A video desensitization method and system based on differential privacy protection

Abstract

The present invention discloses a video desensitization method and system based on differential privacy protection, which is applied to the field of artificial intelligence technology. The method includes: for a video to be desensitized uploaded by a user, segmenting it to obtain a set of video segments; using a recognition algorithm to identify sensitive elements in the obtained video segment set, and erasing the relevant sensitive elements to achieve sensitive information protection; based on the concept of differential privacy, using a video generation algorithm to generate false video segments associated with the original video content; disrupting the order of the original video segments, splicing them to generate a processed desensitized video, inserting the false video segments into the desensitized video, performing relevant processing to restore the original segment order, and providing the desensitized video to an untrusted analyst. The system can achieve privacy protection for video data containing sensitive information, thereby reducing the risk of privacy leakage during the video analysis process.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and more particularly to a video desensitization method and system based on differential privacy protection. Background Art

[0002] With the rapid development and widespread adoption of video technology, massive amounts of video are continuously being generated from personal devices and specialized video recording facilities. As a significant source of unstructured data, detailed and in-depth analysis is essential. Traffic management departments analyze road surveillance videos to gain a clear understanding of urban traffic conditions, while analysis of surveillance videos from various locations effectively monitors suspicious behavior. However, because videos contain a large amount of sensitive information, such as faces and license plates, providing unprocessed videos containing sensitive information to untrusted analysts can pose serious privacy concerns. For analysts, obtaining the original content of the video to be analyzed is unnecessary in most analysis scenarios. Most existing video desensitization techniques rely on detecting areas where sensitive elements appear in the video and simply blur the relevant areas to achieve desensitization. However, because these methods fail to quantify and constrain potential privacy leaks in the video, they are prone to privacy leaks due to insufficient processing or reduced video analysis effectiveness due to excessive processing. Differential privacy aims to evaluate the privacy guarantees provided by a privacy-preserving mechanism or algorithm. Its concept quantifies the abstract concept of privacy. A privacy-preserving mechanism that meets its requirements ensures that analysis results derived from input data with and without a single record are indistinguishable. For video analysis, differential privacy ensures that the presence or absence of any visual element does not affect the final analysis results. Therefore, a video desensitization mechanism based on differential privacy can achieve a balance between video privacy protection and analytical utility.

[0003] Therefore, a video desensitization method and system based on differential privacy protection is proposed to solve the difficulties existing in the existing technology, which is an urgent problem that needs to be solved by those skilled in the art. Summary of the Invention

[0004] In view of this, the present invention provides a video desensitization method and system based on differential privacy protection, which can achieve privacy protection for video data containing sensitive information, so as to reduce the risk of privacy leakage during video analysis.

[0005] In order to achieve the above object, the present invention provides the following technical solutions:

[0006] A video desensitization method based on differential privacy protection includes the following steps:

[0007] Step 1: Segment the video uploaded by the user to be desensitized into segments and obtain a set of video segments;

[0008] Step 2: Using a recognition algorithm to identify sensitive elements in the obtained video clip set, obtain the pixel set that constitutes the sensitive element and its related information, and erase the relevant sensitive elements to achieve sensitive information protection;

[0009] Step 3: Based on the idea of ​​differential privacy, a video generation algorithm is used to generate fake video clips associated with the original video content, so as to add noise to the video clip collection to protect sensitive information.

[0010] Step 4: The order of the original video clips is disrupted and spliced ​​together to generate a processed desensitized video. The false video clips are inserted into the desensitized video and processed to restore the original clip order. The desensitized video is provided to an untrusted analyst.

[0011] In the above method, optionally, step 1 of segmenting the video is as follows:

[0012] For a given video to be desensitized, it is divided into video segments of length L. If the length is less than L, the length is used as the segment length.

