Face recognition method and device, equipment, storage medium and program product

Abstract

The invention discloses a face recognition method and device, equipment, a storage medium and a program product, and the method comprises the steps: carrying out the multi-scale feature extraction of an input face image through a plurality of groups of pre-constructed Gabor filters, obtaining a target feature vector, carrying out the evaluation of each participation node based on a plurality of evaluation dimensions, obtaining a multi-dimensional quantization parameter, and carrying out the recognition of the target feature vector. Aggregating gradient training results of the participating nodes according to the multi-dimensional quantization parameters to obtain global gradient parameters, updating a local face recognition model based on the global gradient parameters, and inputting the target feature vector into the updated local face recognition model for face recognition; due to the fact that the features of the input face image are extracted in a multi-scale mode, the characterization capacity of the image features is enhanced, data contribution weights provided by distributed nodes are effectively balanced through multi-dimensional node evaluation, the convergence speed and the recognition precision of a face recognition model are improved, and the accuracy of face recognition is greatly improved.

Description

Technical Field

[0001] The present invention relates to the field of face recognition, and in particular to a face recognition method, apparatus, device, storage medium and program product. Background Art

[0002] With the rapid development of information technology, facial recognition systems are increasingly being used in security, finance, social networking, and other fields. In particular, achieving efficient, secure, and fair model training in distributed computing environments has become a research hotspot. Traditional facial recognition model training methods face significant challenges in data privacy protection, model performance optimization, and incentive mechanism design for multi-party collaboration across institutions and regions.

[0003] Currently, distributed face recognition model training methods primarily employ centralized aggregation or federated averaging algorithms. These training models are relatively simple, with fixed model weight distribution and no dynamic adjustment based on the data quality of different participants. If the data quality provided by some participants is low or the sample distribution is uneven, this will directly impact the convergence speed and recognition accuracy of the global model. However, if the same aggregation weights are applied to all participants, it will be difficult to fully motivate participants with high-quality data. Furthermore, existing methods typically employ a single convolutional neural network architecture for feature extraction, which has limited ability to express facial features under varying lighting conditions and posture changes, restricting the model's generalization performance. Summary of the Invention

[0004] The main purpose of the present invention is to provide a face recognition method, device, equipment, storage medium and program product, aiming to solve the technical problems that the existing technology cannot effectively balance the contribution weights of different participants in the distributed face recognition process, and the feature extraction method is single, resulting in low face recognition accuracy.

[0005] To achieve the above objectives, the present invention provides a face recognition method, which is applied to a participating node in a federated learning blockchain, wherein the federated learning blockchain includes multiple participating nodes. The method includes the following steps:

[0006] Perform multi-scale feature extraction on the input face image using multiple sets of pre-built Gabor filters to obtain a target feature vector. The Gabor filters are composed of multiple sets of complex filter kernels generated based on different directions and spatial frequencies.

[0007] Evaluate each participating node based on multiple evaluation dimensions to obtain multi-dimensional quantitative parameters, wherein the evaluation dimensions include marginal contribution dimension, risk dimension, and reputation dimension, and the multi-dimensional quantitative parameters include marginal contribution degree, risk weight parameter, and target reputation value;

[0008] Obtaining the gradient training results uploaded by each participating node to the federated learning blockchain, where the gradient training results are obtained by each participating node performing gradient training on the local face recognition model;

[0009] Aggregating the gradient training results of each participating node according to the multidimensional quantization parameter to obtain a global gradient parameter;

[0010] The local face recognition model is updated based on the global gradient parameter, and the target feature vector is input into the updated local face recognition model for face recognition.

[0011] Optionally, performing multi-scale feature extraction on the input face image using multiple sets of pre-built Gabor filters to obtain a target feature vector includes:

[0012] Perform grayscale processing on the input face image to obtain an initial grayscale image;

[0013] Normalizing the initial grayscale image to obtain a candidate grayscale image;

[0014] Generating corresponding complex filter kernels based on multiple sets of different directions and spatial frequencies, and generating multiple sets of Gabor filters based on the complex filter kernels;

[0015] The candidate grayscale image is decomposed at multiple scales by using multiple groups of Gabor filters, and multiple groups of texture feature maps of different directions and spatial frequencies are extracted based on the filter output results. The wavelet function of the Gabor filter refers to the following formula:

[0016]

[0017]

[0018] Among them, z represents the spatial coordinate value (x, y) of the pixel point in the input face image, k u,v represents the center frequency of the Gabor filter, represents the directionality of the two-dimensional Gabor filter, σ 2 Represents the square of the standard deviation of the Gaussian envelope, which is used to control the spatial distribution width of the Gabor filter. represents the Gaussian envelope term, which is used to limit the spatial distribution of the plane wave, exp(ik u,v z) represents the periodicity and directionality of the wavelet, Used for energy spectrum attenuation compensation;

[0019] Feature analysis is performed on the multiple groups of texture feature maps with different directions and spatial frequencies to generate target feature vectors.

[0020] Optionally, performing feature analysis on the multiple groups of texture feature maps of different directions and spatial frequencies to generate a target feature vector includes:

[0021] Traversing each pixel in the texture feature map and calculating the LBP value of each pixel;

[0022] Generate an LBP feature map corresponding to each Gabor filter based on the LBP value of each pixel point in the texture feature map;

[0023] Divide the LBP feature map into multiple sub-blocks, and generate a distribution histogram of the LBP value for each sub-block;

[0024] The distribution histograms of each sub-block are spliced ​​together to generate the target feature vector.

[0025] Optionally, evaluating each participating node based on multiple evaluation dimensions to obtain multi-dimensional quantitative parameters includes:

[0026] The marginal contribution of each participating node is evaluated to obtain the marginal contribution of each participating node. The marginal contribution is calculated based on the following formula:

[0027]

[0028] in, represents the marginal contribution of participating node i, s(i) represents the sub-alliance composed of all participating nodes including participating node i, v(s) represents the profit value of the sub-alliance, v(s / {i}) represents the profit value of the sub-alliance after excluding participating node i, w(|s|) represents the contribution weight of the sub-alliance, |s| represents the size of the sub-alliance, and n represents the total number of nodes in the sub-alliance;

[0029] Based on the integrity of each participating node, the risk weight of each data supply node is evaluated to obtain the risk weight of each participating node. The risk weight is calculated based on the following formula:

[0030]

[0031] Among them, Λ i represents the risk weight of participating node i, represents the integrity of participating node i in the current block time period h, λ i represents the integrity vector of participating node i;

[0032] Conduct a quantitative reputation assessment based on the data quality information and task completion rate of each participating node to obtain the initial reputation value of each participating node;

[0033] The initial reputation value of each participating node is normalized based on the maximum reputation value of the participating nodes to obtain the target reputation value of each participating node.

