Digital e-commerce platform based on cloud computing

Abstract

This application discloses a cloud-based digital e-commerce platform, relating to the field of data processing, including: acquiring multiple frames of facial images; preprocessing the acquired multiple frames of facial images; using a pre-trained convolutional neural network to extract features from the preprocessed multiple frames of facial images to obtain feature vectors; calculating the similarity between the obtained feature vectors and feature vectors in a pre-stored authorized personnel facial feature database to obtain a similarity score; if the calculated similarity score is greater than a threshold, authentication is successful, allowing the user to access the platform; if the calculated similarity score is less than or equal to the threshold, authentication fails; when authentication fails, authentication is performed according to a pre-set second authentication method; addressing the low security of authentication in existing e-commerce platforms, this application employs a second authentication method combining voiceprint recognition and iris recognition, effectively improving authentication security.

Description

Technical Field

[0001] This application relates to the field of data processing, and in particular to cloud-based digital e-commerce platforms. Background Technology

[0002] With the rapid development of internet technology, e-commerce has become an important part of modern business activities. E-commerce platforms provide consumers with a convenient online shopping experience, while also offering merchants a broad sales channel. However, while e-commerce is booming, security issues such as identity fraud and account theft are becoming increasingly prominent, posing serious threats to user privacy and property security, and becoming a bottleneck restricting the further development of e-commerce.

[0003] Existing e-commerce platforms typically use simple username and password authentication methods, which are easily cracked and have low security. Although some platforms have introduced auxiliary authentication methods such as SMS verification codes and hardware keys, they still cannot fundamentally solve the problem of identity theft. Furthermore, existing authentication methods are mostly single-modal, relying solely on a single biometric feature such as face or fingerprint for verification. This single-modal approach lacks reliability and struggles to adapt to complex and ever-changing fraud scenarios. Therefore, there is an urgent need for a multimodal, hierarchical, and integrated authentication solution to comprehensively protect the account and transaction security of e-commerce platforms from multiple angles. Summary of the Invention

[0004] To address the issue of low authentication security in existing e-commerce platforms, this application provides a cloud-based digital e-commerce platform that uses facial recognition as the primary authentication method. When the primary authentication fails, a secondary authentication method combining voiceprint and iris recognition is activated, enabling hierarchical authentication and effectively improving authentication security.

[0005] The purpose of this application is achieved through the following technical solution.

[0006] This specification provides a cloud-based digital e-commerce platform, including: acquiring multiple frames of facial images; preprocessing the acquired multiple frames of facial images; preferably, the image preprocessing further includes: performing illumination equalization processing on the multiple frames of facial images using an adaptive histogram equalization algorithm to reduce the impact of illumination changes on facial feature extraction; performing pose correction on the illumination-equalized multiple frames of facial images using affine transformation to adjust the face to a frontal pose, obtaining a pose-corrected facial image; performing liveness detection on the pose-corrected facial image using a liveness detection algorithm based on blink frequency and facial micro-expression changes to determine whether it is a real face; detecting the blink frequency in the facial region, and if the blink frequency is lower than a set threshold, it is determined to be non-live; analyzing the change features of facial micro-expressions, extracting the motion trajectory of key facial points such as the corners of the mouth and eyebrows, and constructing a facial micro-expression feature vector; inputting the facial micro-expression feature vector into a support vector machine classifier to identify whether it is a natural expression change, and if it is a non-natural expression change, it is determined to be non-live. A pre-trained convolutional neural network is used to extract features from pre-processed multi-frame face images to obtain feature vectors. The convolutional neural network adopts a residual network structure and introduces an attention mechanism to enhance the network's attention to key feature regions such as eyes, nose, and mouth, thereby improving feature representation capabilities. Furthermore, an image quality assessment module is added before the convolutional neural network to remove images that do not meet the standards due to lighting, blurriness, occlusion, etc., thereby improving the accuracy of subsequent recognition.

[0007] The obtained feature vectors are compared with the feature vectors in the pre-stored authorized personnel face feature database to calculate a similarity score. Preferably, the cosine similarity of the extracted feature vectors with multiple feature vectors in the face feature database is calculated to obtain a set of similarity scores. The set of similarity scores is then weighted and averaged, with the weights determined based on the quality score of each feature vector during registration, to obtain the final similarity score. If the calculated similarity score is greater than a threshold, authentication is successful, and the user is allowed to access the platform. If the calculated similarity score is less than or equal to the threshold, authentication fails. When authentication fails, a pre-set second authentication method is used for verification. Preferably, a distributed storage architecture is adopted, dividing the face feature sub-database among multiple service nodes to achieve load balancing and parallel processing. For each sub-database, an index is generated using algorithms such as Locality Sensitive Hashing to accelerate large-scale face feature retrieval. A master-slave hot standby mechanism is established to ensure the high availability and data consistency of the face feature database.