[0013] In the above method, the sensitive element identification in step 2 is optional, and the pixel set constituting the sensitive element and its related information are obtained, and the specific content is as follows:

[0014] For all sensitive elements in the video j ,j∈[1,n], using image recognition algorithm, identify and track its position in each frame of the video clip, and each sensitive element Υ j The occurrences in all frames are decomposed into pixels with RGB values, and the different RGB sets Ψ constituting the sensitive element in the video clip are obtained. j , and use |Ψ j | indicates the number of different RGB elements in the set.

[0015] In the above method, the relevant sensitive elements are optionally erased in step 2, and the specific contents are as follows:

[0016] For all sensitive elements in the video j ,j∈[1,n], the system will accept the erasure list submitted by the user, and erase the elements to be erased in the video clip identified by the system by comparing the elements to be erased in the list with the identified sensitive elements one by one;

[0017] The specific content of the erasing operation is to randomly rearrange the pixel positions and frames based on the acquired pixel set constituting the sensitive element and its related information, and randomly extract pixels and set their RGB to random values ​​to achieve the purpose of erasing the sensitive elements.

[0018] The above method is optional. The differential privacy idea in step 3 is specifically defined as follows:

[0019] For any video grayscale histogram, its grayscale level is, if there is a random algorithm To realize the mapping from the original histogram to the noise histogram, for the histograms H1, H2, if

[0020] ||H1-H2||1=d

[0021] And for any set have

[0022]

[0023] Randomized Algorithm Satisfies (∈,δ)-differential privacy, where ∈ and δ are adjustable privacy parameters, When the histogram H1 is a randomized algorithm The probability that the output belongs to the set O when the input is , d is an adjustable parameter, and ||H1-H2||1 is the 1-norm of the difference between the two histograms.

[0024] In the above method, optionally, in step 3, a false video segment associated with the original video content is generated to add noise to the video segment set, and the specific content is as follows:

[0025] First, based on the content of each intercepted segment, a grayscale histogram of the video is constructed. Then, to ensure that the effectiveness of video analysis is not reduced and the original video information is not deleted, a video generation algorithm is used based on the truncated Laplace mechanism to generate false video segments and add noise information, thereby achieving differential privacy.

[0026] The specific content of constructing the grayscale histogram of the video is that the user specifies the grayscale level of the histogram For each frame of the video clip, its image histogram is constructed separately. The histograms are superimposed frame by frame to obtain the video clip histogram. Then, each clip is superimposed to complete the construction process of the video grayscale histogram. Regarding the truncated Laplace mechanism, its specific definition includes:

[0027] Given a histogram H, the truncated Laplace mechanism Lp(ε,δ,d) adds a non-negative integer noise vector to H where η follows a distribution whose distribution function is

[0028]

[0029] Among them, there are

[0030]

[0031]

[0032] Among them, ∈, δ, and s are all adjustable privacy parameters.

[0033] In the above method, optionally, in step 4, a false video clip is inserted into the desensitized video and relevant processing is performed, the specific contents of which are as follows:

[0034] The timestamp after the fake video clip is inserted into the desensitized video and the timestamp of each video clip in the original video before being scrambled are recorded locally.

[0035] A video desensitization system based on differential privacy protection, applying any of the above-mentioned video desensitization methods based on differential privacy protection, comprising a segmentation module, a sensitive element identification and erasure module, a false video generation module, and a desensitized video generation module connected in sequence; wherein,

[0036] The segmentation module is used to segment the videos uploaded by users to obtain a set of video segments.

[0037] The sensitive element identification and erasure module is used to identify sensitive elements in the obtained video clip set using an identification algorithm, obtain the pixel set that constitutes the sensitive elements and related information, and erase the relevant sensitive elements to achieve sensitive information protection;

[0038] A fake video generation module is used to generate fake video clips associated with the original video content based on the idea of ​​differential privacy and using a video generation algorithm, so as to add noise to the video clip collection to protect sensitive information;

[0039] The desensitized video generation module is used to disrupt the order of the original video clips, splice them together to generate a processed desensitized video, insert false video clips into the desensitized video, perform relevant processing to restore the original clip order, and provide the desensitized video to an untrusted analyst.

