Search

CN-121982497-A - Data set copyright audit method and device for video identification system

CN121982497ACN 121982497 ACN121982497 ACN 121982497ACN-121982497-A

Abstract

The invention discloses a data set copyright audit method and device for a video identification system. The method comprises the steps of firstly training an evaluation model based on an original video data set to obtain output probability of the original video, then carrying out minor modification on the original video data set to obtain a modified data set, inputting the modified data set into the evaluation model to inquire to obtain the output probability of the modified video, then selecting a specified number of samples by comparing difference values of the two output probabilities to obtain a released data set and an unpublished data set, and finally comparing behavior differences of a suspected model on the released data set and the unpublished data set to realize copyright audit of the video identification data set.

Inventors

  • ZHANG ZHIKUN
  • YUAN QUAN
  • HE SHIBO
  • GAO YUNJUN
  • DU LINKANG
  • CHEN MIN
  • CHEN JIMING

Assignees

  • 浙江大学

Dates

Publication Date
20260505
Application Date
20251215

Claims (9)

  1. 1. A data set copyright audit method facing a video recognition system is characterized by comprising the following steps: training based on an original video data set to obtain an evaluation model, and obtaining posterior probability of the original video data based on the evaluation model; Step two, each video in the original video data set is changed to obtain a modified video data set, and posterior probability corresponding to the modified video data set is obtained through evaluation model inquiry; step three, for each video of the video dataset, calculating the difference between posterior probabilities of the original video and the real label corresponding to the modified video, then determining a candidate set based on the difference, obtaining a modified set and a reference set by sampling the candidate set, and further obtaining a published video dataset and a non-published video dataset, comprising the following sub-steps: 3.1, for each video of the video data set, calculating and recording a difference value of posterior probability of a real label corresponding to the original video and the modified video thereof; 3.2, arranging all the differences in a descending order, and then selecting a plurality of video pairs which are ranked at the front as a candidate set; 3.3 random sampling of video pairs from candidate set, selection The individual video pairs are used as a set of modifications, The video pairs are used as a reference set; 3.4 the released video data set is composed of the modified video in the modified set and other original videos, and the unpublished video data set is composed of the original video in the modified set and other modified videos; Querying a suspicion model by utilizing a published video data set and a non-published video data set, and realizing copyright audit of a video identification data set according to behavior differences of the suspicion model, wherein the method comprises the following substeps: 4.1, inputting all video pairs of the reference set into a suspicion model for inquiring, calculating probability output difference values of original video and modified video in each video pair for real labels, and recording to obtain probability sequences Based on which an upper threshold is calculated ; 4.2 Inputting all video pairs of the modification set into the suspicion model for inquiring, calculating the probability output difference value of the original video and the modification video in each video pair for the real label, and carrying out post-processing on the probability output difference value to obtain a probability sequence ; 4.3 Binding upper threshold For probability sequences A hypothesis test is performed to determine if the video identification dataset was compromised.
  2. 2. The method for auditing copyrights of a data set for a video recognition system according to claim 1, wherein in the first step, when an evaluation model is trained, the evaluation model is a neural network architecture, the input of the model is a preprocessed video frame, the output of the model is a set of vectors, the real labels of the video correspond to values of 1, and the other labels correspond to values of 0.
  3. 3. The method for auditing copyrights of a data set for a video recognition system according to claim 1, wherein in the second step, a modified video data set is obtained by injecting a program noise into each frame of a video.
  4. 4. The method for auditing copyrights of data set for video recognition system according to claim 1, wherein in said step 4.1, an upper threshold is calculated The flow of (2) is as follows: first, the mean value of the probability sequence is calculated as follows: Next, the range of the obtained mean value is limited as follows: In the middle of The function representation will be Is limited to a minimum value To a maximum value Between them.
  5. 5. The method for auditing copyrights of a data set for a video recognition system according to claim 1, wherein in the step 4.2, the post-processing flow is as follows: If the probability output value of the original video and the modified video for the real tag is smaller than , To be the reciprocal of the label class number in the original video data set, the probability difference value of the video pair is modified to be Wherein An adjustment factor greater than 0.
  6. 6. The method for auditing copyrights of data set for video recognition system according to claim 1, wherein in step 4.3, the hypothesis test performed is Wilcoxon symbol rank test, original hypothesis Is that Alternative hypothesis Is that 。
  7. 7. A data set copyright audit device for a video recognition system, implemented by the method of any one of claims 1-6, comprising the following modules: the model training module is used for training to obtain an evaluation model based on the original video data set; The video modification module is used for injecting noise into each frame of the original video to obtain a modified video data set; the probability calculation module is used for calculating the output probability of the original video and the modified video on the real label based on the trained evaluation model and calculating the difference value of the original video and the modified video; the video determining module is used for carrying out descending order sequencing on the probability difference values, then carrying out non-repeated random sampling from the video set with the front sequencing, selecting a modification set and a reference set, and the released video data set consists of modified videos of the modification set and other original videos; And the probability judging module is used for judging whether the video identification data set to be audited is stolen or not by inquiring probability output of the suspicion model to the modification set and the reference set and then carrying out Wilcoxon sign rank test.
  8. 8. A video-recognition-system-oriented dataset copyright auditing apparatus comprising a memory and one or more processors, the memory having executable code stored therein, wherein the processor, when executing the executable code, implements the video-recognition-system-oriented dataset copyright auditing method of any of claims 1-6.
  9. 9. A computer-readable storage medium having a program stored thereon, wherein the program, when executed by a processor, implements the video recognition system-oriented dataset copyright audit method of any of claims 1-6.

