User privacy protection method and device for face recognition system

Abstract

The invention discloses a user privacy protection method and device for a face recognition system, and belongs to the technical field of mobile equipment and software security, and the method comprises the steps: carrying out the edge detection and lightweight semantic segmentation parallel processing of an input image, and generating edge strength mapping and semantic confidence mapping; generating a privacy grading mask through dynamic weight fusion edge strength mapping and semantic confidence mapping; dividing the image into three levels of privacy areas through a privacy grading mask; performing hierarchical encryption on the three-level privacy area through three non-periodic chaos sequences generated in advance to complete user privacy protection; wherein the three non-periodic chaotic sequences are iteratively generated through a Young's hyperchaotic system based on a Hash value spliced by a user-defined key and system time; according to the invention, picture encryption according to the privacy level is realized, and the user privacy in face recognition is protected.

Description

Technical Field

[0001] The present invention relates to a user privacy protection method and device for a face recognition system, belonging to the technical field of mobile device and software security. Background Art

[0002] In recent years, with the development of large model technology, the face has become an important biometric feature, and face recognition systems are widely used, especially on mobile devices. However, in the process of interacting with the cloud, there is a risk of leaking user privacy, not only leaking face features, but also possibly leaking background features outside the face.

[0003] Existing face protection measures are mainly based on image encryption or noise addition. They have the following problems: 1. Existing image encryption relies on high-order hyperchaos and multiple rounds of iteration, with low encryption efficiency and high computing power requirements, which are not suitable for deployment on mobile devices and low-computing-power intelligent devices.

[0004] 2. The protection method of noise addition is irreversible and cannot play a role in scenarios where decryption is required, such as data set collection, network video transmission, etc.

[0005] 3. Security protection measures focus on preventing theft and forgery of faces, and do not protect the digital footprints and privacy information carried in the background area outside the face. Summary of the Invention

[0006] The purpose of the present invention is to overcome the deficiencies in the prior art, and provide a user privacy protection method and device for a face recognition system, which realizes image encryption according to privacy levels and protects user privacy in face recognition.

[0007] To achieve the above purpose, the present invention is implemented by the following technical solutions: In the first aspect, the present invention provides a user privacy protection method for a face recognition system, including: Performing parallel processing of edge detection and lightweight semantic segmentation on the input image to generate an edge intensity map and a semantic confidence map; Generating a privacy classification mask by dynamically weighting and fusing the edge intensity map and the semantic confidence map; Dividing the image into three levels of privacy regions through the privacy classification mask; Performing hierarchical encryption on the three levels of privacy regions through three pre-generated aperiodic chaotic sequences to complete user privacy protection; Wherein, the three aperiodic chaotic sequences are iteratively generated by the Young's hyperchaotic system based on the hash value of the concatenation of the user-defined key and the system time.

[0008] Further, generating a privacy classification mask by dynamically weighting and fusing the edge intensity map and the semantic confidence map includes: Based on the local semantic confidence mean and the edge intensity standard deviation , calculate the fusion weight of semantic segmentation and edge detection. The formula is as follows: ; Where, is the fusion weight of semantic segmentation, is the fusion weight of edge detection; Generate a fusion mask through the weighting formula . Where, is the semantic confidence map, is the edge intensity map.

[0009] Further, dividing the image into three levels of privacy regions by the privacy classification mask includes: According to the threshold range into which the fusion mask falls, divide the image into three levels of privacy regions: the facial feature region, the facial contour region, and the background region.

[0010] Further, the generation method of the three aperiodic chaotic sequences includes: Based on the user-defined key and the system time concatenation to obtain a hash value. The formula is as follows: ; Where, is the user input key, is the system time, is the hash value after hash processing, represents the hash function, which is used to perform hash transformation on the input key and output a 256-bit hash sequence; Input the hash value after hash processing into the following Young's hyperchaotic system for iteration: ; Where, is the control variable of the system, , , are the input values for each iteration, , , are the output values for each iteration; Finally, all the elements of the iteration form three aperiodic chaotic sequences.

[0011] Further, performing hierarchical encryption on the three levels of privacy regions by the three pre-generated aperiodic chaotic sequences includes: Apply the stream key diffusion algorithm to the background area, dynamically generate the stream key through the chaotic sequence, and perform diffusion encryption on all image pixels; Apply the improved Arnold transform to the face contour area, and combine the high-frequency trigonometric function to perform coordinate scrambling encryption on the pixel blocks; Apply the channel-coupled chaotic scrambling algorithm to the face feature area, and encrypt by coupling the pixel values of the RGB channels through the chaotic sequence.

