A key-driven dual-domain hybrid revocable face privacy protection method and device

Abstract

The application discloses a kind of key-driven dual-domain hybrid revocable face privacy protection method and device, method is derived from master key by key management module for different purposes Subkey, and support versioned revocation and reissue;In image domain, face image is mapped to latent space using inversion network, protected image is generated based on key hierarchical gating mixing and noise disturbance;In feature domain, revocable binary template is generated by key scrambling, orthogonal projection and grouped random threshold quantization, and correction template is obtained by lightweight error correction coding;Cross-domain alignment module ensures that image domain embedded features and feature domain templates are semantically consistent under the condition of "same person same key";Privacy level control module dynamically adjusts anonymity boundary and noise intensity to achieve multi-level protection;Process control module manages the whole process of registration, verification, revocation and reissue.The application balances identity recognition utility and privacy security, and has irreversibility, unlinkability and revocability.

Description

Technical Field

[0001] This invention relates to the field of biometric information security and privacy computing, and in particular to a key-driven dual-domain hybrid revocable face privacy protection method and apparatus. Background Technology

[0002] With the rapid development of artificial intelligence and deep learning technologies, biometric recognition has been widely applied in various fields such as identity authentication, access control systems, financial payments, and mobile terminals. Among them, facial recognition has become the mainstream solution due to its contactless nature and high convenience. However, facial data is highly unique and immutable. Once the original facial image or feature template is stolen during transmission, storage, or training, it will cause serious privacy leaks and identity abuse risks. Traditional encryption and anonymization methods (such as homomorphic encryption, differential privacy, or image blurring) can alleviate the leakage problem to some extent, but they often lead to excessively high computational costs or significant degradation in recognition performance, making it difficult to promote in practical facial recognition systems. To address this, the concept of Cancelable Biometrics has been proposed. This involves generating revocable templates through key-driven feature transformations. When a template is leaked, a new template can be quickly reissued by changing the key without replacing the biological entity itself, thereby achieving a secure and manageable identity authentication mechanism.

[0003] In recent years, some studies have attempted to achieve revocable protection in the feature domain, such as generating binary templates through random projection, hash encoding, or biohashing. However, these methods typically ignore the semantic consistency between the image domain and the feature domain, easily leading to the loss of identity features or unstable cross-domain matching. Furthermore, single-domain transformations struggle to balance visual anonymity and recognition utility, failing to meet the privacy protection needs of various real-world scenarios. To address these issues, a unified framework capable of achieving dual protection in both the image and feature domains is urgently needed, maintaining original recognition performance while enabling secure revocation and redistribution in the event of template leakage. Summary of the Invention

[0004] The purpose of this invention is to provide a key-driven dual-domain hybrid revocable face privacy protection method and apparatus. By introducing a key conditional hybrid mechanism and random hash transformation of the template domain in the latent space, cross-domain consistent face privacy protection is achieved, effectively balancing identity utility and privacy security.

[0005] The technical solution adopted in this invention is:

[0006] A key-driven, dual-domain hybrid revocable face privacy protection device, comprising the following modules:

[0007] The key management module includes a derivation unit and a version management unit. The derivation unit uses the HKDF / HMAC algorithm, combined with the purpose tag and version number, to derive independent subkeys (k_mix, k_hash, k_noise). Specifically: k_mix is ​​the image domain latent space hierarchical mixing control key, used to drive the generation of synthetic latent codes and the gated mixing process; k_hash is the feature domain template transformation key, used to control the scrambling, orthogonal projection, and group quantization of identity features to generate revocable binary templates; k_noise is the noise and privacy level control key, used to generate reproducible random perturbations and as the control source for level parameters such as anonymity boundaries and noise intensity. The version management unit maintains the key version number ver and its lifecycle, enabling key revocation, rotation, and redistribution.

[0008] The image domain protection module is used to implement key-based conditional hierarchical gating mixing in the latent space of the generative model to generate a protected image from the input image.

[0009] The feature domain template protection module is used to convert facial features into a revocable binary template C, and achieve secure matching through key-driven scrambling and quantization operations.

[0010] The cross-domain alignment module is used to establish a consistent mapping between the image domain and the feature domain, ensuring that the protected image embedding and the template binary code are semantically aligned.

[0011] The privacy level control module is used to dynamically issue privacy protection configurations based on application scenarios, achieving an adaptive balance between privacy and usability;

[0012] The process control module is used to control the logical operations of the entire process of registration, authentication, template revocation and reissue of protected images and binary templates.