[0034] Optionally, performing a risk weight assessment on each data supply node based on the integrity of each participating node to obtain the risk weight of each participating node includes:

[0035] Obtain the historical integrity of each participating node in the historical block time period;

[0036] Evaluate the noise level of the data provided by each participating node and obtain data quality information of each participating node;

[0037] generating an exponential decay coefficient for each participating node based on the data quality information;

[0038] The historical credibility is updated according to the exponential decay coefficient to obtain an updated credibility. The credibility is updated according to the following formula:

[0039]

[0040] Among them, h represents the current block time period, h-1 represents the previous block time period, represents the integrity of participating node i in the previous block time period, α represents the exponential decay coefficient, Indicates the integrity of participating node i in the current block time period, reward i It represents the reward value obtained by participating node i in the local model training of the current block time period, and β represents a positive number. Represents the cumulative historical integrity score of participating node i;

[0041] Based on the updated credibility, a risk weight assessment is performed on each data supply node to obtain the risk weight of each participating node.

[0042] Optionally, aggregating the gradient training results of each participating node according to the multidimensional quantization parameter to obtain a global gradient parameter includes:

[0043] Monitor abnormal behavior of each participating node;

[0044] Mark participating nodes with abnormal behaviors as abnormal nodes;

[0045] The abnormal nodes are weighted according to the multidimensional quantization parameters to obtain a target weight vector, which is calculated based on the following formula:

[0046]

[0047] Where w' represents the target weight vector;

[0048] The gradient training results of each participating node are aggregated based on the target weight vector and the multidimensional quantization parameter to obtain a global gradient parameter.

[0049] In addition, to achieve the above-mentioned purpose, the present invention further provides a face recognition device, comprising:

[0050] A multi-scale feature extraction module is used to extract multi-scale features from the input face image using multiple sets of pre-built Gabor filters to obtain a target feature vector. The Gabor filters are composed of multiple sets of complex filter kernels generated based on different directions and spatial frequencies.

[0051] A multi-dimensional evaluation module is used to evaluate each participating node based on multiple evaluation dimensions to obtain multi-dimensional quantitative parameters. The evaluation dimensions include marginal contribution dimension, risk dimension, and reputation dimension. The multi-dimensional quantitative parameters include marginal contribution degree, risk weight parameter, and target reputation value.

[0052] A distributed gradient training module is used to obtain the gradient training results uploaded by each participating node to the federated learning blockchain. The gradient training results are obtained by each participating node performing gradient training on the local face recognition model.

[0053] A gradient parameter aggregation module is used to aggregate the gradient training results of each participating node according to the multi-dimensional quantization parameter to obtain a global gradient parameter;

[0054] The face recognition module is used to update the local face recognition model based on the global gradient parameter, and input the target feature vector into the updated local face recognition model to perform face recognition.

[0055] In addition, to achieve the above-mentioned purpose, the present application also proposes a face recognition device, which includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program is configured to implement the steps of the face recognition method described above.

[0056] In addition, to achieve the above-mentioned purpose, the present application also proposes a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the face recognition method described above are implemented.

[0057] In addition, to achieve the above-mentioned purpose, the present application also provides a computer program product, which includes a computer program, and when the computer program is executed by a processor, it implements the steps of the face recognition method described above.

[0058] The present invention uses multiple sets of pre-built Gabor filters to extract multi-scale features from the input face image to obtain the target feature vector. The Gabor filter is composed of multiple sets of complex filter kernels generated based on different directions and spatial frequencies. Each participating node is evaluated based on multiple evaluation dimensions to obtain multi-dimensional quantitative parameters. The evaluation dimensions include marginal contribution dimension, risk dimension and reputation dimension. The multi-dimensional quantitative parameters include marginal contribution degree, risk weight parameter and target reputation value. The gradient training results uploaded by each participating node to the federated learning blockchain are obtained. The gradient training results are the local human factors of each participating node. The face recognition model is obtained by gradient training; the gradient training results of each participating node are aggregated according to the multi-dimensional quantization parameter to obtain a global gradient parameter; the local face recognition model is updated based on the global gradient parameter, and the target feature vector is input into the updated local face recognition model for face recognition; because the present invention extracts the features of the input face image through multi-scale, thereby enhancing the representation ability of the image features, through multi-dimensional node evaluation, effectively balancing the data contribution weights provided by the distributed nodes, improving the convergence speed and recognition accuracy of the face recognition model, and greatly improving the accuracy of face recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0060] Figure 1 Schematic diagram of the structure of a face recognition device in the hardware operating environment involved in an embodiment of the present invention;

[0061] Figure 2 This is a flow chart of an embodiment of a face recognition method according to the present invention;

[0062] Figure 3 A schematic diagram of the process of multi-scale feature extraction in an embodiment of the face recognition method of the present invention;

[0063] Figure 4 A schematic diagram of the framework of a face recognition system according to an embodiment of the present invention;

[0064] Figure 5 This is a structural block diagram of an embodiment of a face recognition device of the present invention.

[0065] The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings and in conjunction with the embodiments. DETAILED DESCRIPTION

[0066] It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0067] Reference Figure 1 , Figure 1 This is a schematic diagram of the face recognition device structure of the hardware operating environment involved in the embodiment of the present invention.

[0068] like Figure 1 As shown, the face recognition device may include: a processor 1001, such as a central processing unit (CPU), a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. The communication bus 1002 is used to implement connection and communication between these components. The user interface 1003 may include a display screen (Display) and an input unit such as a keyboard (Keyboard). Optionally, the user interface 1003 may also include a standard wired interface and a wireless interface. The network interface 1004 may optionally include a standard wired interface and a wireless interface (such as a wireless fidelity (WI-FI) interface). The memory 1005 may be a high-speed random access memory (RAM) or a stable non-volatile memory (NVM), such as a disk storage. The memory 1005 may also be a storage device independent of the aforementioned processor 1001.