[0040] A computer device includes: a processor and a memory, wherein the memory stores machine-readable instructions executable by the processor, and the processor is used to execute the machine-readable instructions stored in the memory. When the machine-readable instructions are executed by the processor, the processor performs the steps of the above-mentioned video desensitization method.

[0041] A computer-readable storage medium stores a computer program. When the computer program is executed by a computer device, the computer device executes the steps of the above-mentioned video desensitization method.

[0042] It can be seen from the above technical solution that compared with the existing technology, the present invention discloses a video desensitization method and system based on differential privacy protection: the system can realize privacy protection of video data containing sensitive information to reduce the risk of privacy leakage during video analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0044] Figure 1 A flowchart of a video desensitization method based on differential privacy protection provided by the present invention;

[0045] Figure 2 A system structure diagram of video desensitization based on differential privacy protection provided by the present invention;

[0046] Figure 3 This is a specific flow chart of a video desensitization method based on differential privacy protection provided by the present invention. DETAILED DESCRIPTION

[0047] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0048] The flowcharts shown in the accompanying drawings are merely illustrative and do not necessarily include all steps. For example, some steps may be decomposed, while some steps may be combined or partially combined, so the actual execution order may change according to actual circumstances.

[0049] This paper proposes a video desensitization system based on differential privacy protection. First, the video to be desensitized is segmented to obtain a collection of original video segments. Sensitive elements in each video segment are then identified and erased according to user requirements. For the erased video segments, a video generation algorithm is used to generate fake video segments associated with the original video content that meet differential privacy constraints. Finally, the original video segments are shuffled, fake video segments are inserted, and then spliced ​​together to generate the processed desensitized video, which is then provided to an untrusted analyst.

[0050] Reference Figure 1 As shown, the present invention discloses a video desensitization method based on differential privacy protection, comprising the following steps:

[0051] Step 1: Segment the video uploaded by the user to be desensitized into segments and obtain a set of video segments;

[0052] Step 2: Using a recognition algorithm to identify sensitive elements in the obtained video clip set, obtain the pixel set that constitutes the sensitive element and its related information, and erase the relevant sensitive elements to achieve sensitive information protection;

[0053] Step 3: Based on the idea of ​​differential privacy, a video generation algorithm is used to generate fake video clips associated with the original video content, so as to add noise to the video clip collection to protect sensitive information.

[0054] Step 4: The order of the original video clips is disrupted and spliced ​​together to generate a processed desensitized video. The false video clips are inserted into the desensitized video and processed to restore the original clip order. The desensitized video is provided to an untrusted analyst.

[0055] Furthermore, step 1 segments the video into segments, and the specific content is as follows:

[0056] For a given video to be desensitized, it is divided into video segments of length L. If the length is less than L, the length is used as the segment length.

[0057] Where L is a user-adjustable parameter used to balance the risk of highly sensitive information leakage caused by too long clips and the decrease in video analysis value caused by too short clips.

[0058] Specifically, step 1: for the videos to be desensitized uploaded by users, based on the idea of ​​balancing the risk of highly sensitive information leakage caused by too long clips and the decline in video analysis value caused by too short clips, the original video is divided into video clips of length L according to the adjustable parameter L value set by the user. If the length is less than L, the length is used as the clip length to obtain a set of video clips.

[0059] In this embodiment, L=10 ms is given.

[0060] Furthermore, in step 2, the sensitive element identification is performed to obtain a pixel set constituting the sensitive element and related information. The specific method includes: for the obtained video clip set, for all sensitive elements Y in the video, j ,j∈[1,n], using image recognition algorithm, identify and track its position in each frame of the video clip, and each sensitive element Υ j The occurrences in all frames are decomposed into pixels with RGB values, and the different RGB sets Ψ constituting the sensitive element in the video clip are obtained. j , and use |Ψ j | indicates the number of different RGB elements in the set.