[0013] Furthermore, the image domain protection module includes:

[0014] The inversion and key mapping unit is used to map the input image to the latent space and generate conditional latent features by combining key information;

[0015] The hierarchical mixing generation unit is used to calculate the mixed latent vector in each layer of the latent space based on key-controlled hierarchical gated mixing and noise perturbation, using the following formula:

[0016] ;

[0017] in, for No. Hybrid latent vectors of the layers; For the first Layer hybrid weights, For the first Residual calibration terms of the layer, For the first Layer noise; for No. The latent space vector of the layer; for No. Layered synthetic latent codes;

[0018] The decoding and distillation inference unit is used to decode the mixed latent vectors into protected images using the generator StyleGAN3, and to achieve efficient single-step forward inference through distillation training.

[0019] Specifically, during the training phase, teacher latent codes are introduced. As a goal, a knowledge distillation strategy is employed to train the student network to learn a forward generative mapping; direct input is then used in the inference phase. You can get ,in Using the original image, no iterative optimization is required, enabling fast and high-fidelity privacy-preserving image generation.

[0020] Furthermore, the feature domain template protection module includes:

[0021] The keying preprocessing unit is used to scramble and flip the identity feature vector extracted by the pre-trained face recognition model based on the key k_hash to obtain normalized keyed features;

[0022] The orthogonal projection compression unit is used to map keyed features to a low-dimensional space through orthogonal transformation and Walsh transformation, so as to select several sub-columns or sub-blocks under key control to obtain orthogonal compressed features;

[0023] The grouped random threshold quantization unit is used to convert orthogonal compressed features into revocable binary templates. It generates stable binarization results through key-driven random thresholds and intra-group weights, and combines this with BCH error-correcting coding to obtain a correction template for the binary template.

[0024] Furthermore, the cross-domain alignment module includes:

[0025] The projection head unit is used to map the identity feature vector of the protected image to a continuous vector space consistent with the template codebook.

[0026] The codebook mapping alignment constraint unit is used to map a binary template into a measurable continuous hypervector and establish an alignment constraint with the projection vector by minimizing the Euclidean distance loss between the hypervector and the projection vector.

[0027] Furthermore, the privacy gradation control module includes:

[0028] The tier generation unit is used to generate privacy tier configuration parameters for each tier based on the input strategy and the latest statistical data, according to rules or lightweight strategies, with the noise key k_noise as the tier control parameter. The privacy tier configuration parameters include anonymity boundary m, noise intensity σ, template code length M, quantization threshold interval and error correction intensity.

[0029] Specifically, the gear generation unit parses the input strategy and the latest statistics, calculates the target gear based on the rule or lightweight strategy engine, and uses the noise key k_noise as an important control source for the gearing parameters to drive the noise injection mode under each gear. Then, it solves a set of self-consistent configurations in the parameter space (e.g., light gear: smaller m / lower σ / shorter M / wider threshold interval / weak error correction; medium gear: medium strength global balance; strong gear: larger m / higher σ / longer M / narrower threshold interval / medium or stronger error correction) and generates a deployable profile (including version number and effective window).

[0030] The gear control unit is used to send privacy gear configurations to the image domain protection module and the feature domain template protection module, and to fine-tune the configuration parameters based on online feedback, so as to write back new calibration parameters and gear versions at set intervals.

[0031] Furthermore, the process control module includes:

[0032] The registration and verification unit is used to process user registration and authentication, and to generate and store templates.

[0033] Specifically, during the registration phase, the system receives the original facial image. With master key First, the image domain protection module is invoked to generate the protected image. Then, the corresponding binary template is generated by the feature domain template protection module. The generated result, along with the key version number (ver), is stored on the server for later use. During the verification phase, the system receives the query image and the current version number, and repeatedly generates the corresponding... and Then, the verification domain is selected according to the system strategy: if it is in the template domain, the Jaccard distance is calculated; if it is in the image domain, the face recognition similarity threshold is used to determine the matching result, and the verification score or pass flag is output. The Jaccard coefficient is:

[0034] ;

[0035] in, , It is a binary template. This indicates the same bit count.

[0036] The revocation and reissue unit is used to trigger a version upgrade when a key is leaked or a policy is updated. It regenerates the core template based on the newly derived subkey and isolates the old version.