[0069] Those skilled in the art will understand that Figure 1 The structure shown in the figure does not constitute a limitation on the face recognition device, and may include more or fewer components than shown in the figure, or combine certain components, or arrange the components differently.

[0070] like Figure 1 As shown, the memory 1005 as a computer-readable storage medium may include an operating system, a network communication module, a user interface module, and a face recognition program.

[0071] exist Figure 1 In the face recognition device shown, the network interface 1004 is mainly used for data communication with the network server; the user interface 1003 is mainly used for data interaction with the user; the processor 1001 and the memory 1005 in the face recognition device of the present invention can be set in the face recognition device, and the face recognition device calls the face recognition program stored in the memory 1005 through the processor 1001 and executes the face recognition method provided by the embodiment of the present invention.

[0072] The embodiment of the present invention provides a face recognition method, referring to Figure 2 , Figure 2 FIG. 1 is a flow chart of an embodiment of a face recognition method according to the present invention.

[0073] This embodiment is applied to participating nodes in a federated learning blockchain, which includes multiple participating nodes. In this embodiment, the face recognition method includes the following steps:

[0074] Step S10: performing multi-scale feature extraction on the input face image using multiple sets of pre-built Gabor filters to obtain a target feature vector.

[0075] It should be noted that there are three technical bottlenecks in the training of traditional distributed face recognition models: first, there is a contradiction between efficiency and robustness in the feature extraction link. Although multimodal feature fusion improves the discrimination, it has high computational complexity and weak adaptability to dynamic scenes, resulting in prominent redundancy problems; second, node contribution evaluation relies on static assumptions, and it is difficult to objectively quantify the marginal contribution of nodes during dynamic training, which leads to evaluation bias; third, the model aggregation strategy is sensitive to data heterogeneity, and client features and parameters lack adaptive alignment capabilities.

[0076] Currently, feature extraction is a core component of distributed training for face recognition models. Its efficiency and robustness directly impact model performance. Existing research primarily focuses on feature operator optimization and multimodal feature fusion. Despite this, feature extraction research still faces challenges such as high computational complexity, poor adaptability to dynamic scenarios, and feature redundancy. In particular, the lack of effective dimensionality compression mechanisms for the fusion of heterogeneous features with deep networks limits its large-scale application in distributed training.

[0077] Distributed training node contribution evaluation and model aggregation have garnered widespread attention, helping to identify high-quality training datasets and improve the efficiency of distributed training for face recognition models. However, existing aggregation methods lack the ability to adaptively align heterogeneous client data, limiting their reliability in complex scenarios.

[0078] In response to the above technical deficiencies, in one embodiment of the present invention, a distributed training method for a face recognition model based on data contribution aggregation is proposed. At the feature extraction level, a composite feature fusion mechanism of Gabor transform and local binary pattern (LBP) is adopted, in which the Gabor filter group extracts multi-directional and multi-scale frequency domain features, and the LBP operator captures local texture detail features. Through feature space reconstruction and three-dimensional feature vector dimensionality reduction processing, a low-dimensional feature representation with strong discriminative power is formed. This technical solution effectively enhances the robustness of the model to changes in illumination and posture, and lays a high-quality feature foundation for distributed collaborative training.

[0079] In terms of model collaborative training mechanisms, the Shapley Value theory is introduced to construct a data contribution evaluation model. By calculating each participant's marginal contribution to improving model accuracy in real time, the global model aggregation weight is dynamically adjusted to establish an incentive mechanism that matches "contribution" with "reward." This technology eliminates the negative impact of participants with low-quality data on the global model and stimulates the participation of high-quality data holders.

[0080] It should be understood that the execution subject of this embodiment can be a computing service device with data processing, network communication, and program execution functions, such as a personal computer or server, or a terminal electronic device capable of implementing the above functions. The following uses a facial recognition device as an example to illustrate this embodiment and the following embodiments.

[0081] It should be noted that the Gabor filter is composed of multiple groups of complex filter kernels generated based on different directions and spatial frequencies.

[0082] Furthermore, in order to effectively enhance the robustness of the face model to changes in illumination and posture, and lay a high-quality feature foundation for distributed collaborative training, thereby improving the accuracy of face recognition, refer to Figure 3 , Figure 3 FIG. 5 is a flow chart of multi-scale feature extraction in one embodiment. Step S10 may include:

[0083] Step S101: grayscale processing is performed on the input face image to obtain an initial grayscale image.

[0084] In some embodiments, the face recognition device performs grayscale processing on the input face image, assigns weights of 0.299, 0.587, and 0.114 to the red, green, and blue channels of the image, respectively (in line with the sensitivity of the human eye), and uses the weighted averaging method to merge the three RGB channels into a single-channel grayscale.

[0085] Step S102: performing normalization processing on the initial grayscale image to obtain a candidate grayscale image.

[0086] In some embodiments, the face recognition device divides each pixel value of the grayscale image by 255.0 and converts the data type to float32 format, normalizing and linearly mapping the pixel value from the integer range of [0, 255] to the floating point range of [0, 1].

[0087] Step S103: generating corresponding complex filter kernels based on multiple sets of different directions and spatial frequencies, and generating multiple sets of Gabor filters based on the complex filter kernels.

[0088] In some embodiments, when a face recognition device uses multiple sets of Gabor filters for filtering, it first presets different directions (such as 0°, 45°, 90°, 135°, etc.) and spatial frequencies (parameters related to wavelength) to generate corresponding complex filter kernels.

[0089] Step S104: performing multi-scale decomposition on the candidate grayscale image using the multiple groups of Gabor filters, and extracting multiple groups of texture feature maps of different directions and spatial frequencies based on the filter output results.

[0090] It should be noted that the wavelet function of the Gabor filter refers to the following formula:

[0091]

[0092] Among them, z represents the spatial coordinate value (x, y) of the pixel point in the input face image, k u,v represents the center frequency of the Gabor filter, represents the directionality of the two-dimensional Gabor filter, σ 2 Represents the square of the standard deviation of the Gaussian envelope, which is used to control the spatial distribution width of the Gabor filter. represents the Gaussian envelope term, which is used to limit the spatial distribution of the plane wave, exp(ik u,v z) represents the periodicity and directionality of the wavelet, Used for energy spectrum attenuation compensation.

