Face revocable template protection method based on deep hash

Abstract

The invention relates to the field of biological feature recognition and information security, in particular to a face revocable template protection method based on deep hash. The method comprises the following steps: detecting and aligning an input face image, and extracting and normalizing features; performing projection and symbol quantization to obtain a binary vector with a fixed length; deriving a dimension permutation and index subset by using a user key, and generating a protected template; registration and identification are completed in a protected domain by a Hamming distance; when the template is leaked or needs to be replaced, a new template is regenerated by updating a secret key or an index rule, and template revocation and non-linkability are achieved. The system side only stores the protected template and necessary verification parameters and does not store plaintext images and unprotected features, so that the method is low in calculation overhead, convenient to deploy and suitable for face identity authentication of end-side equipment and a cloud platform.

Description

Technical Field

[0001] This invention relates to the field of biometric identification and information security, specifically to a method for protecting revocable face templates based on deep hashing. Background Technology

[0002] Facial recognition technology is widely deployed in applications such as financial risk control, access control and attendance, and mobile device unlocking. Traditional systems often store real-valued facial features directly as templates, which are difficult to replace once leaked, resulting in insufficient privacy, security, and revocability issues.

[0003] Cancelable biometrics are designed to perform comparisons within a protected domain through irreversible transformations, allowing leaked templates to be invalidated and resent by updating the key or transforming parameters, thereby achieving the goal of being "revocable, unrecoverable, and unlinkable".

[0004] In existing methods, striking a balance between template revocability and recognition performance remains a challenge.

[0005] First, information loss after template quantization leads to a decrease in distinguishability.

[0006] Secondly, key-driven random mappings, if poorly designed, can easily introduce linkability risks.

[0007] Third, if the system stores plaintext features or reversible parameters, there is a risk of them being restored.

[0008] Therefore, there is an urgent need for a face template protection method that balances robustness, revocability, and non-linkability without preserving plaintext features. Summary of the Invention

[0009] The purpose of this invention is to provide a face revocable template protection method based on deep hashing. The method completes the comparison within the protection domain through feature extraction, symbol quantization and key derivation dimension substitution / selection, supports template revocation and retransmission, and avoids storing plaintext images and unprotected features.

[0010] To achieve the above objectives, the present invention provides the following technical solution: a method for revocable face template protection based on deep hashing, comprising:

[0011] Perform face detection and alignment preprocessing on the face image to be processed;

[0012] A normalized face feature vector is obtained by forward inference of the image using a feature extraction network;

[0013] The feature vector is symbolically quantized according to a preset projection mapping to obtain a binary vector of length B;

[0014] Based on key-derived dimension index selection and dimension permutation, binary vectors are rearranged and subsets are selected to generate protected face templates.

[0015] Calculate the Hamming distance between protected templates within the protected domain and make a comparison judgment based on a threshold.

[0016] When a template is revoked, the key or index selection rules are updated to regenerate a new protected template for the same identity, thereby achieving template revocation and unlinkability.

[0017] The system only stores protected templates and key derivation parameters, and does not save plaintext images and unprotected features.

[0018] Furthermore, during the training phase, quality-adaptive margin constraints based on sample quality scores can be introduced to improve feature discriminative power.

[0019] Error correction coding can be used before / after template generation to enhance robustness in bit-flipping scenarios.

[0020] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0021] It enables template-level comparison without exposing plaintext features, supports template revocation and resending, and reduces the risk of link attacks;

[0022] Security is enhanced through key-derived permutations and index selection, while maintaining good recognition performance and deployability in the binary domain. Attached Figure Description

[0023] To more intuitively describe the technical solution of the present invention and clearly demonstrate the technical flow of this solution, the accompanying drawings involved in the description of the technical solution are briefly introduced below. The accompanying drawings are only used to illustrate a brief flow of the technical solution and are not intended to limit this application.

[0024] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0025] This invention targets facial recognition applications. Through deep feature extraction, template binarization, and key-based bit-level protection, it completes identity registration, comparison, and revocation within a protected domain. The system only stores the protected template and necessary key verification information, and does not store plaintext facial images or unprotected facial features, thereby reducing the risk of leakage and achieving non-linkability.

[0026] Preprocessing and Feature Extraction: The input face image first undergoes face detection, keypoint alignment, and scale and channel normalization to obtain a standardized face patch. This patch is fed into a finely tuned MobileNetV3-Large backbone network to obtain the face feature vector f. Subsequently, ℓ2 normalization is performed to obtain the unit norm feature f̂ = f / ‖f‖2. During the training phase, a cosine classifier is used to normalize both the feature vector and the class center, and cosθ = ŵᵀ·f̂ is used as the similarity metric. During the inference phase, only the normalized feature vector f̂ is retained for subsequent generation of the protected template.

[0027] Quality-Adaptive Margin Training: To improve the discriminative power of features and robustness to low-quality face samples, a quality score q (normalized from feature norms and other information) is estimated for each sample during training, and the classification margin m is dynamically adjusted based on this quality score. It can be written as m = m_min + (m_max − m_min) · σ(γ · (q − q0)), where m_min and m_max are the upper and lower margin limits, q0 is the quality baseline, γ is the slope, and σ is the logistic function σ(t) = 1 / (1+e^(−t)). High-quality samples receive larger margins, and low-quality samples receive smaller margins. This mechanism is only used during the training phase; q or m does not need to be calculated during the inference phase.