[0061] Furthermore, in step 2, the relevant sensitive elements are erased, and the specific contents are as follows:

[0062] For all sensitive elements in the video j ,j∈[1,n], the system will accept the erasure list submitted by the user, and erase the elements to be erased in the video clip identified by the system by comparing the elements to be erased in the list with the identified sensitive elements one by one;

[0063] The specific content of the erasing operation is to randomly rearrange the pixel positions and frames based on the acquired pixel set constituting the sensitive element and its related information, and randomly extract pixels and set their RGB to random values ​​to achieve the purpose of erasing the sensitive elements.

[0064] In this embodiment, a multi-target tracking algorithm (SORT) is used to identify sensitive elements. The algorithm is based on the Faster RCNN target detection method and uses the Kalman filter algorithm and the Hungarian algorithm to greatly improve the speed of multi-target tracking. Based on the recognition results, the system will accept the erasure list submitted by the user. In this embodiment, the given erasure list is all the sensitive elements identified. By comparing the elements to be erased in the list with the identified sensitive elements one by one, the elements to be erased in the video clips identified by the system are randomly rearranged in terms of their pixel locations and frames, and randomly extract pixels and set their RGB to random values.

[0065] Furthermore, the differential privacy concept in step 3 is specifically defined as follows:

[0066] For any video grayscale histogram, its grayscale level is, if there is a random algorithm To realize the mapping from the original histogram to the noise histogram, for the histograms H1, H2, if

[0067] ||H1-H2||1=d

[0068] And for any set have

[0069]

[0070] Randomized Algorithm Satisfies (∈,δ)-differential privacy, where ∈ and δ are adjustable privacy parameters, When the histogram H1 is a randomized algorithm The probability that the output belongs to the set O when the input is , d is an adjustable parameter, and ||H1-H2||1 is the 1-norm of the difference between the two histograms.

[0071] Furthermore, in step 3, a false video segment associated with the original video content is generated to add noise to the video segment set, and the specific content is as follows:

[0072] First, based on the content of each intercepted segment, a grayscale histogram of the video is constructed. Then, to ensure that the effectiveness of video analysis is not reduced and the original video information is not deleted, a video generation algorithm is used based on the truncated Laplace mechanism to generate false video segments and add noise information, thereby achieving differential privacy.

[0073] The specific content of constructing the grayscale histogram of the video is that the user specifies the grayscale level of the histogram For each frame of the video clip, its image histogram is constructed separately. The histograms are superimposed frame by frame to obtain the video clip histogram. Then, each clip is superimposed to complete the construction process of the video grayscale histogram. Regarding the truncated Laplace mechanism, its specific definition includes:

[0074] Given a histogram H, the truncated Laplace mechanism Lp(ε,δ,d) adds a non-negative integer noise vector to H where η follows a distribution whose distribution function is

[0075]

[0076] Among them, there are

[0077]

[0078]

[0079] Where ∈, δ, and d are all adjustable privacy parameters.

[0080] In this embodiment, given ∈ = 1, δ = 0.05, d = 2; in this embodiment, given

[0081] In this example, the SVG (stochastic video generation) algorithm is used to predict and generate videos for the next period of time based on user-provided video content, thereby generating fake video clips. The algorithm encodes existing video frames into a latent space and then uses an LSTM to learn temporal connections between the video frames in this latent space. Finally, the latent vector output by the LSTM is decoded into the image space.

[0082] Step 4: The original video clips are shuffled and concatenated to generate a processed, desensitized video. The fake video clips are then inserted into the desensitized video. The timestamps of the fake video clips after insertion and the timestamps of each video clip in the original video before shuffling are recorded locally. This allows for the subsequent separation of the fake clips from the original video clips and the restoration of the original video. Finally, the desensitized video is provided to an untrusted analyst.