[0093] Step S105: performing feature analysis on the multiple groups of texture feature maps of different directions and spatial frequencies to generate target feature vectors.

[0094] In some embodiments, the face recognition device performs a convolution operation on each Gabor filter and the grayscale image to extract the texture response in a specific direction and frequency, and takes the amplitude of the filtering result as a feature.

[0095] In some embodiments, facial recognition devices use Gabor transforms to decompose images at multiple scales, effectively extracting local information at different scales and directions. By selecting different frequencies and directions, multiple sets of two-dimensional Gabor wavelets can be generated, allowing for analysis of the image from multiple perspectives during feature extraction.

[0096] Furthermore, in order to accurately extract local features and global features of the image, thereby enhancing the feature representation capability of the image, the above step S105 may include:

[0097] Step S1051: traverse each pixel in the texture feature map and calculate the LBP value of each pixel;

[0098] Step S1052: generating an LBP feature map corresponding to each Gabor filter based on the LBP value of each pixel point in the texture feature map;

[0099] Step S1053: Divide the LBP feature map into multiple sub-blocks, and generate a distribution histogram of the LBP value for each sub-block;

[0100] Step S1054: concatenate the distribution histograms of the sub-blocks to generate a target feature vector.

[0101] In some embodiments, the face recognition device performs a local binary pattern (LBP) calculation on each filtered image, takes each pixel as the center, compares the difference between its neighboring pixels and the center value, generates a binary code, and then converts it into a decimal value to form an LBP feature map.

[0102] It should be noted that the local binary pattern (LBP) is calculated according to the following formula:

[0103]

[0104] Among them, LBP P,R represents the LBP value, P represents the number of neighborhood points around the central pixel, R represents the distance between the neighborhood point and the central pixel, g p represents the gray value of the pth neighborhood point, g c Represents the gray value of the central pixel, s(c) is used to compare the gray value of the neighborhood point with the central pixel, 2 p It is used to weight the comparison result (0 or 1) of each neighborhood point according to the bit weight value, and finally form a binary number (which is the LBP value after conversion to decimal).

[0105] In some embodiments, the face recognition device divides the LBP feature maps of all Gabor filter channels into blocks and calculates the histogram, divides each Gabor filtered image into several small blocks, and calculates the LBP histogram in each small block (that is, assuming that the LBP pattern has b bins, the histogram of each block is a b-dimensional vector. It is concatenated into a high-dimensional feature vector to comprehensively describe the local texture and global structural information of the image).

[0106] In some embodiments, the face recognition device can use a lightweight convolutional network structure CNN for algorithm optimization. In the CNN network, the input layer is set to adapt to the standardized image size (64×64 pixels), and two layers of convolution (kernel size 3×3, step size 1) and maximum pooling (kernel size 2×2) are defined for the feature extraction part. The output dimension of the fully connected layer is consistent with the number of classification task categories, and the final target feature vector (i.e., the comprehensive feature vector) is obtained. Figure 3 , Figure 3This is a schematic diagram of the feature extraction process in one embodiment. The face recognition device grayscales and normalizes the input face image, filters the processed image through multiple Gabor filters of different scales, obtains texture feature maps of multiple scales, calculates and statistically analyzes the texture distribution through LBP features, and determines whether each Gabor filter has completed feature map processing. If not, feature processing continues; if so, the feature histograms output by the Gabor filters of each scale are connected in series to output a comprehensive feature vector (i.e., the target feature vector).

[0107] It should be noted that Figure 3 This paper presents a schematic diagram of the feature extraction technology framework designed in this solution. The entire processing flow begins with the raw image input. In the preprocessing stage, grayscale conversion and data normalization are performed. To further enhance texture representation capabilities, the system uses a multi-channel Gabor filter bank for feature enhancement. By configuring multiple filter banks with different frequency and orientation parameters, local texture details in the image are effectively extracted. Furthermore, an improved LBP operator is introduced for secondary processing of the feature map. This algorithm establishes a grayscale contrast relationship between the central pixel and its annular neighboring pixels, ultimately forming a highly discriminative binary feature code. A histogram is calculated for the LBP features output by each Gabor filter, and the distribution of different texture patterns is statistically analyzed to form a texture description using a single filter. The final output, a comprehensive feature vector, incorporates multi-scale and multi-directional texture information, making it suitable for tasks such as image classification and recognition. This method combines the frequency domain analysis capabilities of Gabor with the local texture description advantages of LBP, enhancing feature representation capabilities.

[0108] Step S20: Evaluate each participating node based on multiple evaluation dimensions to obtain multi-dimensional quantitative parameters.

[0109] It should be noted that the evaluation dimensions include marginal contribution dimension, risk dimension and reputation dimension, and the multi-dimensional quantitative parameters include marginal contribution degree, risk weight parameter and target reputation value.

[0110] It should be noted that in the contribution evaluation mechanism, in order to screen high-quality data from data stored in different nodes and reduce the impact of low-quality data on the performance of the final training model, quality indicators and data freshness are designed for each round of model updates. The contribution evaluation mechanism is designed based on the Shapley value theory, which quantifies the marginal contribution of nodes and introduces risk weight correction to achieve dynamic matching of data quality and aggregation weight.

[0111] Furthermore, in order to accurately assess the contribution, risk, and reputation of each participating node, in one embodiment, the above step S20 may include:

[0112] Step S201: perform marginal contribution evaluation on each participating node to obtain the marginal contribution degree of each participating node.

[0113] It should be noted that this embodiment can use Shapley value sampling to calculate the marginal contribution value of each data provider in different federated learning sub-alliances to quantify the marginal contribution of each data provider. The marginal contribution is calculated based on the following formula:

[0114]

[0115] in, represents the marginal contribution of participating node i, s(i) represents the sub-alliance consisting of all participating nodes that include participating node i, v(s) represents the revenue of the sub-alliance, v(s / {i}) represents the revenue of the sub-alliance after excluding participating node i, w(|s|) represents the contribution weight of the sub-alliance, |s| represents the size of the sub-alliance, and n represents the total number of nodes in the sub-alliance. In federated learning, a collaborative group consisting of a group of nodes (members) is called a sub-alliance. Member i in sub-alliance s(i) participates in local training of federated learning by sharing data or model parameters, jointly optimizing the global model. s(i) is the set of all sub-alliances that include member i, with each member i providing data or computation.