[0012] Furthermore, the step of applying the stream key diffusion algorithm to the background area, dynamically generating the stream key through the chaotic sequence, and performing diffusion encryption on all image pixels includes: Based on the elements of the chaotic sequence Generate the stream key , the formula is as follows: ; Where sin(10 ); Perform diffusion encryption on all image pixels through the formula ; Wherein, represents the value of the th pixel after diffusion, respectively represent the th and th elements in the stream key sequence, is the original pixel.

[0013] Furthermore, the step of applying the improved Arnold transform to the face contour area, and combining the high-frequency trigonometric function to perform coordinate scrambling encryption on the pixel blocks includes: Divide the face contour area into 32×32 pixel blocks, and through the dynamic parameter formula ; Perform non-linear scrambling encryption on the pixel coordinates, wherein, represents the pixel block coordinates after transformation, is the original block coordinates, is the control parameter value, is the width and height of the image.

[0014] Furthermore, the step of applying the channel-coupled chaotic scrambling algorithm to the face feature area, and encrypting by coupling the pixel values of the RGB channels through the chaotic sequence, the formula is as follows: ; Wherein, represents the three color channels of the picture, then represents the three channels after transformation.

[0015] In a second aspect, the present invention provides a user privacy protection device for a face recognition system, including: A memory for storing computer programs / instructions; A processor for executing the computer programs / instructions to implement the steps of the method described in any one of the foregoing.

[0016] In a third aspect, the present invention provides a computer-readable storage medium having stored thereon a computer program, which when executed by a processor implements the steps of the method described in any one of the foregoing.

[0017] Compared with the prior art, the beneficial effects achieved by the present invention are as follows: The present invention provides a user privacy protection method and device for a face recognition system, and realizes efficient and secure protection through the following innovations: (1) Fast face recognition and privacy grading algorithm design: A fast face recognition and privacy grading algorithm based on edge detection and lightweight semantic segmentation is proposed to achieve fast recognition and positioning of the face area, and by calculating the privacy weight, the pictures containing faces collected by the face recognition system are divided into regional blocks with different privacy levels; (2) Strong chaotic sequence generation method: Based on the Young's hyperchaotic system, a hash generation anti-collision chaotic sequence that fuses user-defined keys and system time is proposed to solve the threat that traditional chaotic algorithms are vulnerable to differential attacks, providing the possibility for realizing lightweight image encryption; (3) Lightweight hierarchical encryption algorithm: For regions with different privacy degrees, based on the designed strong privacy chaotic sequence generation method, a stream key diffusion model, a lightweight Arnold transformation method, and a channel-coupled chaotic scrambling method are respectively proposed to achieve picture encryption according to the privacy level and protect the user privacy in face recognition.

[0018] This solution solves the problems of low encryption efficiency and irreversibility in the prior art while protecting the privacy of background information, and is applicable to high-real-time scenarios such as face recognition and video transmission. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 is a flowchart of a user privacy protection method for a face recognition system provided by an embodiment of the present invention; Figure 2 is a schematic diagram of the collaborative work process of each module provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0020] The present invention will be further described below with reference to the drawings. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention and should not be used to limit the protection scope of the present invention.

[0021] Example 1. This example introduces a user privacy protection method for a face recognition system, including: Perform parallel processing of edge detection and lightweight semantic segmentation on the input image to generate an edge intensity map and a semantic confidence map; Fuse the edge intensity map and the semantic confidence map through dynamic weights to generate a privacy grading mask; Divide the image into three levels of privacy regions through the privacy grading mask; Perform hierarchical encryption on the three levels of privacy regions through three pre-generated aperiodic chaotic sequences to complete user privacy protection; Among them, the three aperiodic chaotic sequences are iteratively generated by the Young's hyperchaotic system based on the hash value of the concatenation of the user-defined key and the system time.