[0083] Reference Figure 2 As shown, a video desensitization system based on differential privacy protection applies the above-mentioned video desensitization method based on differential privacy protection, including a segmentation module, a sensitive element identification and erasure module, a false video generation module, and a desensitized video generation module connected in sequence; wherein,

[0084] The segmentation module segments the videos uploaded by users to be desensitized into segments and obtains a set of video segments;

[0085] The sensitive element identification and erasure module uses an identification algorithm to identify sensitive elements in the obtained video clip set, obtains the pixel set that constitutes the sensitive elements and related information, and erases the relevant sensitive elements to achieve sensitive information protection;

[0086] The fake video generation module, based on the idea of ​​differential privacy, uses a video generation algorithm to generate fake video clips associated with the original video content, thereby adding noise to the video clip collection to protect sensitive information;

[0087] The desensitized video generation module disrupts the order of the original video clips and splices them to generate a processed desensitized video. It also inserts false video clips into the desensitized video and performs relevant processing to restore the original clip order. The desensitized video is provided to an untrusted analyst.

[0088] A computer device includes: a processor and a memory, wherein the memory stores machine-readable instructions executable by the processor, and the processor is used to execute the machine-readable instructions stored in the memory. When the machine-readable instructions are executed by the processor, the processor performs the steps of the above-mentioned video desensitization method.

[0089] A computer-readable storage medium stores a computer program. When the computer program is executed by a computer device, the computer device executes the steps of the above-mentioned video desensitization method.

[0090] Reference Figure 3 As shown, the present invention discloses a specific flow chart of a method for video desensitization based on differential privacy protection.

[0091] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Reference can be made to the common and similar parts between the various embodiments. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method description.

[0092] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not limited to the embodiments shown herein but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A video desensitization method based on differential privacy protection, characterized in that: The steps include: Step 1: Segment the video uploaded by the user to be desensitized into segments and obtain a set of video segments; Step 2: Using a recognition algorithm to identify sensitive elements in the obtained video clip set, obtain the pixel set that constitutes the sensitive element and its related information, and erase the relevant sensitive elements to achieve sensitive information protection; Step 3: Based on the idea of ​​differential privacy, a video generation algorithm is used to generate fake video clips associated with the original video content, so as to add noise to the video clip collection to protect sensitive information. Step 4: The original video clips are disrupted and spliced ​​together to generate a processed desensitized video. The fake video clips are inserted into the desensitized video and processed to restore the original clip order. The desensitized video is provided to an untrusted analyst. In step 3, a false video segment associated with the original video content is generated to add noise to the video segment set. The specific content is: First, based on the content of each intercepted segment, a grayscale histogram of the video is constructed. Then, to ensure that the effectiveness of video analysis is not reduced and the original video information is not deleted, a video generation algorithm is used based on the truncated Laplace mechanism to generate false video segments and add noise information, thereby achieving differential privacy. The specific content of constructing the grayscale histogram of the video is that the user specifies the grayscale level of the histogram For each frame of the video clip, its image histogram is constructed separately, and the histograms are superimposed frame by frame to obtain the video clip histogram, and then each clip is superimposed to complete the video grayscale histogram construction process; Regarding the truncated Laplace mechanism, its specific definition includes: Given a histogram H, the truncated Laplace mechanism Lp(ε,δ,d) adds a non-negative integer noise vector to H where η follows a distribution whose distribution function is Among them, there are Among them, ε, δ, and d are all adjustable privacy parameters.

2. The video desensitization method based on differential privacy protection according to claim 1, characterized in that: Step 1 is to segment the video into segments, and its specific content is as follows: For a given video to be desensitized, it is divided into video segments of length L. If the length is less than L, the current length is used as the segment length.

3. The video desensitization method based on differential privacy protection according to claim 1, characterized in that: In step 2, sensitive element identification is performed to obtain the pixel set that constitutes the sensitive element and its related information. The specific contents are as follows: For all sensitive elements in the video γ j ,j∈[1,n], using image recognition algorithm, identify and track its position in each frame of the video clip, and convert each sensitive element γ j The occurrences in all frames are decomposed into pixels with RGB values, and the different RGB sets Ψ constituting the sensitive element in the video clip are obtained. j , and use |Ψ j | indicates the number of different RGB elements in the set.