[0028] Binary Template Generation: To facilitate comparison within the protected domain, this invention converts the normalized feature f̂ into a fixed-length binary template. Specifically: the global mean μ of the normalized feature is calculated on the development set; for a single sample, (f̂ − μ) is calculated, and bits are extracted by sign to obtain a vector h, whose elements are only +1 or −1; then +1 is mapped to bit 1, and −1 is mapped to bit 0, resulting in a binary template b of length L. The template length L can be a fixed value for convenient unified storage and fast comparison.

[0029] Protected Template Generation and Revocability: To achieve template revocability and cross-scenario non-linkability, this invention uses a user key k to protect a binary template b. First, b is rearranged according to the bit permutation order generated by key k. Then, it is XORed with a bit string of the same length derived from key k to obtain the final protected template t̃. This can be written as t̃ = P(k)(b) ⊕ K(k), where P(k)(·) represents the bit rearrangement determined by key k, K(k) represents the bitmask derived from the same key, and ⊕ represents the bitwise XOR. Changing key k will generate a completely new t̃, and the old template can be immediately invalidated. Therefore, the template can be revoked and reissued at any time.

[0030] Matching within the protected domain: The system directly compares the protected templates within the protected domain, rather than reconstructing them into real-valued features. Given two protected templates A and B, their normalized Hamming distance is defined as the proportion of bits that differ between them in the total length: D_H(A,B) = (1 / L) × number of differing bits. When D_H(A,B) ≤ γ, they are determined to be the same identity; otherwise, they are determined to be different identities. Hamming distance can be implemented using bitwise XOR plus bitwise counting, which has extremely high speed and is suitable for large-scale, low-latency deployments.

[0031] Optional Mechanism 1: Introduce constraints such as quantization error suppression, bit balancing, and bit decorrelation during the training phase to improve the discriminability and stability of the binary template within the protected domain. Disabling this mechanism does not affect the implementation of the core process of this invention.

[0032] Option 2: Introduce error-correcting coding before or after template protection to improve robustness in scenarios with a small number of bit flips. This mechanism does not change the measurement and determination method of the protected field.

[0033] Implementation Recommendations: The template length B is coupled with the threshold setting. Increasing B generally benefits operation with a low false recognition rate, but increases storage and computational overhead; edge devices preferably use B of 128 or 256. The comparison implementation can use bitwise XOR and bit-counting instructions to improve throughput.

[0034] Security and Compliance: Keys are used only for derivation permutations and indexing rules, and are not persisted in plaintext in insecure environments; the registry does not store reversible recovery paths. The generation, revocation, and version changes of templates may record necessary metadata for auditing purposes, but information that could be used to deduce the original image or unprotected features is not recorded.

[0035] The identification process of this invention is consistent with the registration process: the queried image obtains a protected template according to the above steps, and then compares it with the protected template in the registration database using Hamming distance to make a judgment. The system only stores the protected template and key derivation verification information, and does not store reversible restoration paths; when the template needs to be updated or leakage is suspected, only a new key needs to be generated for the user and the protected template needs to be recalculated, and the old template is marked as invalid at the same time. The above scheme does not depend on a specific network backbone, template length L, or specific implementation platform. As long as the process of "normalized features → fixed-length binary template → key-driven protection → Hamming distance judgment within the protected domain → support for key replacement and revocation" is followed, it falls within the protection scope of this invention.

Claims

1. A method for protecting revocable face templates based on deep hashing, characterized in that, include: Perform face detection and alignment preprocessing on the face images to be processed; Normalized facial feature vectors are obtained using a feature extraction network; The feature vector is symbolically quantized according to a preset projection mapping to obtain a binary vector of length B; Based on key-derived dimension index selection and dimension permutation, the binary vector is rearranged and subsets are selected to generate a protected face template. The Hamming distance between protected templates is calculated in the protected domain and compared and judged based on the threshold.

2. The method according to claim 1, characterized in that, The method applies a quality-adaptive margin constraint based on the sample quality score during the training phase. The quality score is obtained by normalizing the feature norm, the angular distance to the class center, or a combination thereof, in order to improve the discriminative power of samples of different quality.

3. The method according to claim 1, characterized in that, The projection mapping is a linear projection, a nonlinear learnable mapping, or a combination of both; the symbol quantization uses a symbol function with a zero threshold or a threshold function with a learnable bias.

4. The method according to claim 1, characterized in that, The length B of the binary vector is any value between 64 and 1024, preferably 128, 256 or 512.

5. The method according to claim 1, characterized in that, The key-derived dimension index selection generates a dimension index sequence and permutation table by inputting the user key as a seed into a pseudo-random function, and then rearranges and selects subsets of the binary vector accordingly.

6. The method according to claim 1, characterized in that, To improve the robustness of the comparison, error correction coding is introduced before or after the generation of the protected template, so that the correct judgment can still be made as long as the preset bit flip ratio is not exceeded.

7. The method according to claim 1, characterized in that, The method supports generating new protected templates for the same user by updating the key or index selection rules to achieve template revocation and non-linkability. The system only stores the protected template and key derivation parameters and does not save the original image and unprotected feature vectors.

Images