[0037] Furthermore, this invention discloses a key-driven dual-domain hybrid revocable face privacy protection method, employing the aforementioned key-driven dual-domain hybrid revocable face privacy protection device, the method comprising the following steps:

[0038] S1, Key Derivation and Version Configuration: Derive subkeys (k_mix, k_hash, k_noise) from the master key k using the HKDF / HMAC algorithm and bind them to the version number ver;

[0039] S2, Image Domain Protected Image Generation: The input image is mapped to the latent space, and layered mixing and noise perturbation are performed based on the key k_mix to generate the protected image. and identity embedding ,in This is a pre-trained Arcface model used to extract image identity features;

[0040] S3, Reversible Template Generation in the Feature Domain: Features are extracted from the image, and scrambling, orthogonal projection, and group quantization are performed based on the key k_hash to generate a binary template C. The binary template C is then encoded using BCH error correction to obtain the correction template. ;

[0041] S4, Cross-domain alignment: Embedding identity Input projection head Obtain the projection vector , will the correction template Input codebook mapping head Obtain the supervector By minimizing the Euclidean distance loss between the hypervector and the projection vector, an alignment constraint is established with the projection vector to ensure that the image domain embedding and feature domain template semantics are consistent under the same person and key.

[0042] S5, Registration and Verification Process: During the registration phase, the protected image and its corresponding binary template are stored. During the verification phase, the query object is matched with the object stored in the database using Jaccard similarity or identity similarity.

[0043] S6, Privacy Level Control: Dynamically adjust privacy parameters to achieve different levels of protection;

[0044] S7, Revocation and Reissue: Update the version and regenerate a new template when the key is leaked, ensuring that the old and new templates cannot be linked.

[0045] Further, step S2 includes:

[0046] S2-1, The input image is processed through an inversion network. Mapping to latent space representation ;

[0047] S2-2, Generating synthetic latent codes based on k_mix Mixed weights α and noise ε;

[0048] S2-3, calculate the mixed latent vector at each layer, using the following formula:

[0049] ;

[0050] in, for No. Hybrid latent vectors of the layers; For the first Layer hybrid weights, For the first Residual calibration terms of the layer, For the first Layer noise; for No. The latent space vector of the layer; for No. Layered synthetic latent codes;

[0051] S2-4 decodes the hybrid latent vector input generator to generate the protected image. ;

[0052] S2-5, inject controllable noise into the latent space or image domain based on the key k_noise;

[0053] S2-6, Output protected image and identity embedding .

[0054] Further, step S3 includes:

[0055] S3-1, Extract identity feature vector from input image or protected image. ;

[0056] S3-2, Extracting identity feature vectors based on k_hash pairs By performing random scrambling and symbol flipping, the keyed features are obtained. ;

[0057] S3-3, Keying Features Apply an orthogonal transformation and in the key Under control, several sub-columns or sub-blocks are selected to obtain orthogonal compression features;

[0058] S3-4, the orthogonal compression features are grouped into fixed-size groups, and the weight vector of each group is derived from the key k_hash. and non-zero threshold , so as to quantize each group to generate a binary template C;

[0059] S3-5, perform BCH error correction coding on the binary template C to obtain the correction template. And bind the version number ver.

[0060] Furthermore, S6 includes the following steps:

[0061] S6-1, Receive the query image and the current version ver, (execute S2 and S3) respectively to calculate the protected image of the query image. Identity Embedding And new calibration templates ;

[0062] S6-2, Calculate the new correction template Correction templates in the database The Jaccard similarity is calculated and compared with a preset first threshold, and the judgment result is output.

[0063] S6-3, Calculate the new identity embedding Identity embedding in database The identity similarity is calculated and compared with a preset second threshold to output the judgment result;

[0064] S6-4. Based on the strategy, the final certification conclusion is obtained by selecting the best or merging the results from the two domains.

[0065] Furthermore, in S5, the user registration request and the current version ver are received; steps S2 and S3 are executed to generate... With correction template ; Correction template and The version number, along with the existing version, is stored in the template library to form a registration entry.

[0066] Furthermore, in S7, ver+1 is triggered when the key is leaked / expired or the policy is upgraded; subkeys are re-derived according to S1 and new templates are generated in batches or lazily according to S3 and written to the new version library; the old version library is set to read-only and cross-library comparison is prohibited, so that the old / new templates are not linked and are traceable.