[0022] As Figure 1 and Figure 2 shown, the application process of the user privacy protection method for the face recognition system provided in this example specifically involves the following steps: S1. Fast face recognition and privacy grading: For the picture captured by the camera, use the edge detection-semantic segmentation model to quickly and efficiently divide the entire picture into three privacy levels: (Ⅰ) Generate the face geometric contour through the edge detection modality, where the edge detection threshold is dynamically calculated based on the local statistical characteristics of the image, and the following calculation formula is designed for the dynamic fusion weight: ； ； Among them, is the local semantic confidence mean, is the edge intensity standard deviation; is the fusion weight of semantic segmentation, is the fusion weight of edge detection; (Ⅱ) Generate a pixel-level semantic confidence map through the lightweight semantic segmentation network MobileNetV3-Small, and the lightweight semantic segmentation confidence model outputs ∈[0,1]; (Ⅲ) Perform weighted fusion on the edge intensity map and the semantic confidence map to obtain three levels of regions, and its fusion mask is: ; Among them, the semantic segmentation confidence map is the output value of the semantic segmentation model, ∈[0,1], and is the output of the Canny algorithm (an edge detection algorithm); When When it is marked as the first-level area, 0.6 ≤ ≤ 0.9 is marked as the second-level area, and all pixels of the picture belong to the third-level area range; the first-level privacy area is the area with the highest privacy level and belongs to the face feature area in the picture. The feature vector values extracted from this part of the area during the face collection process contain all the data for biometric individual authentication and are extremely sensitive. Therefore, relatively complex compound chaotic encryption will be performed.

[0023] The second-level privacy area is the area within the face frame. Since the face picture involves personal privacy, this part has the second-priority privacy protection. Considering the balance between security and computational overhead, the improved Arnold transform (an image scrambling algorithm) can better adapt to this scenario.

[0024] The third-level privacy area includes the above areas and the background area. The digital footprints carried by the background area, such as scene information, landmark buildings, etc., will expose the privacy information of the photographer. However, considering the balance between the efficiency and security of the entire system, it is marked as a privacy area with a lower priority for quick processing. At the same time, all pixels are defined as the third-level privacy area to participate in the processing, so that all pixels participate in a round of diffusion to avoid exposing the pixel statistical law after the first- and second-level areas are scrambled.

[0025] S2. Pre-generation of hyperchaotic sequences. At the module design level, the chaotic sequence is parallel to other modules and pre-computed to improve system efficiency. By iterating the Young's system, 3 hyperchaotic sequences will be obtained for encryption, and their mathematical expressions are as follows:

[0026] 、 、 are the input values for each iteration, and the outputs 、 、 , and the next iteration is performed.

[0027] Finally, all the elements of all iterations form three mutually coupled chaotic sequences . is the control parameter of the system. The initial value selection of the system depends on an input value of the user, as well as the system time at that time. and are hashed and concatenated to obtain a collision-resistant hash value, which is input into the system for iteration after normalization processing.

[0028] At the beginning of the implementation of the entire scheme, this module will pre-compute a sufficient amount of sequences to improve efficiency, and then through a parallel architecture, it will be asynchronous with the fast face recognition and privacy classification and hierarchical encryption module.

[0029] S3. Disturbance and hierarchical encryption. After receiving the image to be encrypted marked with the privacy level, the hierarchical encryption module performs the following steps: S301. Perform random disturbance on several bits of the least significant bit plane of the picture. This disturbance can be amplified by a chaotic system to effectively counter differential attacks and plaintext attacks.

[0030] S302. Perform fast stream key encryption on the third-level privacy area, that is, all pixels of the picture. The mathematical formula for generating the stream key is: ; where is a defined non-linear combination operation, specifically sin(10 ); represents the stream key, is an element of the chaotic sequence.

[0031] The mathematical formula for stream diffusion through the stream key is: ; where represents the value of the th pixel after diffusion, respectively represent the th, th elements in the stream key sequence, is the original pixel. Calculate , and perform a simple addition transformation on the plaintext and the key stream. Then calculate introduce the sine function to make the key stream more complex and non-linear.

[0032] S303. For the second-level privacy area, that is, the part within the face frame. Since this area has privacy information in the visual sense, pixel block scrambling is designed to eliminate the visual meaning of the picture. Specifically, based on the classic Arnold transformation, by introducing high-frequency trigonometric functions as coefficients, the linear characteristics of the classic transformation itself are eliminated, and at the same time, floor discretization enhances chaos. The entire transformation process adopts a block mode. According to the pixel size of the picture, a 32-pixel size block transformation is designed to achieve stronger versatility and higher efficiency. The formula for the designed lightweight transformation is as follows: ； where represent the pixel coordinates before and after the transformation respectively, , are control parameters, is the width and height of the picture.