[0008] In this context, a feature vector represents the features of an object or sample as a vector. In pattern recognition and machine learning, image and speech data are often converted into feature vectors as input to algorithms. Feature vectors extract key information from the data, making similar samples appear closer together in the feature space. A deep neural network (DNN) is an artificial neural network model containing multiple hidden layers. Through hierarchical feature learning, DNNs can automatically extract high-level semantic representations of data. In tasks such as speech recognition and image classification, DNNs have achieved groundbreaking progress due to their powerful modeling capabilities. Common DNN models include convolutional neural networks (CNNs) and recurrent neural networks (RNNs).

[0009] Furthermore, the second authentication method includes: extracting user voice data, performing voiceprint recognition using a deep neural network model to obtain voice recognition similarity; acquiring user iris images, extracting iris features, calculating the similarity with pre-stored iris features to obtain iris recognition similarity; inputting the voice recognition similarity and iris recognition similarity into a Bayesian network fusion algorithm to obtain the user's identity verification result; preferably, constructing a dynamic Bayesian network model, introducing time series factors on the basis of a static Bayesian network, using the voice recognition similarity and iris recognition similarity verified multiple times as input, to model the dynamic change process of authentication; each time slice of the dynamic Bayesian network includes a voice recognition similarity node, an iris recognition similarity node, a voice credibility node, an iris credibility node, and an authentication result node, with time slices connected by state transition probabilities; based on the user's historical verification data, the expected maximization (EM) algorithm is used to learn the parameters of the dynamic Bayesian network, including the conditional probability distribution and state transition probability distribution of each node; wherein, the forward-backward algorithm is used to calculate the expected step in the EM algorithm, utilizing Junction The Tree algorithm is used for inference calculations. During identity verification, the similarity scores of speech recognition and iris recognition from multiple consecutive user interactions are obtained to construct an observation sequence. Then, the Viterbi algorithm is used to infer on a dynamic Bayesian network to find the most likely hidden state sequence, i.e., the dynamic change process of identity authenticity. Based on the identity authenticity probability at the final moment, combined with a global threshold and decision rules, it is determined whether the user's identity has passed verification. The decision rules comprehensively consider factors such as the global threshold, historical verification results, and suspicious user behavior, and adopt a weighted voting mechanism to adaptively adjust the standard for identity verification. If the identity verification is successful, the verification data is added to the user's historical dataset, and the parameters of the dynamic Bayesian network are updated using an incremental learning algorithm to achieve online adaptive optimization of the model. Otherwise, user access is denied, and an early warning mechanism is activated to record suspicious behavior for subsequent security analysis. Based on the user's identity verification results, user access to the platform is controlled.

[0010] The iris, a ring-shaped structure with rich colors in the human eye, contains abundant detailed information such as spots and grooves. Iris features refer to a series of numerical indicators extracted from iris images to characterize the visual attributes of iris texture. Commonly used iris features include Gabor wavelet features and Local Binary Pattern (LBP) features. Iris features are unique and stable, and are widely used in iris recognition. A Bayesian network is a reasoning framework based on a probabilistic graphical model. It uses a directed acyclic graph to represent the dependencies between variables, with nodes representing random variables and edges representing conditional dependencies. Bayesian network fusion algorithms utilize Bayesian networks to fuse multi-source information. By modeling the causal relationships between features and comprehensively considering the degree of support of each piece of evidence for the hypothesis, a final decision is reached.