[0067] This invention employs the above technical solutions, specifically targeting facial image data, to generate protected facial representations that simultaneously enhance privacy and identify identities. While effectively identifying users, it effectively inhibits attackers from inferring soft biometric attributes (such as gender, age, and race) from protected images or feature templates, thereby preventing unauthorized disclosure of sensitive user characteristics. By introducing a key-driven dual-domain hybrid mechanism in the image and feature domains, this invention achieves dual-layer revocable protection of facial data: the image domain achieves controllable anonymization and visual consistency through key-conditional layered hybridization, while the feature domain generates revocable binary templates through keyed scrambling and orthogonal quantization, achieving a synergistic balance between identity utility and privacy. This method significantly improves template security and version manageability while maintaining facial recognition accuracy and cross-domain matching stability. This invention can be widely applied to the security protection of biometric identification systems such as facial recognition, identity verification, and privacy computing. It is particularly suitable for intelligent terminals and cloud authentication systems that transmit, compare, and store facial data in open or semi-trusted environments, effectively enhancing the overall privacy protection capabilities and controllable revocability of the system. Attached Figure Description

[0068] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments;

[0069] Figure 1 This is a schematic diagram of the structure of a key-driven dual-domain hybrid revocable face privacy protection device according to the present invention;

[0070] Figure 2 This is a flowchart illustrating a key-driven, dual-domain hybrid revocable face privacy protection method according to the present invention. Detailed Implementation

[0071] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings.

[0072] like Figure 1 As shown in Figure 2, this invention discloses a key-driven dual-domain hybrid revocable face privacy protection device, comprising:

[0073] The key management module 110 is used to derive subkeys (k_mix, k_hash, k_noise) for different purposes from the master key k, and to realize the revocation, rotation and redistribution of keys through a version management mechanism, thereby ensuring the independence, traceability and controllable access of each key in the system.

[0074] The image domain protection module 120 is used to implement key-based conditional hierarchical gated mixing in the latent space of the generative model, generating a protected image from the input image in a single forward pass. By controlling the mixing ratio and noise intensity of the latent vectors, it achieves perturbation of facial identity and privacy protection, while maintaining visual naturalness and a controllable level of anonymization.

[0075] The feature domain template protection module 130 is used to convert facial features into a revocable binary template. This enables efficient matching and secure auditing. Through key-driven scrambling and grouping quantization operations, the template maintains identity discriminativeness while possessing irreversibility and revocability, making it suitable for scenarios requiring rapid identity verification and privacy protection.

[0076] Cross-domain alignment module 140 is used to establish a consistent mapping between the image domain and the feature domain, ensuring that the embedded representation of the protected image matches the corresponding template binary code under the "same person, same key" condition. Semantic alignment reduces the false rejection rate (FRR) caused by cross-domain differences, thereby improving the stability and reliability of the system's authentication.

[0077] The privacy level control module 150 is used to dynamically issue light, medium or strong privacy protection configurations according to different application scenarios and policy requirements, so as to achieve an adaptive balance between privacy and usability in the system.

[0078] The process control module 160 is used for the logical control of the entire process of registration, identity verification, template revocation, and reissue, ensuring the security, consistency, and traceability of the system in multi-stage operations. Through modular process scheduling, it effectively connects sub-modules such as key management, image domain protection, and feature template protection to achieve closed-loop management from registration to verification to revocation.

[0079] In a specific embodiment, the key management module 110 includes a derivation unit 111 and a version management unit 112;

[0080] Derivation unit 111 is used to derive independent subkeys (k_mix, k_hash, k_noise) for different purposes from the master key k. It is generated by combining the purpose tag and version number with the HKDF / HMAC algorithm to ensure that the keys of each functional domain are independent and cannot be mutually derived, thus avoiding cross-domain reuse and association risks.

[0081] Version management unit 112 is used to maintain the key version number (ver) and its lifecycle, and supports key revocation and redistribution. When the version (ver) is updated, the system automatically regenerates a new key and archives the old version, ensuring that different versions are not linked and are traceable.

[0082] In a specific embodiment, the image domain protection module 120 includes an inversion and key mapping unit 121, a hierarchical mixing generation unit 122, and a decoding and distillation inference unit 123.

[0083] Inversion and key mapping unit 121 is used to convert the input image Mapped to the latent space, and conditional latent features generated by combining key information, this provides a basis for subsequent mixing and perturbation, fusing identity representation and key features in the latent layer to form a controllable source of privacy perturbation;

[0084] The hierarchical mixing generation unit 122 is used to realize key-controlled hierarchical gating mixing and noise perturbation to generate latent vectors that correspond to the original image but have been enhanced for privacy. While maintaining the image structure and identity recognizability, privacy ambiguity and differential features are introduced through multi-layer modulation.

[0085] The decoding and distillation inference unit 123 is used to restore the mixed latent vectors to the protected image and achieve efficient single-step forward inference through distillation training, ensuring the naturalness, consistency and inference speed of the generated image.