[0116] Step S202: Perform a risk weight assessment on each data supply node based on the integrity of each participating node to obtain the risk weight of each participating node.

[0117] It should be noted that this embodiment can introduce a set of additional data risk weight vectors Λ=(Λ1, Λ2,…, Λ n ), where: i Represents the data privacy risk weight of the i-th distributed training participant. Taking into account the integrity of the current participating node's past experience, the calculation formula of the Shapley value has been modified based on the characteristics of the participating distributed training data source, so that the weight is related to the integrity of the participating node.

[0118] set up is the integrity of miner node i in block time period h. Since all source data have been identified before the distributed training task begins, the integrity vector λ is consistent in a single distributed training task and the sum of the weights is 1. The risk weight is calculated based on the following formula:

[0119]

[0120] Among them, Λ i represents the risk weight of participating node i, represents the integrity of participating node i in the current block time period h, which is dynamically updated over time, λ iRepresents the integrity vector of participating node i.

[0121] Step S203: Perform a reputation quantitative evaluation based on the data quality information and task completion rate of each participating node to obtain an initial reputation value of each participating node.

[0122] It should be noted that the data quality information may be the data quality level provided by each participating node in the current cycle and historical cycles. For example, the data quality information may be obtained based on an assessment of the noise level of the data provided by the participating nodes. The aforementioned task completion rate may be the task completion rate of distributed training performed by the participating nodes in historical cycles.

[0123] Step S204: normalizing the initial reputation value of each participating node based on the maximum reputation value of the participating nodes to obtain a target reputation value of each participating node.

[0124] In some embodiments, the face recognition device may set a random hash value Convert(h i ), mapping the participant's unique identification hash value to a random number (0,1]. According to the parameter value, the mapping priority is where r i Represents the node reputation value, max(r) represents the maximum reputation value among all nodes. The final priority is determined by the reputation value and the random hash value Convert(h i ) jointly decide.

[0125] Furthermore, in order to accurately evaluate the integrity of each participating node and effectively reduce the risk of facial recognition data leakage, in some embodiments, step S202 may include:

[0126] Step S2021: Obtain the historical integrity of each participating node in the historical block time period;

[0127] Step S2022: performing noise level assessment on the data provided by each participating node to obtain data quality information of each participating node;

[0128] Step S2023: generating an exponential decay coefficient for each participating node based on the data quality information;

[0129] Step S2024: updating the historical credibility according to the exponential decay coefficient to obtain an updated credibility;

[0130] Step S2025: Perform risk weight assessment on each data supply node based on the updated credibility to obtain the risk weight of each participating node.

[0131] It should be noted that face recognition devices can use dynamic weight integration to make dynamic adjustments based on the data quality (noise level) of the participating nodes. High-quality data will receive higher weights. At the same time, the influence of historical contributions can be adjusted through the exponential decay coefficient, and the earlier the contribution, the more obvious the decay.

[0132] In some embodiments, the integrity of participating nodes can be recorded in the main chain's smart contract and bound to the data contributor's account one by one. It is continuously updated with the training rounds and evaluation scores. The integrity update refers to the following formula:

[0133]

[0134] Among them, h represents the current block time period, h-1 represents the previous block time period, It represents the integrity of participating node i in the previous block time period, and α represents the exponential decay coefficient. The larger α is, the higher the weight of the current cycle reward is, and the faster the historical contribution decays. Indicates the integrity of participating node i in the current block time period, reward i It represents the reward value obtained by participating node i in the local model training of the current block time period, and β represents a positive number. Represents the cumulative historical integrity score of participating node i.

[0135] Step S30: Obtain the gradient training results uploaded by each participating node to the federated learning blockchain.

[0136] It should be noted that the gradient training results are obtained by each participating node performing gradient training on the local face recognition model.

[0137] In some embodiments, during the model aggregation process, each participating node is assigned a value according to the formula Train the local model gradient by finding the minimum parameters Minimize the loss function Build a decentralized training architecture. Represents the data set x ik ,y ik ∈D i The main parameters trained. Each node propagates its gradient results to the entire network.

[0138] In some embodiments, each participating node may introduce a verifiable random function (VRF) to encrypt data and randomly select an aggregation node to ensure fairness in node selection.

[0139] In some embodiments, the key generation may be based on the formula (pk i ,sk i )←VRF.KeyGen(1 λ) Generate a public-private key pair for each participating node, (pk i ,sk i ) represent the public and private key pairs respectively, and λ is a security parameter; the cryptographic sorting algorithm calculates the priority of each node in the election for leader or committee member based on the private key of each node and a global seed random number and outputs a sorting certificate. The calculation formula is expressed as: (π i ,h i )←VRF.Eval(sk i ,seed|r i ). where h i Represents the priority hash value used for sorting, π i A verifiable credential.

[0140] In some embodiments, the face recognition device can verify whether the node priority is legal, and the verification formula is VRF.Verify(pk i ,seed|r i ,π i ,h i )→{0,1}, the formula outputs 1 if the verification is successful, and 0 if it is failed.

[0141] Step S40: Aggregating the gradient training results of each participating node according to the multi-dimensional quantization parameter to obtain a global gradient parameter.

[0142] In a specific implementation, the face recognition device can aggregate multi-dimensional quantitative parameters to obtain the aggregation weight of each participating node, and aggregate the gradient training results of each participating node based on the aggregation weight to obtain the global gradient parameter.

[0143] In some embodiments, the face recognition device can complete each round of training by sending the aggregate weight w' and risk weight Λ of the current round to the smart contract. i , marginal contribution and dynamic integrity weight Key parameters such as the value of the data contributors (nodes) are permanently written to the blockchain. This data not only provides a historical basis for weight calculations in subsequent rounds but also ensures the transparency and auditability of the entire aggregation process through the blockchain's immutable nature. Data contributors (nodes) are quantitatively scored based on their data quality and task completion rate, generating a reputation score that serves as cumulative proof of long-term contribution. This score, normalized by dividing by the maximum reputation value (max(r),) eliminates absolute numerical differences and establishes a relative evaluation system. This normalized result can be adjusted proportionally with changes in the highest reputation value, enabling rapid adjustments to priorities even with the addition of new nodes.

