rating difference sequence can effectively verify the ownership of the video dataset. Attached Figure Description
[0015] Figure 1 A flowchart of a video dataset ownership verification method for Top-K output provided in an embodiment of the present invention; Figure 2 This is an example of the Top-K output of a suspect model for an original video set provided in an embodiment of the present invention; Figure 3 This invention provides an example of modifying the Top-K output of a suspect model for a video set. Figure 4 This is a structural block diagram of a video dataset ownership verification device for Top-K output provided in an embodiment of the present invention. Detailed Implementation
[0016] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.
[0017] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.
[0018] like Figure 1 As shown in the figure, this embodiment provides a method for verifying ownership of video datasets for Top-K output, and the steps are as follows: Step 1: Randomly select a specified proportion of original videos from the original video dataset, and then modify a portion of frames from each selected video based on a pre-trained model. This includes the following sub-steps: 1.1 Sample the original video dataset and randomly select a proportion of it. The original video, such as settings ; 1.2 For each selected original video, the random sampling ratio is... The original video frames, such as the settings For each extracted frame, optimization is performed based on a pre-trained model (e.g., a ResNet18 model pre-trained on the ImageNet dataset) to obtain the corresponding modified frame. The specific implementation details corresponding to the resulting optimization problem are as follows: In the formula Represents the original frame. Represents modifying frames, function For distance measurement function, This is an amplitude limiting function, meaning it maximizes the original frame and the modified frame. The optimization objective is to optimize the distance between the original frame and the modified frame (i.e., the Euclidean distance between the two sets of results obtained by inputting them into the pre-trained model). The maximum modification magnitude for each pixel (ranging from 0 to 255) shall not exceed [a certain value]. For constraints, such as setting If the original value of a pixel is 100, then the modified value of the pixel must be between 90 and 110.
[0019] 1.3 These modified frames are combined with the unselected original frames to obtain a modified video. The modified video is then smoothed using a filter to obtain the final modified video. The specific implementation details are as follows: A Gaussian filter is used to smooth the modified video, which involves weighting the neighboring pixels.
[0020] 1.4 Combine the modified videos obtained in step 1.3 with their corresponding original videos to form a benchmark dataset; Step 2: Replace the relevant original videos in the original video dataset with these modified videos to obtain a published dataset, which consists of the replaced modified videos and other original videos. Step 3: Randomly sample the original video dataset again, ensuring that the sampled videos do not overlap with the original videos selected in Step 1. Referring to Step 1.2, modify some frames of the selected videos based on the pre-trained model, including the following sub-steps: 3.1 Sample the original video dataset, ensuring the sampled videos do not overlap with the videos selected in step one, and randomly select videos with a proportion of [missing information]. The original video, such as settings ; 3.2 For each selected original video, randomly sample a proportion using the same processing method as in step 1.2. The original video frames are extracted, and each frame is optimized based on a pre-trained model to obtain the corresponding modified frame. 3.3 Following the same processing method as step 1.3, combine these modified frames with the unselected original frames to obtain a modified video. Then, use a filter to smooth the modified video to obtain the final modified video. 3.4 Combine the modified videos obtained in step 3.3 with their corresponding original videos to form a comparison dataset; Step four involves inputting the original videos and corresponding modified videos from the benchmark dataset into the suspect model, and inputting the original videos and corresponding modified videos from the comparison dataset into the suspect model, performing queries respectively, and verifying the ownership of the video dataset based on the Top-K output of the suspect model. This includes the following sub-steps: 4.1 For each original video and its corresponding modified video in the benchmark dataset, input them into the suspect model to obtain and record the Top-K query results; The specific implementation details are as follows: Top-K results are the K tags with the highest confidence ranking, for example, K = 3. Figure 2 As shown, after querying the suspect model, the original video 1 outputs running, basketball, and tennis in order of confidence ranking. Figure 3 As shown, after querying the suspect model, the modified video 1 outputs diving, basketball, and running in order of confidence ranking.
[0021] 4.2 Combining the real tags of the original videos, the Top-K query results are transformed to obtain the ratings of each original video and its corresponding modified video. The difference between the ratings of the original video and its corresponding modified video is then calculated to obtain a rating difference sequence. ; The specific implementation details for transforming Top-K query results to obtain corresponding scores are as follows: If the video's corresponding real tag is among the top K tags, then the video's rating is the reciprocal of its tag confidence ranking. In other words, if the real tag's confidence ranking is [missing value], then the video's rating is [missing value]. If there are 1, then the corresponding score is 1. Otherwise, the video will receive a score of 0.