Description

Data set copyright audit method and device for video identification system Technical Field The invention relates to the technical field of data security, in particular to a data set copyright audit method and device for a video identification system. Background Video recognition systems are increasingly used in reality. For example, through activity recognition in a monitoring system, whether abnormal conditions occur can be effectively judged, and timely early warning information is provided for users. To facilitate video recognition research, related research institutions may publish a series of high quality data sets. However, these video data sets are issued using strict open source licensing agreements, aimed at protecting the intellectual property rights of the data owners. Unauthorized commercial use is prone to violating relevant agreement terms and can raise serious legal and ethical concerns. Therefore, it is important to protect copyrights of the video identification data set. The existing research on copyright protection of data sets is mainly focused on the image field, and the methods cannot be effectively applied to video data. On the one hand, video data has an additional time dimension compared to image data, and on the other hand, neural network models for video recognition have higher robustness. Therefore, how to design an effective data set copyright audit scheme aiming at the characteristics of the video identification data set is a problem to be solved urgently at present. Disclosure of Invention The invention aims to provide a data set copyright auditing method and device for a video identification system, aiming at the defect that the existing data set copyright auditing method does not consider the characteristics of video data. The aim of the invention is realized by the following technical scheme: in a first aspect, the present invention provides a method for auditing copyrights of a data set for a video recognition system, the method comprising: training based on an original video data set to obtain an evaluation model, and obtaining posterior probability of the original video data based on the evaluation model; Step two, each video in the original video data set is changed to obtain a modified video data set, and posterior probability corresponding to the modified video data set is obtained through evaluation model inquiry; step three, for each video of the video dataset, calculating the difference between posterior probabilities of the original video and the real label corresponding to the modified video, then determining a candidate set based on the difference, obtaining a modified set and a reference set by sampling the candidate set, and further obtaining a published video dataset and a non-published video dataset, comprising the following sub-steps: 3.1, for each video of the video data set, calculating and recording a difference value of posterior probability of a real label corresponding to the original video and the modified video thereof; 3.2, arranging all the differences in a descending order, and then selecting a plurality of video pairs which are ranked at the front as a candidate set; 3.3 random sampling of video pairs from candidate set, selection The individual video pairs are used as a set of modifications,The video pairs are used as a reference set; 3.4 the released video data set is composed of the modified video in the modified set and other original videos, and the unpublished video data set is composed of the original video in the modified set and other modified videos; Querying a suspicion model by utilizing a published video data set and a non-published video data set, and realizing copyright audit of a video identification data set according to behavior differences of the suspicion model, wherein the method comprises the following substeps: 4.1, inputting all video pairs of the reference set into a suspicion model for inquiring, calculating probability output difference values of original video and modified video in each video pair for real labels, and recording to obtain probability sequences Based on which an upper threshold is calculated; 4.2 Inputting all video pairs of the modification set into the suspicion model for inquiring, calculating the probability output difference value of the original video and the modification video in each video pair for the real label, and carrying out post-processing on the probability output difference value to obtain a probability sequence; 4.3 Binding upper thresholdFor probability sequencesA hypothesis test is performed to determine if the video identification dataset was compromised. Further, in the first step, when the evaluation model is trained, the evaluation model is a neural network architecture, the input of the model is a preprocessed video frame, the output of the model is a set of vectors, the value corresponding to the real video tag is 1, and the values corresponding to other tags are all 0. Further,