[0011] Furthermore, obtaining speech recognition similarity includes: collecting user speech data; denoising the collected user speech data using a speech denoising algorithm based on a deep neural network; performing endpoint detection and silence removal on the denoised speech data to obtain preprocessed speech data; extracting linear prediction cepstral coefficient features from the preprocessed speech data using linear predictive analysis; extracting formant frequency features from the preprocessed speech data using a formant estimation algorithm; concatenating the extracted linear prediction cepstral system features and formant frequency features to obtain a voiceprint feature vector; and using the obtained voiceprint feature vector as input, employing a pre-trained deep neural network model to obtain speech similarity. The voiceprint feature vector is input into a pre-trained deep neural network model. Through forward propagation of the network, the similarity to pre-stored voiceprints is calculated and used as the speech recognition similarity. The deep neural network model employs an attention mechanism and a connection-temporal classification (CTC) loss function to improve the accuracy and robustness of voiceprint recognition. Simultaneously, an adversarial learning strategy is introduced during model training, using a generative adversarial network (GAN) to synthesize highly realistic fake voiceprint samples, enhancing the model's resistance to deception. The input to the deep neural network model is the voiceprint feature vector, and the output is the first verification similarity. The network structure of the deep neural network model includes multiple convolutional layers, pooling layers, fully connected layers, and a Softmax classification layer. The deep neural network model is obtained through pre-training, which includes using a large amount of user speech data as training samples, optimizing network parameters using the backpropagation algorithm and stochastic gradient descent algorithm, minimizing the recognition loss function, and obtaining the optimal model parameters. Preferably, when calculating speech recognition similarity, not only the static similarity of voiceprint features is considered, but also the similarity measure of dynamic voiceprint features is introduced; a sequence matching algorithm based on dynamic time bending (DTW) is used to calculate the optimal alignment distance between two voiceprint feature sequences to characterize the dynamic change characteristics of voiceprints in the time dimension; the static similarity and dynamic similarity are weighted and fused to obtain the final speech recognition similarity.

[0012] Speech denoising refers to the process of removing background noise from noisy speech signals to restore clean speech. Speech denoising algorithms model the statistical characteristics of speech and noise, adaptively adjusting filter parameters to suppress noise components. Commonly used speech denoising algorithms include spectral subtraction, Wiener filtering, and statistical models such as MMSE. Deep learning methods, such as convolutional denoising autoencoders, are also used for speech denoising tasks. Endpoint detection refers to automatically determining the start and end positions of effective speech segments in a speech signal. Silence removal removes silent segments from a speech signal, retaining only segments containing speech content. Both are common steps in speech preprocessing, aiming to extract effective speech information and reduce data redundancy. Features such as short-time energy and zero-crossing rate are often used for endpoint detection. Methods based on energy thresholds or statistical models can be used for silence removal.

[0013] Furthermore, linear predictive analysis is used to extract linear predictive cepstral coefficient features from the preprocessed speech data. This includes: performing frame segmentation on the preprocessed speech data to obtain a speech frame sequence; applying a Hamming window function to the speech frame sequence to obtain windowed speech frames; calculating the autocorrelation function on the windowed speech frames to obtain autocorrelation function coefficients; calculating linear predictive coefficients using the Levinson-Durbin recursive algorithm based on the autocorrelation function coefficients; performing a fast Fourier transform on the linear predictive coefficients to obtain frequency domain predictive coefficients; taking the logarithm and inverse Fourier transform on the frequency domain predictive coefficients to obtain cepstral coefficients; and extracting multidimensional cepstral coefficients as linear predictive cepstral coefficient features.

[0014] Preferably, the linear predicted cepstral coefficient features are normalized to obtain normalized linear predicted cepstral coefficient features; wherein, the normalization process uses the Zero Mean and Unit Variance (ZMUV) algorithm; the mean and standard deviation of the linear predicted cepstral coefficient features are calculated; the mean is subtracted from the linear predicted cepstral coefficient features and then divided by the standard deviation to obtain the normalized features; the normalized linear predicted cepstral coefficient features are compensated for by channel length using a channel length normalization algorithm to obtain channel length normalized linear predicted cepstral coefficient features; the channel length parameter of the normalized linear predicted cepstral coefficient features is estimated; based on the channel length parameter, the normalized linear predicted cepstral coefficient features are subjected to frequency scaling transformation to compensate for channel length differences; the channel length normalized linear predicted cepstral coefficient features are subjected to discrete cosine transform (DCT) to obtain the DCT coefficients of the cepstral features; and the low-order coefficients in the DCT coefficients of the cepstral features are extracted as the final linear predicted cepstral coefficient features.