[0022] For example Figure 2 The true label for the original video 1 is "running," and the first output of the suspect model is "running," therefore the corresponding transformation score is 1. Figure 3The real label for modified video 1 is "running," and the second output of the suspect model is also "running," therefore the corresponding conversion score is... The corresponding score difference sequence .
[0023] 4.3 For each original video and its corresponding modified video in the comparison dataset, obtain and record the Top-K query results in the same manner as in step 4.1. Then, obtain the scores of the original video and its corresponding modified video in the comparison dataset in the same manner as in step 4.2, and calculate the difference between the scores of the original video and its corresponding modified video to obtain the difference sequence. ; 4.4 Based on difference sequences and Hypothesis testing was conducted to verify the ownership of the video dataset.
[0024] The specific implementation details are as follows: The null hypothesis of the hypothesis test is The mean is not lower than The mean, the alternative hypothesis is The mean is significantly smaller than The mean of the sample was determined using a one-sided independent samples t-test. An example of the test is shown below: The first step is to calculate the basic statistics (sample size, mean, standard deviation) of the two difference sequences. The second step is to perform a homogeneity of variance test, which involves calculating the variances of the two samples, then dividing the larger variance by the smaller variance to obtain the F statistic, thereby determining the degrees of freedom. Next, based on the F distribution, the two-sided significance p-value is calculated. If this value is less than the significance level (e.g., 0.05), the null hypothesis of equal variances is rejected, and the variances are considered to be unequal; otherwise, the variances are considered to be equal. The third step is to select the t-test method based on the results of the homogeneity of variance test: if the variances are equal, use the equal variance t-test; otherwise, use Welch's t-test. The fourth step is to calculate the t-statistic and the corresponding degrees of freedom; Fifth step, calculate the one-sided p-value (in the t-distribution, find the cumulative probability corresponding to the t-statistic). Step 6: Make a decision based on the one-sided p-value: If the one-sided p-value is less than the significance level (e.g., 0.05), then reject the null hypothesis and conclude that... The mean is significantly lower than If the video dataset has been stolen, the null hypothesis cannot be rejected, and the video dataset is considered not to have been stolen.
[0025] Another aspect of this invention discloses a device for verifying ownership of video datasets for Top-K output, such as... Figure 4 As shown, the device includes the following modules: Video optimization module: Samples a specified proportion of video, then optimizes some frames of the original video based on a pre-trained model to obtain a modified video dataset for publication, and then performs non-overlapping sampling again to obtain a comparison dataset through optimization; Query and transformation module: By inputting the modified video and the corresponding original video into the suspect model, the Top-K output of the suspect model is obtained, and then it is converted into the corresponding score to obtain a score sequence. Then, the modified video and the corresponding original video in the comparison dataset are queried and transformed in the same way to obtain another score sequence. Hypothesis testing module: Based on two rating sequences, perform hypothesis testing to verify ownership of the video dataset.
[0026] The present invention also discloses a video dataset ownership verification device for Top-K output, including a memory and one or more processors. The memory stores executable code, and when the processor executes the executable code, it implements the video dataset ownership verification method for Top-K output as described above.
[0027] The present invention also discloses a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the above-described method for verifying ownership of video datasets for Top-K output.
[0028] The above description is merely a preferred embodiment of the present invention. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make many possible variations and modifications to the technical solutions of the present invention using the methods and techniques disclosed above, or modify them into equivalent embodiments with equivalent changes, without departing from the scope of the technical solutions of the present invention. Therefore, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solutions of the present invention shall still fall within the protection scope of the technical solutions of the present invention.
Claims
1. A method for verifying ownership of video datasets with Top-K output, characterized in that, The method includes: Step 1: Randomly select several original videos from the original video dataset, and then modify a portion of frames from each selected video based on a pre-trained model. This includes the following sub-steps: 1.1 Sample the original video dataset and randomly select several original videos from it; 1.2 For each selected original video, several original video frames are randomly selected. For each selected frame, optimization is performed based on the pre-trained model to obtain the corresponding modified frame. 1.3 Combine these modified frames with the unselected original frames to obtain a modified video. Use a filter to smooth the modified video to obtain the final modified video. 1.4 Combine the modified videos obtained in step 1.3 with their corresponding original videos to form a benchmark dataset; Step 2: Replace the relevant original videos in the original video dataset with these modified videos to obtain a published dataset, which consists of the replaced modified videos and other original videos. Step 3: Randomly sample the original video dataset again, ensuring that the sampled videos do not overlap with the original videos selected in Step 1. Then, obtain the modified videos using the same processing methods as Steps 1.2 and 1.