[0144] Furthermore, in order to accurately assign data weights to each participating node, the above step S40 may include:

[0145] Step S401: The face recognition device monitors abnormal behavior of each participating node;

[0146] Step S402: Mark participating nodes with abnormal behavior as abnormal nodes;

[0147] Step S403: adjusting the weight of the abnormal node according to the multi-dimensional quantization parameter to obtain a target weight vector;

[0148] Step S404: Aggregating the gradient training results of each participating node based on the target weight vector and the multi-dimensional quantization parameter to obtain a global gradient parameter.

[0149] It's important to note that within the distributed federated learning architecture, the collaborative paradigm among participating nodes utilizes a consensus-like mechanism, and its operational logic is comparable to the node verification system of a blockchain network. The reputation assessment system utilizes a three-state judgment mechanism: when a node generates positive feedback, its reputation weight increases; in steady-state judgment, the current evaluation coefficient is maintained; and if a negative event is detected, reputation decay is triggered.

[0150] In practice, to defend against malicious node attacks, facial recognition devices can preset a dynamic penalty protocol: when verifying abnormal behavior, first perform account permission suspension and incentive call-back operations, and then implement zeroing of reputation weight; if the violation is serious, initiate a ban on participating in federated learning penalties. Based on the Shapley value calculation formula, a comprehensive weight evaluation is performed to obtain the aggregate weight (i.e., the target weight vector), which is calculated based on the following formula:

[0151]

[0152] Where w' represents the target weight vector, which adjusts the contribution distribution ratio of each node (participant) in federated learning, combining the dynamic weight of integrity and Shapley value. The numerator represents the basic contribution weight of node i in the sub-alliance s, while taking into account its integrity weight Λ i , the marginal contribution is used in the denominator to ensure The adjusted weights w' sum to 1, achieving fair distribution.

[0153] Step S50: updating the local face recognition model based on the global gradient parameter, and inputting the target feature vector into the updated local face recognition model for face recognition.

[0154] In some embodiments, facial recognition devices use a lightweight convolutional network (CNN) for algorithm optimization. The CNN network's input layer is configured to adapt to the standardized image size (64×64 pixels). Two layers of convolution (kernel size 3×3, stride 1) and maximum pooling (kernel size 2×2) are defined for feature extraction. The fully connected layer output dimensions are aligned with the number of classification task categories to obtain the final features.

[0155] In some embodiments, the face recognition device may set a random hash value to map the participant's unique identification hash value to a random number (0, 1]. Based on the parameter values, the mapping priority is . Where represents the node reputation value, representing the maximum reputation value among all nodes. The final priority is determined by the reputation value and the random hash value.

[0156] In some embodiments, when a node i is selected as a verification node, the working node generates signed gradient data to verify the gradient information of other nodes. First, the system checks the marginal contribution of the participating nodes. If a dynamic threshold (such as the median contribution of all nodes in the current round) is reached, only nodes with qualified contributions are allowed to participate in gradient verification. A report is submitted based on the verification results. Working nodes are responsible for broadcasting gradient collection requests, sending signed gradients and metadata, and encapsulating and storing gradient blocks in the blockchain / database. Ultimately, all components collaborate to complete gradient verification and data storage tasks.

[0157] Once the newly generated block is verified by the majority of nodes, each node will Updates its local model to use the new global gradients in the verified block.

[0158] This embodiment uses multiple sets of pre-built Gabor filters to extract multi-scale features from the input face image to obtain a target feature vector. The Gabor filters are composed of multiple sets of complex filter kernels generated based on different directions and spatial frequencies. Each participating node is evaluated based on multiple evaluation dimensions to obtain multi-dimensional quantitative parameters. The evaluation dimensions include marginal contribution dimension, risk dimension and reputation dimension. The multi-dimensional quantitative parameters include marginal contribution degree, risk weight parameter and target reputation value. The gradient training results uploaded by each participating node to the federated learning blockchain are obtained. The gradient training results are the evaluation results of each participating node on the local The face recognition model is obtained by gradient training; the gradient training results of each participating node are aggregated according to the multi-dimensional quantization parameter to obtain a global gradient parameter; the local face recognition model is updated based on the global gradient parameter, and the target feature vector is input into the updated local face recognition model for face recognition; since this embodiment extracts the features of the input face image through multi-scale, thereby enhancing the representation ability of the image features, through multi-dimensional node evaluation, effectively balancing the data contribution weights provided by the distributed nodes, improving the convergence speed and recognition accuracy of the face recognition model, and greatly improving the accuracy of face recognition.

[0159] In addition, the embodiment of the present invention also proposes a face recognition system, referring to Figure 4 , Figure 4 : is a schematic diagram of the framework of a face recognition system in one embodiment, Figure 4 This paper presents a comprehensive framework for a face recognition system, aiming to improve the efficiency and fairness of distributed collaboration through model aggregation and incentive mechanisms. Its content can be divided into four parts: First, based on literature analysis and current status analysis, a local feature extraction scheme based on a convolutional neural network (CNN) is constructed. This scheme includes modules such as convolutional layers, GCS-LBP feature extraction (a hybrid approach that combines traditional texture features with deep learning), and pooling layers, enabling efficient data feature mining. Second, during the model aggregation phase, a blockchain-based federated learning (BCFL) framework is introduced, incorporating a verifiable random function (VRF) to ensure the security and randomness of inter-node communication and prevent malicious tampering. Furthermore, a fair incentive mechanism is designed to address the differences in contributions among participants in federated learning. This mechanism evaluates node contributions through risk weight adjustment and a modified Shapley value, balancing the impact of computing resources and data quality on the results, thereby optimizing collaborative fairness. Finally, through experimental verification, the robustness, convergence efficiency, and incentive rationality of the framework are systematically tested by setting up an environment, preprocessing data, and configuring multiple nodes, ensuring the compatibility of the theoretical design with practical scenarios. Overall, the framework of the face recognition system takes secure aggregation and fair incentives as its core innovations, taking into account the accuracy of feature extraction and the credibility of distributed collaboration, and provides methodological support for the application of federated learning in complex scenarios.