[0015] Linear predictive analysis (LDE) is a speech signal processing technique that establishes a linear combination relationship between the current speech sample and several historical samples through short-time analysis of the speech signal, thereby achieving linear prediction of the speech signal. LDE coefficients characterize the formant properties of the speech signal and have important applications in speech coding, speech synthesis, and other fields. The autocorrelation function describes the correlation between sample values ​​of a signal at different times. For a discrete signal x(n), its autocorrelation function R(k) is defined as: R(k) = ∑x(n)x(nk), representing the sum of the product of the signal and its k delayed samples. The autocorrelation function has properties such as symmetry and a maximum value at the origin. The short-time autocorrelation function of speech is often used to extract features such as pitch and formants. The Levinson-Durbin algorithm is an efficient recursive algorithm for solving linear prediction coefficients. Given the autocorrelation function of the speech signal, this algorithm recursively calculates the reflection coefficient and linear prediction coefficients, avoiding the complexity of solving matrix equations. The Levinson-Durbin algorithm guarantees the stability of the prediction coefficients, ensuring they meet the requirements of a minimum-phase system.

[0016] Furthermore, a formant estimation algorithm is used to extract formant frequency features from the preprocessed speech data, including: calculating the linear prediction spectrum based on the obtained linear prediction coefficients; smoothing the linear prediction spectrum using a triangular window function to obtain a smoothed linear prediction spectrum; obtaining the local maxima of the smoothed linear prediction spectrum as the initial formant frequencies; performing harmonic detection on the initial formant frequencies to remove harmonic peaks and obtain optimized formant frequencies; obtaining the top N formant frequencies with the highest energy among the optimized formant frequencies as formant frequency features; performing a discrete Fourier transform on the formant frequency features to obtain frequency domain formant features; filtering the frequency domain formant features using a triangular Mel filter bank to obtain Mel frequency domain formant features; performing logarithmic and discrete cosine transforms on the Mel frequency domain formant features to obtain the MFCC features of the formants; and concatenating the MFCC features of the formants with the corresponding first-order and second-order difference parameters to obtain the final formant frequency features.

[0017] The Fourier transform converts a time-domain signal to a frequency-domain signal, while the inverse Fourier transform restores the frequency-domain signal to the time-domain signal. For a frequency-domain signal X(f), its inverse Fourier transform is: x(t)=∫X(f)e j2πftThe inverse Fourier transform (IF) is commonly used in frequency domain filtering and spectral analysis. In speech processing, the IF transform is used to convert frequency domain prediction coefficients into time domain linear prediction coefficients. Cepstral analysis obtains cepstral coefficients by taking the logarithm of the speech signal's spectrum and then performing an IF transform. Multidimensional cepstral coefficients refer to cepstral coefficients extracted from multiple dimensions (such as time and frequency) to comprehensively characterize the spectral envelope characteristics of the speech signal. Commonly used multidimensional cepstral coefficients include Mel-frequency cepstral coefficients (MFCC) and perceptual linear predictive coefficients (PLP). A triangular window is a time-domain windowing function with a linearly ascending and descending triangular shape. Its mathematical definition is: Where N is the window length. The triangular window's frequency domain characteristics are low side lobes and a wide main lobe. In speech processing, triangular windows are often used to smooth the speech spectrum and suppress rapid spectrum changes. A local maximum is a point where a function reaches its maximum value within a local range. For a discrete sequence x(n), if x(n) > x(n-1) and x(n) > x(n+1), then x(n) is a local maximum. In speech processing, local maxima are often used to detect features such as formants and pitch period in speech signals. Harmonics are sine waves whose frequencies are integer multiples of the fundamental frequency. Speech signals consist of a fundamental frequency and harmonics, and their spectrum shows concentrated energy peaks at the fundamental and harmonic frequencies, called harmonic peaks. Harmonic peaks reflect the periodic structure of speech but can interfere with formant estimation. The influence of harmonic peaks needs to be removed during formant extraction. The Discrete Fourier Transform (DFT) is a Fourier transform for discrete signals, converting time-domain signals into frequency-domain signals. For a discrete signal x(n) of length N, its DFT is: k = 0, 1, ..., N-1.