[0086] In a specific embodiment, the feature domain template protection module 130 includes a keying preprocessing unit 131, an orthogonal projection compression unit 132, and a grouped random threshold quantization unit 133.

[0087] The keying preprocessing unit 131 is used to extract stable identity feature vectors from the input image or the protected image, and scramble and flip them based on the key to overturn the distribution structure of the original features, weaken the statistical correlation between fixed dimensions, and realize feature randomization and key binding.

[0088] The orthogonal projection compression unit 132 maps the keyed features to an approximately equidistant low-dimensional subspace through orthogonal transformation Walsh transformation, and performs sub-column or sub-block selection under key control, thereby obtaining an intermediate representation that retains the main identity information and has secure compression characteristics.

[0089] The grouped random threshold quantization unit 133 is used to convert orthogonal compressed features into revocable binary templates. It generates stable binary results through key-driven random thresholds and intra-group weights, and combines lightweight error correction coding to improve the consistency and fault tolerance of the templates.

[0090] In a specific embodiment, the cross-domain alignment module 140 includes a projection head unit 141 and a codebook mapping alignment constraint unit 142:

[0091] The projection head unit 141 embeds the identity of the protected image into a representation space consistent with the template codebook, eliminating inter-domain distribution differences and providing a comparable continuous vector representation for cross-domain metrics.

[0092] Codebook mapping alignment constraint unit 142, which sets the binary template Mapped to measurable continuous hypervectors and with Establish alignment constraints to ensure consistency between the two domains under the condition of "same person, same key", thereby reducing false rejections across domains.

[0093] In a specific embodiment, the privacy gear control module 150 includes a gear generation unit 151 and a gear control unit 152:

[0094] The privacy profile generation unit 151 generates standardized privacy profile configurations, covering the boundaries of anonymity. Noise intensity Template code length By incorporating key parameters such as quantification threshold intervals and error correction strength, a controllable trade-off between privacy and utility can be achieved.

[0095] The gear control unit 152 uniformly distributes the generated privacy gears to the image domain protection module and the feature domain template protection module, and calibrates the threshold and intensity through online feedback closed-loop calibration.

[0096] In a specific embodiment, the process control module 160 includes a registration and verification unit 161 and a revocation and reissue unit 162:

[0097] The 161 Registration and Verification Unit is used for new user registration and authentication, ensuring that the user template generation, storage, and matching process is secure, controllable, and consistent.

[0098] Unit 162, Revocation and Reissue, is used for version management of keys and templates. It ensures that the system can securely revoke old templates and generate new templates after key leakage or policy updates, thereby achieving template revocation and version isolation.

[0099] This device can be used by designers of facial recognition systems to set privacy trade-offs to train the network based on different privacy protection levels or relevant laws and regulations. This device achieves dual-layer revocable protection of facial data: in the image domain, controlled anonymization and visual consistency are achieved through key-conditional layered mixing; in the feature domain, revocable binary templates are generated through key-based scrambling and orthogonal quantization, achieving a synergistic balance between identity utility and privacy. This method significantly improves template security and version manageability while maintaining facial recognition accuracy and cross-domain matching stability. This invention can be widely applied to the security protection of biometric identification systems such as facial recognition, identity verification, and privacy computing. It is particularly suitable for smart terminals and cloud authentication systems that transmit, compare, and store facial data in open or semi-trusted environments, effectively improving the overall privacy protection capabilities and controllable revocability of the system.

[0100] Furthermore, the present invention also discloses a key-driven dual-domain hybrid revocable face privacy protection method, which includes the following steps:

[0101] S1, Key Derivation and Version Configuration;

[0102] S2, Image domain protected image generation;

[0103] S3, Feature domain reversible template generation;

[0104] S4, cross-domain alignment;

[0105] S5, Registration and Verification Process;

[0106] S6, privacy settings control;

[0107] S7, Withdrawal and Reissue;

[0108] Furthermore, S1 includes the following steps:

[0109] S1-1, Receive master key k, purpose label (k_mix, k_hash, k_noise) and version number ver;

[0110] S1-2, key derivation is performed using the HKDF / HMAC algorithm, i.e.:

[0111] ;

[0112] S1-3: Register the key derivation parameters and version mapping relationship, and enable access control policies to restrict keys from leaving the domain; when the policy triggering conditions are met, execute... Updated to support revocation and reissue.

