A face recognition privacy protection method based on dictionary learning
By optimizing face recognition through dictionary learning and Laplace noise mechanism, the problems of insufficient privacy protection and low data utility are solved, achieving stronger privacy protection and higher data availability.
Patent Information
- Application Number
- CN202211552958.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-12-06
- Publication Date
- 2025-11-25
- Estimated Expiration
- 2042-12-06
AI Technical Summary
Existing facial recognition technology suffers from low data utility and serious privacy leaks, especially since adversaries can use machine learning to analyze publicly available facial images to infiltrate the system.
A dictionary-based learning approach is adopted to obtain pixel vectors of face images through compressed sensing technology, learn the dictionary basis matrix and encoding coefficient vector, add Laplacian noise to the encoding coefficient vector, and combine the Lagrange multiplier method to optimize the privacy protection and data availability.
It effectively hides facial features to prevent intrusion by adversaries, while improving the usefulness of data with the same level of privacy protection, thus achieving a higher level of privacy protection and data utility.
Smart Images
Figure CN116343282B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the field of computer, and particularly relates to a face recognition privacy protection method based on dictionary learning. BACKGROUND
[0002] Face recognition is widely used in biometric authentication in many application fields, such as public security, finance and e-commerce, which brings a lot of convenience to people's life. However, an enemy can use artificial intelligence technology to train publicly available face images, analyze the face features of an individual of interest, and then invade the corresponding face recognition-based application system. Therefore, the face recognition-based application system faces a serious privacy leakage problem. At present, image privacy protection methods mainly focus on mosaicking and blurring, but the existing methods do not focus on face recognition privacy protection problems and have the problem of low data utility. We first propose a face recognition privacy protection method based on dictionary learning, which can provide greater privacy protection under the condition of given data utility. And compared with the existing methods, the data utility of the method proposed by us is higher, and the privacy protection is stronger.
[0003] Glossary:
[0004] Privacy budget: a form of privacy protection strength, the smaller the set value, the greater the privacy protection strength.
[0005] Face feature: face information is represented by characteristic features, that is, the gray information of the face image is used to extract global or local features through a corresponding algorithm. SUMMARY
[0006] To solve the above problems, the application provides a face recognition privacy protection method based on dictionary learning. The application proposes a face image privacy protection method that can effectively hide face features, which can provide greater privacy protection under the condition of given data utility, and the data utility is higher under the same privacy protection strength.
[0007] To achieve the above technical effects, the technical scheme of the application is:
[0008] A face recognition privacy protection method based on dictionary learning, comprising the following steps:
[0009] Step 1: assuming that there is a face image F, the image is represented by an image pixel vector of length N . Through the compression sensing technology, the framework is: wherein, is an MxN measurement matrix, which is randomly generated and recorded; is a measurement vector of length M. Thus, the pixel vector of the face image can be obtained where N represents the length of the pixel vector.
[0010] Step two: learn the dictionary basis matrix by dictionary learning and the corresponding encoding coefficient vector The learning process is divided into two steps: the first step is stochastic gradient descent; the second step is -norm operation; the specific operation is as shown below.
[0011] 1) Stochastic gradient descent: through the chain rule, the gradient of the local dictionary basis matrix and the gradient of the local encoding coefficient vector can be obtained. Then the overall gradient of the dictionary basis matrix and the encoding coefficient vector is and Therefore, the dictionary basis matrix and the encoding coefficient vector can be updated, and the update process is as follows: and where η is the learning rate.
[0012] 2) -norm operation: using the Iterative Soft Thresholding Algorithm (ISTA), the -norm operation is performed on the updated encoding coefficient vector, so that the sparsity of the encoding coefficient vector is where λ is the threshold value, and values exceeding the threshold value are considered invalid values, which are set to 0.
[0013] After the training of the dictionary learning network, the overall dictionary basis matrix and the overall encoding coefficient vector
[0014] Step three: decompose the face image vector into the product of the selected dictionary basis and the encoding coefficient vector that is, Therefore, our privacy protection framework is to add Laplace noise to the encoding coefficient vector, that is, where the Laplace scale parameter vector
[0015] Step four: calculate the dictionary basis matrix The sum of the squares of the first N elements in the mth (m = 1, 2,..., M) row, so as to obtain the weight vector Optimization for data usability.
[0016] Step five: using the encoding coefficient vector decomposed in step three, calculate the minimum difference of face feature vectors between adjacent users to obtain a partial sensitivity vector, where m is the element of the feature vector, and adjacent users refer to the two users with the smallest average distance between face feature vectors; where min{·,·} represents the minimum of the two, and a is the adjustment parameter, is the lower bound of sensitivity, is the upper bound of sensitivity.
[0017] Step six: using the Lagrange multiplier method, weight vector and partial sensitivity vector, calculate the optimal noise size parameter
[0018] According to the definition of differential privacy, the privacy budget where, is the pixel vector of the face image B' after adding noise belonging to the face feature is the pixel vector of the face image A' after adding noise belonging to the face feature , and further, where,
[0019] The definition of face image data utility is as follows.
[0020] Definition (face image data utility) face image data utility is the visual effect of the image, which is measured by the bias variance of the image pixel vector, and its definition is as follows:
[0021]
[0022] According to our face image privacy protection framework, Therefore, the data utility can be simplified as where the weight vector σ m is the standard deviation of Laplace noise, and this Laplace noise is independent. Since the variance of Laplace distribution is Further, the data utility can be simplified as
[0023] Using the Lagrange multiplier method, the face image data utility is optimized under the given privacy budget ε Under the given initial condition and the initial value of the privacy budget ε is ε0, then solve optimal solution.