[0033] S304, the first-level privacy area, i.e., the face feature area, not only contains visual meaning but also can extract individual feature values, so it has a higher privacy level. However, the computational overhead of the system needs to be compatible with low-computing-power devices, so some traditional high-strength and multi-round image encryptions are not feasible. To solve this problem, we obtain higher security performance by coupling three color channels while controlling the computational overhead. The proposed channel-coupled chaotic scrambling method is as follows: ; respectively represent the three color channels of the original image, 、 、 are elements in the chaotic sequence generated by the chaotic system, is the changed color channel.

[0034] This change couples the three color channels, enhances chaos, has good security performance, and since there are fewer pixels in the first-level privacy area, it can effectively save computational resources.

[0035] Next, in combination with a preferred embodiment, the content involved in the above embodiments will be described.

[0036] Application scenario: Real-time face privacy protection in an intelligent access control system requires real-time encryption of face and background information in the video stream.

[0037] Hardware configuration: 1. Main control platform: NVIDIA Jetson AGX Xavier (main control platform under NVIDIA): GPU: 512-core Volta architecture (NVIDIA GPU microarchitecture), CPU: 8-core ARM v8.2 64-bit, memory: 16GB LPDDR4x.

[0038] 2. Camera module: Sony IMX477 (model) sensor, supporting 1920×1080@30fps input, ROI (region of interest) programmable configuration.

[0039] 3. Encryption acceleration: Use the CUDA cores (NVIDIA GPU parallel computing units) of the main control platform to perform parallel computing on chaotic sequences and image scrambling.

[0040] 4. Software environment: a) Operating system: Ubuntu 20.04 LTS (Ubuntu operating system).

[0041] b) Deep learning framework: PyTorch 1.10 deep learning framework + TensorRT 8.2 inference acceleration engine.

[0042] c) Encryption library: OpenSSL 3.0 open-source encryption toolkit (used for SHA3-256 hash calculation).

[0043] Implementation steps: S1. Fast face recognition and privacy grading; S101. Edge detection mode implementation: Use improved Canny edge detection to optimize the threshold adaptive mechanism. Its parameter configuration is as follows: Gaussian filtering: Kernel size 5×5, standard deviation , suppressing high-frequency noise.

[0044] Gradient calculation: Sobel operator (edge detection operator, horizontal kernel [−1,0,1;−2,0,2;−1,0,1], vertical kernel transposed) Dynamic double threshold: High threshold , low threshold where and are the grayscale mean and standard deviation of the local 8×8 region.

[0045] Edge connection: Adopt the two-way scanning method to preferentially connect strong edge pixels.

[0046] Output: Edge intensity map Sedge∈[0,1], Sedge∈[0,1], normalized to a floating-point matrix.

[0047] S102. Semantic segmentation mode implementation, using lightweight MobileNetV3-Seg (lightweight image segmentation network), with the following optimizations: Backbone network: MobileNetV3-Small (miniaturized network model), the number of depthwise separable convolutional layers reduced to 8 layers, with 1.2M parameters.

[0048] Decoder design: Adopt transposed convolutional operations with a 3×3 convolutional kernel and stride = 2, and gradually upsample to the original resolution.

[0049] Simplification of the Atrous Spatial Pyramid Pooling (ASPP) module: Only retain 3 parallel branches (atrous rate = 6, 12, 18), and compress the output channels to 64.

[0050] Output layer: The Sigmoid activation function generates a pixel-level semantic confidence map ∈[0,1].

[0051] Dataset: CelebAMask-HQ dataset (30,000 face images, including annotations of facial features, contours, and backgrounds).

[0052] Loss function: Use Combined loss function to address class imbalance.

[0053] Training parameters: Adam optimizer (Adaptive Moment Estimation), lr = 1e-4, batch_size = 16.

[0054] S103, Dynamic weight fusion and region marking; Fusion formula: ; ; Implementation steps: Local statistics calculation: Divide the image into 32×32 blocks and calculate (mean of semantic confidence) and (standard deviation of edge intensity) for each block.

[0055] High semantic confidence area ( >0.7): Increase the α weight to 0.7 - 0.9.

[0056] High edge complexity area ( >0.4): Increase the β weight to 0.6 - 0.8.

[0057] Region marking rules: First-level region: >0.9, covering the facial features (eyes, nose, mouth), with an example size of 80×80 pixels.

[0058] Second-level region: 0.6 ≤ ≤0.9, covering the facial contour (example 320×320 pixels).

[0059] Third-level region: All pixels in the whole image, forced to participate in the lowest-intensity encryption.