[0113] Furthermore, S2 includes the following steps:

[0114] S2-1, Input face image Normalization is performed; the ReStyle inversion network is used to... Mapping to latent space representation ;

[0115] S2-2, based on Synthetic latent codes are generated by a key mapping network. Layered mixed weights and hierarchical decoupling noise ;

[0116] S2-3, in each layer ,calculate: ;

[0117] S2-4, Decoding and Generation: [The sentence is incomplete and requires more context to be translated.] The input generator produces a protected image. ;

[0118] S2-5, based on the noise subkey Lightweight random perturbations are injected into the latent space or image domain to adjust the privacy level and differential strength. This key controls the distribution characteristics of the noise (such as variance, spatial correlation, and random source selection), making the perturbation reproducible under the same person and key, and the perturbation pattern unpredictable between different keys, thereby achieving controllable privacy enhancement.

[0119] S2-6, Output and its identity embedding

[0120] Furthermore, S3 includes the following steps:

[0121] S3-1, from or Extract identity feature vectors ;

[0122] S3-2, Keying Preprocessing: Based on right Perform random scrambling and symbol flipping to obtain keyed features. ;

[0123] S3-3, for Apply the orthogonal transformation Walsh, and in Under control, sub-column / sub-block selection is performed to obtain an intermediate representation. ;

[0124] S3-4, Divide the data into several groups of a fixed size, with each group containing a weight vector. and non-zero threshold The key k_hash is derived and then the weighted inner product of each group is calculated and compared with a threshold. If the result is greater than zero, 1 is output; otherwise, 0 is output. Finally, the results are concatenated to form a complete binary template. ;

[0125] S3-5 pairs Lightweight error correction coding (BCH) is performed to obtain the correction template. And bind the version number ver;

[0126] Furthermore, S4 includes the following steps:

[0127] S4-1, Projection Alignment: ... Input projection head get ;

[0128] S4-2, Codebook Mapping: ... Input codebook mapping head get ;

[0129] S4-3, Minimize This ensures that the image domain and feature domain representations are consistent under the condition of "same person, same key".

[0130] Furthermore, S5 includes the following steps:

[0131] S5-1, receives user registration requests and the current version (ver);

[0132] S5-2, Execute steps S2 and S3 to generate and ;

[0133] S5-3, and The verb, along with the filename, is stored in the template library to form a registration entry.

[0134] Furthermore, S6 includes the following steps:

[0135] S6-1, Receive the query image and the current version ver, execute S2 and S3 to obtain... and ;

[0136] S6-2, Template Domain Alignment: Calculate the new... With the library The Jaccard similarity is compared with the threshold to output the judgment;

[0137] S6-3, Image Domain Alignment: Calculating New With the library The identity similarity is compared with a threshold, and the result is output.

[0138] S6-4. Based on the strategy, the final certification conclusion is obtained by selecting the best or merging the results from the two domains.

[0139] Furthermore, S7 includes the following steps: S7-1, triggering ver+1 when the key is leaked / expired or the policy is upgraded;

[0140] S7-2, re-derive the subkey according to S1 and generate new templates for existing objects in batches or lazily according to S3, and write them to the new version library;

[0141] S7-3 sets the old version library to read-only and prohibits cross-library comparison, thus making the old / new templates unlinkable and traceable.

[0142] Example 2, see Figure 2This embodiment proposes a revocable face privacy protection method based on dual-domain protection. During registration, the system derives multiple subkeys from a master key to protect the image and feature domains, and generates and stores protected images and binary templates. During verification, the system compares the query image with the protected images and templates in the database and returns the verification result. If a template needs to be revoked, the system generates a new subkey and updates the template, while ensuring that the old and new templates are not associated, thereby protecting user privacy.

[0143] This invention employs the aforementioned technical solution, specifically targeting facial image data, to generate protected facial representations that simultaneously enhance privacy and identify identities. While effectively identifying users, it effectively inhibits attackers from inferring soft biometric attributes (such as gender, age, and race) from protected images or feature templates, thereby preventing unauthorized disclosure of sensitive user characteristics. By introducing a key-driven dual-domain hybrid mechanism in the image and feature domains, this invention achieves dual-layer revocable protection of facial data: the image domain achieves controllable anonymization and visual consistency through key-conditional layered hybridization, while the feature domain generates revocable binary templates through keyed scrambling and orthogonal quantization, achieving a synergistic balance between identity utility and privacy. This method significantly improves template security and version manageability while maintaining facial recognition accuracy and cross-domain matching stability. This invention can be widely applied to the security protection of biometric identification systems such as facial recognition, identity verification, and privacy computing. It is particularly suitable for intelligent terminals and cloud authentication systems that transmit, compare, and store facial data in open or semi-trusted environments, effectively enhancing the overall privacy protection capabilities and controllable revocability of the system.