[0160] In addition, an embodiment of the present invention further provides a computer-readable storage medium, on which a face recognition program is stored. When the face recognition program is executed by a processor, the steps of the face recognition method described above are implemented.

[0161] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, systems or devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, system or device. The program code contained on the computer-readable storage medium may be transmitted using any appropriate medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0162] The computer-readable storage medium may be included in the face recognition device, or may exist independently without being incorporated into the face recognition device.

[0163] In addition, an embodiment of the present invention further provides a computer program product, including a face recognition program, which implements the steps of the face recognition method described above when executed by a processor.

[0164] The specific implementation of the computer program product of the present invention is basically the same as the embodiments of the above-mentioned face recognition method, and will not be repeated here.

[0165] Reference Figure 5 , Figure 5 This is a structural block diagram of an embodiment of a face recognition device of the present invention.

[0166] like Figure 5 As shown, the face recognition device proposed in the embodiment of the present invention is applied to a participating node in a federated learning blockchain, wherein the federated learning blockchain includes multiple participating nodes, and the face recognition device includes:

[0167] A multi-scale feature extraction module 10 is used to extract multi-scale features from the input face image using multiple sets of pre-built Gabor filters to obtain a target feature vector. The Gabor filters are composed of multiple sets of complex filter kernels generated based on different directions and spatial frequencies.

[0168] A multi-dimensional evaluation module 20 is configured to evaluate each participating node based on multiple evaluation dimensions to obtain multi-dimensional quantitative parameters, wherein the evaluation dimensions include a marginal contribution dimension, a risk dimension, and a reputation dimension, and the multi-dimensional quantitative parameters include a marginal contribution degree, a risk weight parameter, and a target reputation value;

[0169] A distributed gradient training module 30 is used to obtain the gradient training results uploaded by each participating node to the federated learning blockchain. The gradient training results are obtained by each participating node performing gradient training on the local face recognition model.

[0170] A gradient parameter aggregation module 40 is used to aggregate the gradient training results of each participating node according to the multi-dimensional quantization parameter to obtain a global gradient parameter;

[0171] The face recognition module 50 is configured to update the local face recognition model based on the global gradient parameter, and input the target feature vector into the updated local face recognition model for face recognition.

[0172] This embodiment uses multiple sets of pre-built Gabor filters to extract multi-scale features from the input face image to obtain a target feature vector. The Gabor filters are composed of multiple sets of complex filter kernels generated based on different directions and spatial frequencies. Each participating node is evaluated based on multiple evaluation dimensions to obtain multi-dimensional quantitative parameters. The evaluation dimensions include marginal contribution dimension, risk dimension and reputation dimension. The multi-dimensional quantitative parameters include marginal contribution degree, risk weight parameter and target reputation value. The gradient training results uploaded by each participating node to the federated learning blockchain are obtained. The gradient training results are the evaluation results of each participating node on the local The face recognition model is obtained by gradient training; the gradient training results of each participating node are aggregated according to the multi-dimensional quantization parameter to obtain a global gradient parameter; the local face recognition model is updated based on the global gradient parameter, and the target feature vector is input into the updated local face recognition model for face recognition; since this embodiment extracts the features of the input face image through multi-scale, thereby enhancing the representation ability of the image features, through multi-dimensional node evaluation, effectively balancing the data contribution weights provided by the distributed nodes, improving the convergence speed and recognition accuracy of the face recognition model, and greatly improving the accuracy of face recognition.

[0173] The facial recognition device provided in this application utilizes the facial recognition method described in the aforementioned embodiments to address the technical challenges of facial recognition. Compared to the prior art, the facial recognition device provided in this application achieves the same beneficial effects as the facial recognition method described in the aforementioned embodiments. Other technical features of the facial recognition device are the same as those disclosed in the aforementioned embodiments and are not further elaborated upon here.

[0174] It should be understood that the above is only an example and does not constitute any limitation to the technical solution of the present invention. In specific applications, those skilled in the art can make settings as needed, and the present invention does not impose any limitation on this.

[0175] It should be noted that the workflow described above is merely illustrative and does not limit the scope of protection of the present invention. In practical applications, technicians in this field can select part or all of it according to actual needs to achieve the purpose of the embodiment scheme, and no limitation is made here.

[0176] In addition, for technical details not fully described in this embodiment, please refer to the face recognition method provided in any embodiment of the present invention, and will not be repeated here.

[0177] It should be noted that, in this document, the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, article, or system comprising a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or system. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or system comprising the element.

[0178] The serial numbers of the above embodiments of the present invention are for description only and do not represent the advantages or disadvantages of the embodiments.

[0179] Through the description of the above embodiments, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus the necessary general hardware platform, and of course can also be implemented by hardware, but in many cases the former is a better embodiment. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as read-only memory / random access memory, magnetic disk, optical disk), and includes a number of instructions for enabling a terminal device (which can be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in each embodiment of the present invention.

[0180] The above are only preferred embodiments of the present invention and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention description and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A face recognition method, characterized in that: The face recognition method is applied to a participating node in a federated learning blockchain, wherein the federated learning blockchain includes multiple participating nodes. The face recognition method includes: Perform multi-scale feature extraction on the input face image using multiple sets of pre-built Gabor filters to obtain a target feature vector. The Gabor filters are composed of multiple sets of complex filter kernels generated based on different directions and spatial frequencies. Evaluate each participating node based on multiple evaluation dimensions to obtain multi-dimensional quantitative parameters, wherein the evaluation dimensions include marginal contribution dimension, risk dimension, and reputation dimension, and the multi-dimensional quantitative parameters include marginal contribution degree, risk weight parameter, and target reputation value; Obtaining the gradient training results uploaded by each participating node to the federated learning blockchain, where the gradient training results are obtained by each participating node performing gradient training on the local face recognition model; Aggregating the gradient training results of each participating node according to the multidimensional quantization parameter to obtain a global gradient parameter; The local face recognition model is updated based on the global gradient parameter, and the target feature vector is input into the updated local face recognition model for face recognition.