[0060] Post-processing: Perform 3×3 median filtering on adjacent blocks to eliminate region fragmentation.

[0061] S2, Hyperchaotic sequence generation and key management; S201, User key processing; User key K: In string format, with a length of 8 - 64 characters, and the allowed character set is [A-Za-z0-9@#$%^&+=].

[0062] System time T: UTC timestamp with millisecond precision (Coordinated Universal Time timestamp), in the Unix Epoch format.

[0063] Hash concatenation: Call the EVP_sha3_256 function of the OpenSSL library (the calculation interface provided by the encryption library) to output a 64-byte hash value.

[0064] Initial value generation: Divide the hash value into three parts: ； ; ; Normalize it to the interval [−1, 1].

[0065] S202. Iterate the Yang's hyperchaotic system; Dynamic equation: ; Through Lyapunov exponent analysis ( = 0.52, = 0.18, = -1.02), ensure that the system is in a hyperchaotic state.

[0066] Control parameter is determined by optimizing the bifurcation diagram. Pre-generate 10,000 chaotic values and store them in a circular buffer.

[0067] S3. Hierarchical encryption process; S301. Encrypt the three-level area; Stream key generation: ; Encryption operation: ; Pixel block division: Divide the image into 64×64 blocks, and each block is assigned an independent chaotic sequence segment to avoid duplicate keys.

[0068] Parallel computing: Utilize CUDA to accelerate the stream key generation and XOR operation, with a throughput of 1.2 GPixel / s.

[0069] S4. Encrypt the secondary area (improved Arnold transform); Transformation formula: ； Parameter dynamic adjustment: ; is the chaotic sequence value.

[0070] Number of iterations: Number of iterations = ; Block scrambling: Divide the secondary region (320×320) into 16×16 sub - blocks and perform the transformation in parallel.

[0071] S5. Encryption of the primary region (channel - coupled chaotic scrambling); Channel - coupling formula: ; Quantization noise injection: Introduce chaotic quantization noise in equal terms to destroy the pixel correlation.

[0072] GPU acceleration: Texture memory caches RGB channel data, and the calculation time per single pixel ≤ 5ns.

[0073] The image output at this time is a secure image after three - round hierarchical encryption.

[0074] Embodiment 2. This embodiment provides a user privacy protection device for a face recognition system, including: A memory for storing computer programs / instructions; A processor for executing the computer programs / instructions to implement the steps of the method described in any one of Embodiment 1.

[0075] Embodiment 3. This embodiment provides a computer - readable storage medium, on which a computer program is stored. When the program is executed by a processor, it implements the steps of the method described in any one of Embodiment 1.

[0076] The above - mentioned are only the preferred embodiments of the present invention. It should be noted that for those of ordinary skill in the art in this technical field, without departing from the technical principle of the present invention, several improvements and modifications can be made, and these improvements and modifications should also be regarded as the protection scope of the present invention.

[0077] Those skilled in the art should understand that the embodiments of the present disclosure can be provided as a method, a system, or a computer program product. Therefore, the present disclosure can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present disclosure can take the form of a computer program product implemented on one or more computer - usable storage media (including but not limited to disk memories, CD - ROMs, optical memories, etc.) containing computer - usable program codes.

[0078] This disclosure is described with reference to the flowcharts and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the disclosure. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and combinations of flows and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to the processors of general-purpose computers, special-purpose computers, embedded processors, or other programmable data processing devices to produce a machine, such that the instructions executed by the processors of the computer or other programmable data processing devices produce means for implementing the functions specified in one or more of the flows Figure 1 one or more of the flows and / or blocks Figure 1 or means for implementing the functions specified in one or more of the blocks.

[0079] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to operate in a specific manner, such that the instructions stored in the computer-readable memory produce a manufacture including instruction means for implementing the functions specified in one or more of the flows Figure 1 one or more of the flows and / or blocks Figure 1 or means for implementing the functions specified in one or more of the blocks.

[0080] These computer program instructions can also be loaded onto a computer or other programmable data processing device, such that a series of operational steps are performed on the computer or other programmable device to produce a computer-implemented process, and thus the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one or more of the flows Figure 1 one or more of the flows and / or blocks Figure 1 or means for implementing the functions specified in one or more of the blocks.

[0081] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present disclosure and not to limit the scope of its protection. Although the present disclosure has been described in detail with reference to the above embodiments, those of ordinary skill in the art should understand that: after reading the present disclosure, those skilled in the art can still make various changes, modifications, or equivalent replacements to the specific implementation manners of the invention, but these changes, modifications, or equivalent replacements are all within the scope of protection of the claims pending for publication of the present disclosure.