[0144] Obviously, the described embodiments are only a part of the embodiments of this application, not all of them. Without conflict, the embodiments and features in the embodiments of this application can be combined with each other. The components of the embodiments of this application described and illustrated herein can generally be arranged and designed in various different configurations. Therefore, the detailed description of the embodiments of this application is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

Claims

1. A key-driven, dual-domain hybrid revocable face privacy protection device, characterized in that: It includes the following modules: The key management module includes a derivation unit and a version management unit. The derivation unit is used to derive independent subkeys k_mix, k_hash, and k_noise from the master key k using the HKDF or HMAC algorithm combined with the purpose tag and version number. The version management unit maintains the key version number ver and its lifecycle, and realizes key revocation, rotation, and redistribution. Among them: k_mix is ​​the image domain latent space hierarchical mixing control key, used to drive the generation of synthetic latent codes and the gated mixing process; k_hash is the feature domain template transformation key, used to control the scrambling, orthogonal projection, and group quantization of identity features to generate revocable binary templates; k_noise is the noise and privacy level control key, used to generate reproducible random perturbations and as the control source for anonymity boundaries and noise intensity level parameters. The image domain protection module is used to implement key-based conditional hierarchical gating mixing in the latent space of the generative model to generate a protected image from the input image. The feature domain template protection module is used to convert facial features into a revocable binary template C, and achieve secure matching through key-driven scrambling and quantization operations. The cross-domain alignment module is used to establish a consistent mapping between the image domain and the feature domain, ensuring that the protected image embedding and the template binary code are semantically aligned; the privacy level control module is used to dynamically issue privacy protection configurations according to the application scenario, achieving an adaptive balance between privacy and usability. The process control module is used to control the logical operations of the entire process of registration, authentication, template revocation and reissue of protected images and binary templates.

2. The key-driven dual-domain hybrid revocable face privacy protection device according to claim 1, characterized in that: The image domain protection module includes: The inversion and key mapping unit is used to map the input image to the latent space and generate conditional latent features by combining key information; The hierarchical mixing generation unit is used to calculate the mixed latent vector in each layer of the latent space based on key-controlled hierarchical gated mixing and noise perturbation, as shown in the formula: ； in, For the first Hybrid latent vectors of the layers; For the first Layer hybrid weights, For the first Residual calibration terms of the layer, For the first Layer noise; For the first The latent space vector of the layer; For the first Layered synthetic latent codes; The decoding and distillation inference unit is used to decode the mixed latent vectors into a protected image using the generator StyleGAN3, while simultaneously generating identity embeddings. ,in This is a pre-trained Arcface model used to extract identity features from images.

3. The key-driven dual-domain hybrid revocable face privacy protection device according to claim 1, characterized in that: The feature domain template protection module includes: The keying preprocessing unit is used to scramble and flip the identity feature vector extracted by the pre-trained face recognition model based on the key k_hash to obtain normalized keyed features; The orthogonal projection compression unit is used to map keyed features to a low-dimensional space through orthogonal transformation and Walsh transformation, so as to select several sub-columns or sub-blocks under key control to obtain orthogonal compressed features; The grouped random threshold quantization unit is used to convert orthogonal compressed features into revocable binary templates, generate stable binary results through key-driven random thresholds and intra-group weights, and obtain the correction template of the binary template by combining BCH error correction coding.

4. The key-driven dual-domain hybrid revocable face privacy protection device according to claim 1, characterized in that: The cross-domain alignment module includes: The projection head unit is used to map the identity feature vector of the protected image to a continuous vector space consistent with the template codebook. The codebook mapping alignment constraint unit is used to map a binary template into a measurable continuous hypervector and establish an alignment constraint with the projection vector by minimizing the Euclidean distance loss between the hypervector and the projection vector.

5. The key-driven dual-domain hybrid revocable face privacy protection device according to claim 1, characterized in that: The privacy settings control module includes: The tier generation unit is used to generate privacy tier configuration parameters for each tier based on the input strategy and the latest statistical data, according to rules or lightweight strategies, with the noise key k_noise as the tier control parameter. The privacy tier configuration parameters include anonymity boundary m, noise intensity σ, template code length M, quantization threshold interval and error correction intensity. The gear control unit is used to send privacy gear configurations to the image domain protection module and the feature domain template protection module, and to fine-tune the configuration parameters based on online feedback, so as to write back new calibration parameters and gear versions at set intervals.