[0024] According to the definition of data utility,
[0025] According to the definition of data utility,
[0026] According to the definition of data utility,
[0027] Therefore, the optimal Laplace scale parameter
[0028] The advantages of the present application are as follows:
[0029] 1. By using the perturbation mechanism subject to the Laplace distribution, the facial features in the face image can be effectively hidden, and the enemy can be prevented from training and learning the publicly released face image through machine learning method, so as to invade the face recognition system;
[0030] 2. By using the constraint optimization method, the best data availability can be realized under the given privacy protection strength. BRIEF DESCRIPTION OF DRAWINGS
[0031] Figure 1 The data availability comparison chart of the method of the present application and the existing Standard-DP and Partial-DP methods on two typical face image data sets LFW and PubFig respectively;
[0032] Figure 2 The comparison chart of the method of the present application and the Standard-DP and Partial-DP methods in the privacy protection strength on two typical face image data sets LFW and PubFig respectively.
[0033] Figure 3 The visual effect of the face image after adding noise. DETAILED DESCRIPTION
[0034] The present application will be further described below in combination with the drawings and examples.
[0035] The development language of the present application is Python3.7, and the operating system is win10. The experimental data is a face image data set containing multiple users. The specific steps are as follows:
[0036] Step one: obtaining the pixel vector of the face image
[0037] The pixel vector of the face image is obtained through the compressed sensing technology.
[0038] Step two: learning the dictionary basis matrix and the encoding coefficient vector of the face image
[0039] The dictionary learning is used to learn the overall dictionary basis matrix and the overall encoding coefficient vector.
[0040] Step three: adding Laplace noise
[0041] The face image vector is decomposed into the product of the selected dictionary basis and the encoding coefficient vector, and the Laplace noise is added to the encoding coefficient vector, i.e.
[0042] Step four: obtaining the privacy-protected weight
[0043] The sum of squares of the first to Nth elements in the mth (m=1, 2, …, M) row of the dictionary basis matrix is calculated, thereby obtaining the privacy-protected weight.
[0044] Step five: calculating the partial sensitivity vector
[0045] The partial sensitivity vector is calculated using the encoding coefficient vector obtained in step three. Wherein, min{·,·} represents the minimum value of two, and α is an adjustment parameter, is the lower bound of the sensitivity, is the upper bound of the sensitivity.
[0046] Step six: calculating the optimal noise scale parameter
[0047] The optimal noise scale parameter is calculated using the Lagrange multiplier method, the weight vector and the partial sensitivity vector.
[0048] The face recognition differential privacy protection algorithm is as follows:
[0049]
[0050] The above examples are only used to help understand the core idea of the present application; meanwhile, for those skilled in the art, according to the idea of the present application, the specific implementation and application range will be changed, and the above description should not be understood as the limitation of the present application.
Claims
1. A face recognition privacy protection method based on dictionary learning, characterized in that, Includes the following steps: Step 1: Obtain the pixel vector of the face image using compressed sensing technology. Where N represents the length of the pixel vector; Step 2: Use a dictionary learning neural network method to learn the face image dataset and obtain the corresponding sparse dictionary basis matrix. Where N represents the number of rows in the matrix, and M represents the number of columns in the matrix; Step 3: Convert the face image vector Decomposed into selected dictionary bases and coding coefficient vector The product of, that is, Step 4: Using the dictionary base obtained in Step 3, calculate the weights for optimizing data availability; Step 5: Using the coding coefficient vector obtained in Step 3, calculate the partial sensitivity; Step Six: Calculate the optimal noise scale parameter by combining the Lagrange multiplier method, weights, and partial sensitivity. By combining the Laplace distribution, the coding coefficient vector after adding noise is obtained. Thus, the image with added noise is obtained.
2. The face recognition privacy protection method based on dictionary learning as described in claim 1, characterized in that, The process of compressed sensing technology in step one is as follows: in, It is a pixel vector of a human face image. The measurement matrix is randomly generated (M×N). Let M be the measurement vector of length M.
3. The face recognition privacy protection method based on dictionary learning as described in claim 1, characterized in that, The dictionary learning process in step two is as follows: First, the gradients of the dictionary bases are updated using the chaining method; then, the L1-norm operation is performed on the coding coefficients updated during stochastic gradient descent using an iterative soft thresholding algorithm to obtain the sparsity of the coding coefficients.
4. The face recognition privacy protection method based on dictionary learning as described in claim 1, characterized in that, In step four, the weight w m The calculation process involves calculating the matrix. The sum of the squares of the first to Nth elements in the m-th row (m = 1, 2, ..., M).
5. The face recognition privacy protection method based on dictionary learning as described in claim 1, characterized in that, Partial sensitivity s in step five m The calculation process involves calculating the minimum difference between the facial feature vectors of adjacent users, where m is an element of the feature vector, and adjacent users refer to the two users with the smallest mean distance between their facial feature vectors.
6. The face recognition privacy protection method based on dictionary learning as described in claim 1, characterized in that, The calculation process for the optimal noise level parameter in step six is as follows: Where ε0 is the privacy budget value input by the user, and δ is a parameter with a very small value.