Claims

1. A user privacy protection method for a face recognition system, characterized in that Including: Performing edge detection and lightweight semantic segmentation on the input image in parallel to generate an edge intensity map and a semantic confidence map; Generating a privacy grading mask by dynamically fusing the edge intensity map and the semantic confidence map; Dividing the image into three levels of privacy regions through the privacy grading mask; Performing hierarchical encryption on the three levels of privacy regions through three pre-generated aperiodic chaotic sequences to complete user privacy protection; Among them, the three aperiodic chaotic sequences are iteratively generated by the Young's hyperchaotic system based on the hash value obtained by concatenating the user-defined key and the system time.

2. The user privacy protection method for a face recognition system according to claim 1, wherein The generating of the privacy grading mask by dynamically fusing the edge intensity map and the semantic confidence map includes: Based on the mean of local semantic confidence and the standard deviation of edge intensity , calculate the fusion weight of semantic segmentation and edge detection, and the formula is as follows: ； Among them, is the fusion weight of semantic segmentation, is the fusion weight of edge detection; Generate a fused mask through a weighted formula, where is the semantic confidence map, and is the edge intensity map. ​ 3. The user privacy protection method for a face recognition system according to claim 2, wherein The dividing of the image into three levels of privacy regions through the privacy grading mask includes: According to the fusion mask Based on the falling threshold range, the image is divided into three levels of privacy regions: the facial feature region, the facial contour region, and the background region.

4. The user privacy protection method for a face recognition system according to claim 3, characterized in that, The generating method of the three aperiodic chaotic sequences includes: Obtaining a hash value by concatenating the user-defined key and the system time, and the formula is as follows: ； Among them, is the user input key, is the system time, is the hash value after hash processing, represents the hash function, which is used to perform hash transformation on the input key and output a 256-bit hash sequence; The hash value after hash processing , is input into the following Young's hyperchaotic system for iteration: ； Among them, is the control variable of the system, , , are the input values for each iteration, , , are the output values for each iteration; Finally, all the elements of the iterations form three aperiodic chaotic sequences.

5. The user privacy protection method for a face recognition system according to claim 4, wherein The hierarchical encryption of the three levels of privacy regions through the three pre-generated aperiodic chaotic sequences includes: Adopting a stream key diffusion algorithm for the background region, dynamically generating a stream key through the chaotic sequence and performing diffusion encryption on all the pixels of the image; Adopting an improved Arnold transform for the face contour region, and performing coordinate scrambling encryption on the pixel blocks by combining high-frequency trigonometric functions; Adopting a channel-coupled chaotic scrambling algorithm for the face feature region, and encrypting by coupling the pixel values of the RGB channels through the chaotic sequence.

6. The user privacy protection method for a face recognition system according to claim 5, wherein, The adopting of the stream key diffusion algorithm for the background region, dynamically generating a stream key through the chaotic sequence and performing diffusion encryption on all the pixels of the image includes: Based on the elements of the chaotic sequence Generate a stream key , and the formula is as follows: ； Among them sin(10 ) ; Through the formula perform diffusion encryption on all image pixels; Among them, represents the value after diffusion of the -th pixel, respectively represent the -th and -th elements in the stream key sequence, is the original pixel.

7. The user privacy protection method for a face recognition system according to claim 6, characterized in that, The adopting of the improved Arnold transform for the face contour region, and performing coordinate scrambling encryption on the pixel blocks by combining high-frequency trigonometric functions includes: Dividing the face contour region into 32×32 pixel blocks, through the dynamic parameter formula ； Perform non-linear scrambling encryption on pixel coordinates, where, represents the pixel block coordinates after transformation, is the original block coordinates, is the control parameter value, is the width and height of the image.

8. The user privacy protection method for a face recognition system according to claim 7, characterized in that, The adopting of the channel-coupled chaotic scrambling algorithm for the face feature region, and encrypting by coupling the pixel values of the RGB channels through the chaotic sequence, and the formula is as follows: ； Among them, among them represents the three color channels of the picture, while represents the three channels after transformation.

9. A user privacy protection device for a face recognition system, characterized in that, Including: A memory for storing computer programs / instructions; A processor for executing the computer programs / instructions to implement the steps of the method according to any one of claims 1-8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by the processor, it implements the steps of the method according to any one of claims 1-8.

Images