4. The video desensitization method based on differential privacy protection according to claim 3 is characterized in that: In step 2, the relevant sensitive elements are erased, and the specific contents are as follows: For all sensitive elements in the video γ j ,j∈[1,n], the system will accept the erasure list submitted by the user, and erase the elements to be erased in the video clip identified by the system by comparing the elements to be erased in the list with the identified sensitive elements one by one; The specific content of the erasing operation is to randomly rearrange the pixel positions and frames based on the acquired pixel set constituting the sensitive element and its related information, and randomly extract pixels and set their RGB to random values ​​to achieve the purpose of erasing the sensitive elements.

5. The video desensitization method based on differential privacy protection according to claim 4, characterized in that: The differential privacy concept in step 3 is specifically defined as follows: For any video grayscale histogram, its grayscale level is, if there is a random algorithm To realize the mapping from the original histogram to the noise histogram, for the histograms H1 and H2, if ||H1-H2||1=d And for any set have Randomized Algorithm Satisfies (ε,δ)-differential privacy, where ε and δ are adjustable privacy parameters, When the histogram H1 is a randomized algorithm The probability that the output belongs to the set O when the input is , d is an adjustable parameter, and ||H1-H2||1 is the 1-norm of the difference between the two histograms.

6. The video desensitization method based on differential privacy protection according to claim 5, characterized in that: In step 4, a fake video clip is inserted into the desensitized video and relevant processing is performed, the specific contents of which are as follows: The timestamp after the fake video clip is inserted into the desensitized video and the timestamp of each video clip in the original video before being scrambled are recorded locally.

7. A video desensitization system based on differential privacy protection, characterized by A video desensitization method based on differential privacy protection according to any one of claims 1 to 6 is applied, comprising a segmentation module, a sensitive element identification and erasure module, a false video generation module, and a desensitized video generation module connected in sequence; wherein, The segmentation module is used to segment the videos uploaded by users to obtain a set of video segments. The sensitive element identification and erasure module is used to identify sensitive elements in the obtained video clip set using an identification algorithm, obtain the pixel set that constitutes the sensitive elements and related information, and erase the relevant sensitive elements to achieve sensitive information protection; A fake video generation module is used to generate fake video clips associated with the original video content based on the idea of ​​differential privacy and using a video generation algorithm, so as to add noise to the video clip collection to protect sensitive information; A desensitized video generation module is used to disrupt the order of the original video clips, splice them together to generate a processed desensitized video, insert fake video clips into the desensitized video, perform relevant processing to restore the original clip order, and provide the desensitized video to an untrusted analyst; The specific contents of the fake video generation module are as follows: First, based on the content of each intercepted segment, a grayscale histogram of the video is constructed. Then, to ensure that the effectiveness of video analysis is not reduced and the original video information is not deleted, a video generation algorithm is used based on the truncated Laplace mechanism to generate false video segments and add noise information, thereby achieving differential privacy. The specific content of constructing the grayscale histogram of the video is that the user specifies the grayscale level of the histogram For each frame of the video clip, its image histogram is constructed separately. The histograms are superimposed frame by frame to obtain the video clip histogram. Then, each clip is superimposed to complete the construction process of the video grayscale histogram. Regarding the truncated Laplace mechanism, its specific definition includes: Given a histogram H, the truncated Laplace mechanism Lp(ε,δ,d) adds a non-negative integer noise vector to H where η follows a distribution whose distribution function is Among them, there are Among them, ε, δ, and d are all adjustable privacy parameters.

8. A computer device, characterized in that: include: A processor and a memory, wherein the memory stores machine-readable instructions executable by the processor, and the processor is used to execute the machine-readable instructions stored in the memory. When the machine-readable instructions are executed by the processor, the processor performs the steps of the video desensitization method according to any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that A computer program is stored on a computer-readable storage medium. When the computer program is executed by a computer device, the computer device performs the steps of the video desensitization method according to any one of claims 1 to 6.

Images