3. Combine these modified videos with their corresponding original videos to form a comparison dataset. Step four involves inputting the original videos and corresponding modified videos from the benchmark dataset into the suspect model, and inputting the original videos and corresponding modified videos from the comparison dataset into the suspect model, performing queries respectively, and verifying the ownership of the video dataset based on the Top-K output of the suspect model. This includes the following sub-steps: 4.1 For each original video and its corresponding modified video in the benchmark dataset, input them into the suspect model to obtain and record the Top-K query results; 4.2 Combining the real tags of the original videos, the Top-K query results are transformed to obtain the ratings of each original video and its corresponding modified video. The difference between the ratings of the original videos and their corresponding modified videos is then calculated to obtain a rating difference sequence. ; 4.3 For each original video and its corresponding modified video in the comparison dataset, obtain and record the Top-K query results in the same manner as in step 4.
1. Obtain the scores of the original video and its corresponding modified video in the comparison dataset in the same manner as in step 4.2, and calculate the difference between the scores of the original video and its corresponding modified video to obtain the difference sequence. ; 4.4 Based on difference sequences and Hypothesis testing was conducted to verify the ownership of the video dataset.
2. The method for verifying ownership of video datasets for Top-K output according to claim 1, characterized in that, The optimization problem formed in step 1.2 is as follows: In the formula Represents the original frame. Represents a modified frame. For distance measurement function, This is an amplitude limiting function, i.e., maximizing the original frame. and modify frames The distance is the optimization target, and the maximum modification range for each pixel does not exceed [the specified value]. These are constraints.
3. The method for verifying ownership of video datasets for Top-K output according to claim 1, characterized in that, In step 1.3, a Gaussian filter is used to smooth the modified video, that is, a weighted average of the neighboring pixels is applied.
4. The method for verifying ownership of video datasets for Top-K output according to claim 1, characterized in that, In step 4.1, the Top-K query result is the top K tags with the highest confidence ranking.
5. The method for verifying ownership of video datasets for Top-K output according to claim 1, characterized in that, In step 4.2, the process of transforming the Top-K query results to obtain the corresponding scores is as follows: If the video's corresponding real tag is among the top K tags, then the video's rating is the reciprocal of its tag confidence ranking. In other words, if the real tag's confidence ranking is [missing value], then the video's rating is [missing value]. If there are 1, then the corresponding score is 1. Otherwise, the video will receive a score of 0.
6. The method for verifying ownership of video datasets for Top-K output according to claim 1, characterized in that, In step 4.4, the null hypothesis for hypothesis testing is: The mean is not lower than The mean, the alternative hypothesis is The mean is significantly smaller than The mean of the values was determined using a one-sided independent samples t-test.
7. A device for verifying ownership of a video dataset oriented towards Top-K output, implemented using the method described in any one of claims 1-6, characterized in that, Includes the following modules: Video optimization module: Samples several original videos from the original video dataset, then optimizes some frames of the original videos based on a pre-trained model to obtain modified videos. Based on the modified videos, a benchmark dataset and a release dataset are obtained. Then, non-overlapping sampling is performed again, and a comparison dataset is obtained through optimization. Query and transformation module: By inputting the modified videos and corresponding original videos in the benchmark dataset into the suspect model, the Top-K output is obtained, and then it is converted into the corresponding scores to obtain a score sequence. Then, the modified videos and corresponding original videos in the comparison dataset are queried and transformed in the same way to obtain another score sequence. Hypothesis testing module: Based on two rating sequences, perform hypothesis testing to verify ownership of the video dataset.
8. A device for verifying ownership of a video dataset for Top-K output, comprising a memory and one or more processors, wherein the memory stores executable code, characterized in that, When the processor executes the executable code, it implements the video dataset ownership verification method for Top-K output as described in any one of claims 1-6.
9. A computer-readable storage medium having a program stored thereon, characterized in that, When the program is executed by the processor, it implements the video dataset ownership verification method for Top-K output as described in any one of claims 1-6.