2. The face recognition method according to claim 1, wherein: The method of performing multi-scale feature extraction on the input face image by using multiple sets of pre-built Gabor filters to obtain a target feature vector includes: Perform grayscale processing on the input face image to obtain an initial grayscale image; Normalizing the initial grayscale image to obtain a candidate grayscale image; Generating corresponding complex filter kernels based on multiple sets of different directions and spatial frequencies, and generating multiple sets of Gabor filters based on the complex filter kernels; The candidate grayscale image is decomposed at multiple scales by using multiple groups of Gabor filters, and multiple groups of texture feature maps of different directions and spatial frequencies are extracted based on the filter output results. The wavelet function of the Gabor filter refers to the following formula: Among them, z represents the spatial coordinate value (x, y) of the pixel point in the input face image, k u,v represents the center frequency of the Gabor filter, represents the directionality of the two-dimensional Gabor filter, σ 2 Represents the square of the standard deviation of the Gaussian envelope, which is used to control the spatial distribution width of the Gabor filter. represents the Gaussian envelope term, which is used to limit the spatial distribution of the plane wave, exp(ik u,v z) represents the periodicity and directionality of the wavelet, Used for energy spectrum attenuation compensation; Feature analysis is performed on the multiple groups of texture feature maps with different directions and spatial frequencies to generate target feature vectors.

3. The face recognition method according to claim 2, wherein: The step of performing feature analysis on the plurality of groups of texture feature maps of different directions and spatial frequencies to generate a target feature vector includes: Traversing each pixel in the texture feature map and calculating the LBP value of each pixel; Generate an LBP feature map corresponding to each Gabor filter based on the LBP value of each pixel point in the texture feature map; Divide the LBP feature map into multiple sub-blocks, and generate a distribution histogram of the LBP value for each sub-block; The distribution histograms of each sub-block are spliced ​​together to generate the target feature vector.

4. The face recognition method according to any one of claims 1 to 3, characterized in that: The method of evaluating each participating node based on multiple evaluation dimensions to obtain multi-dimensional quantitative parameters includes: The marginal contribution of each participating node is evaluated to obtain the marginal contribution of each participating node. The marginal contribution is calculated based on the following formula: in, represents the marginal contribution of participating node i, s(i) represents the sub-alliance composed of all participating nodes including participating node i, v(s) represents the profit value of the sub-alliance, v(s / {i}) represents the profit value of the sub-alliance after excluding participating node i, w(|s|) represents the contribution weight of the sub-alliance, |s| represents the size of the sub-alliance, and n represents the total number of nodes in the sub-alliance; Based on the integrity of each participating node, the risk weight of each data supply node is evaluated to obtain the risk weight of each participating node. The risk weight is calculated based on the following formula: Among them, Λ i represents the risk weight of participating node i, represents the integrity of participating node i in the current block time period h, λ i represents the integrity vector of participating node i; Conduct a quantitative reputation assessment based on the data quality information and task completion rate of each participating node to obtain the initial reputation value of each participating node; The initial reputation value of each participating node is normalized based on the maximum reputation value of the participating nodes to obtain the target reputation value of each participating node.

5. The face recognition method according to claim 4, wherein: The risk weight assessment of each data supply node based on the integrity of each participating node is performed to obtain the risk weight of each participating node, including: Obtain the historical integrity of each participating node in the historical block time period; Evaluate the noise level of the data provided by each participating node and obtain data quality information of each participating node; generating an exponential decay coefficient for each participating node based on the data quality information; The historical credibility is updated according to the exponential decay coefficient to obtain an updated credibility. The credibility is updated according to the following formula: Among them, h represents the current block time period, h-1 represents the previous block time period, represents the integrity of participating node i in the previous block time period, α represents the exponential decay coefficient, Indicates the integrity of participating node i in the current block time period, reward i It represents the reward value obtained by participating node i in the local model training of the current block time period, and β represents a positive number. Represents the cumulative historical integrity score of participating node i; Based on the updated credibility, a risk weight assessment is performed on each data supply node to obtain the risk weight of each participating node.

6. The face recognition method according to claim 5, wherein: The step of aggregating the gradient training results of each participating node according to the multi-dimensional quantization parameter to obtain a global gradient parameter includes: Monitor abnormal behavior of each participating node; Mark participating nodes with abnormal behaviors as abnormal nodes; The abnormal nodes are weighted according to the multidimensional quantization parameters to obtain a target weight vector, which is calculated based on the following formula: Among them, w ' represents the target weight vector; The gradient training results of each participating node are aggregated based on the target weight vector and the multidimensional quantization parameter to obtain a global gradient parameter.

7. A face recognition device, characterized in that: The face recognition device is applied to a participating node in a federated learning blockchain, wherein the federated learning blockchain includes multiple participating nodes, and the face recognition device includes: A multi-scale feature extraction module is used to extract multi-scale features from the input face image using multiple sets of pre-built Gabor filters to obtain a target feature vector. The Gabor filters are composed of multiple sets of complex filter kernels generated based on different directions and spatial frequencies. A multi-dimensional evaluation module is used to evaluate each participating node based on multiple evaluation dimensions to obtain multi-dimensional quantitative parameters. The evaluation dimensions include marginal contribution dimension, risk dimension, and reputation dimension. The multi-dimensional quantitative parameters include marginal contribution degree, risk weight parameter, and target reputation value. A distributed gradient training module is used to obtain the gradient training results uploaded by each participating node to the federated learning blockchain. The gradient training results are obtained by each participating node performing gradient training on the local face recognition model. A gradient parameter aggregation module is used to aggregate the gradient training results of each participating node according to the multi-dimensional quantization parameter to obtain a global gradient parameter; The face recognition module is used to update the local face recognition model based on the global gradient parameter, and input the target feature vector into the updated local face recognition model to perform face recognition.

8. A face recognition device, characterized in that: The face recognition device includes: a memory, a processor, and a face recognition program stored in the memory and executable on the processor, wherein the face recognition program is configured to implement the face recognition method according to any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a face recognition program, which, when executed by a processor, implements the face recognition method according to any one of claims 1 to 6.

10. A computer program product, characterized in that The computer program product includes a face recognition program, and when the face recognition program is executed by a processor, the steps of the face recognition method according to any one of claims 1 to 6 are implemented.

Images