6. The key-driven dual-domain hybrid revocable face privacy protection device according to claim 1, characterized in that: The process control module includes: The registration and verification unit is used to process user registration and authentication, generate and store templates; the registration phase is based on the original facial image. With master key Generate protected images The corresponding binary template C, along with the key version number ver, is stored on the server. During the verification phase, a new protected image is repeatedly generated based on the queried image and the current version number. With correction template In the image domain or feature domain, the matching of queried images with objects stored in the server database is achieved through Jaccard similarity or identity similarity. The revocation and reissue unit is used to trigger a version upgrade when a key is leaked or a policy is updated. It regenerates the core template based on the newly derived subkey and isolates the old version.

7. A key-driven dual-domain hybrid revocable face privacy protection method, employing the key-driven dual-domain hybrid revocable face privacy protection device according to any one of claims 1 to 6, characterized in that: The method includes the following steps: S1, Key Derivation and Version Configuration: Derive subkeys (k_mix, k_hash, k_noise) from the master key k using the HKDF / HMAC algorithm and bind them to the version number ver; S2, Image Domain Protected Image Generation: The input image is mapped to the latent space, and layered mixing and noise perturbation are performed based on the key k_mix to generate the protected image. and identity embedding ,in This is a pre-trained Arcface model used to extract identity features from images; S3, Reversible Template Generation in the Feature Domain: Features are extracted from the image, and scrambling, orthogonal projection, and group quantization are performed based on the key k_hash to generate a binary template C. The binary template C is then encoded using BCH error correction to obtain the correction template. ; S4, Cross-domain alignment: Embedding identity Input projection head Obtain the projection vector , will the correction template Input codebook mapping head Obtain vector By minimizing the Euclidean distance loss between the hypervector and the projection vector, an alignment constraint is established with the projection vector to ensure that the image domain embedding and feature domain template semantics are consistent under the same person and key. S5, Registration and Verification Process: During the registration phase, the protected image and its corresponding binary template are stored. During the verification phase, the image is matched with the objects stored in the server database in the image domain or feature domain using Jaccard similarity or identity similarity. S6, Privacy Level Control: Dynamically adjust privacy parameters to achieve different levels of protection; S7, Revocation and Reissue: Update the version and regenerate a new template when the key is leaked, ensuring that the old and new templates cannot be linked.

8. The key-driven dual-domain hybrid revocable face privacy protection method according to claim 7, characterized in that: Step S2 includes: S2-1, The input image is processed through an inversion network. Mapping to latent space representation ; S2-2, Generating synthetic latent codes based on k_mix Mixed weights α and noise ε; S2-3, calculate the mixed latent vector at each layer, using the following formula: ； in, For the first Hybrid latent vectors of the layers; For the first Layer hybrid weights, For the first Residual calibration terms of the layer, For the first Layer noise; For the first The latent space vector of the layer; For the first Layered synthetic latent codes; S2-4 decodes the hybrid latent vector input generator to generate the protected image. ; S2-5, inject controllable noise into the latent space or image domain based on the key k_noise; S2-6, Output protected image and identity embedding .

9. The key-driven dual-domain hybrid revocable face privacy protection method according to claim 8, characterized in that: Step S3 includes: S3-1, Extract identity feature vector from input image or protected image. ; S3-2, Extracting identity feature vectors based on k_hash pairs By performing random scrambling and symbol flipping, the keyed features are obtained. ; S3-3, Keying Features Apply an orthogonal transformation and in the key Under control, several sub-columns or sub-blocks are selected to obtain orthogonal compression features; S3-4, the orthogonal compression features are grouped into fixed-size groups, and the weight vector of each group is derived from the key k_hash. and non-zero threshold , so as to quantize each group to generate a binary template C; S3-5, perform BCH error correction coding on the binary template C to obtain the correction template. And bind the version number ver.

10. The key-driven dual-domain hybrid revocable face privacy protection method according to claim 7, characterized in that: S6 includes the following steps: S6-1, Receive the queried image and calculate the protected image of the queried image using the current version ver. Identity Embedding And new calibration templates ; S6-2, Calculate the new correction template Correction templates in the database The Jaccard similarity is calculated and compared with a preset first threshold, and the judgment result is output. S6-3, Calculate the new identity embedding Identity embedding in database The identity similarity is calculated and compared with a preset second threshold to output the judgment result; S6-4. Based on the strategy, the final certification conclusion is obtained by selecting the best or merging the results from the two domains.

Images