[0018] The Discrete Fourier Transform (DFT) decomposes a signal into sine waves of different frequencies, obtaining a spectral representation of the signal. The Fast Fourier Transform (FFT) is an efficient implementation algorithm of the DFT. The Mel filter bank is a set of triangular bandpass filters whose center frequencies are arranged according to the Mel frequency scale. The Mel frequency takes into account the nonlinear characteristics of human hearing, with high resolution in the low-frequency range and low resolution in the high-frequency range. The triangular Mel filter bank extracts the frequency components relevant to perception by weighted averaging the energy of the speech spectrum, obtaining the Mel spectrum. The logarithmic transform converts the amplitude spectrum of the Mel spectrum into a logarithmic spectrum, compressing the dynamic range of the amplitude to make it closer to the response characteristics of the human ear. The Discrete Cosine Transform (DCT) is an orthogonal transform in the real domain, converting the logarithmic Mel spectrum into the cepstral domain, obtaining the Mel frequency cepstral coefficients (MFCCs). The DCT has excellent properties such as energy concentration and decorrelation, making MFCCs a commonly used feature in speech recognition. Differential parameters characterize the dynamic changes of speech features and are often used as a supplement to static features. Let the static feature vector be x(n). The first-order difference Δx(n) = x(n+1) - x(n) reflects the instantaneous rate of change of the feature; the second-order difference ΔΔx(n) = Δx(n+1) - Δx(n) reflects the trend of the rate of change of the feature. The difference parameters are generally estimated by methods such as linear regression and are then combined with the original features for use.

[0019] Furthermore, the iris recognition similarity is obtained by: using a convolutional neural network-based iris detection algorithm to locate the iris in the iris image and obtain the iris region; using an image segmentation algorithm to segment the iris region and obtain the segmented iris image; wherein, the image segmentation algorithm obtains the globally optimal segmentation curve by constructing an energy function of the iris region and then performs iris segmentation; using the Gabor wavelet transform method to extract the iris texture features of the segmented image; using a Zernike moment-based method to extract the iris shape features of the segmented image; using a Markov random field-based method to extract the iris texture statistical features of the segmented image; fusing the extracted iris texture features, iris shape features, and iris texture statistical features to obtain the iris feature vector; and using the iris feature vector as input, calculating the iris recognition similarity using the Hamming distance metric.

[0020] Iris detection refers to the process of automatically locating the iris region from an eye image. Convolutional neural network-based iris detection algorithms utilize deep learning methods to automatically extract salient features of the iris region by training a convolutional neural network model, achieving precise iris localization. This algorithm overcomes the sensitivity of traditional methods to image quality and lighting changes through end-to-end feature learning. Image segmentation refers to the process of dividing an image into several regions with specific meanings and extracting the target of interest. Image segmentation algorithms achieve semantic image segmentation by defining region division criteria and optimizing region boundaries. Common image segmentation algorithms include thresholding, region growing, edge detection, and graph-based segmentation. In iris recognition, image segmentation is used to extract a clear iris texture image from the iris region. Gabor wavelets are a local frequency analysis tool that can simultaneously obtain local frequency and phase information of a signal in the time-frequency domain. Two-dimensional Gabor wavelets are obtained by modulating a Gaussian function and have directional selectivity and multi-scale characteristics. Gabor wavelet transform extracts multi-directional, multi-scale texture features from an image by convolving it with a set of Gabor wavelets of different scales and orientations, and has important applications in iris texture feature extraction. Zernike moments are orthogonal moment functions defined by the product of a Zernike polynomial and a radial polynomial. Zernike moments possess advantages such as rotation invariance, noise robustness, and low information redundancy, and are commonly used for image shape description and recognition. Methods based on Zernike moments extract global features related to iris shape by calculating the Zernike moments of the iris image, which are then used for feature representation in iris recognition. Markov random fields are probabilistic graphical models that describe spatial context dependencies. By defining the local dependency structure between pixels, they characterize the texture properties of an image. Methods based on Markov random fields model the iris image as a Markov random field, and by defining the interactions between pixels and the clustering potential function, extract statistical features of iris texture, such as the gray-level co-occurrence matrix and local binary patterns, for iris feature representation.

[0021] Furthermore, an image segmentation algorithm is used to segment the iris region, obtaining a segmented iris image. The image segmentation algorithm constructs an energy function for the iris region to obtain the globally optimal segmentation curve, and then performs iris segmentation. This includes: representing the iris region as a weighted undirected graph, where the nodes are pixels in the iris region and the edge weights are the similarity between adjacent pixels; constructing an energy function for the weighted undirected graph, which includes a region term and a boundary term; where the region term represents the probability that a pixel belongs to the iris region or a non-iris region, and the boundary term represents the probability that adjacent pixels have consistent labels; minimizing the energy function using a maximum flow minimum cut algorithm to obtain the globally optimal segmentation curve; and segmenting the iris region according to the globally optimal segmentation curve to obtain the segmented iris image; where the area inside the segmentation curve is the iris region, and the area outside is the non-iris region.

[0022] Weighted undirected graphs are a data structure in graph theory, consisting of nodes and weighted edges connecting them. In image segmentation, image pixels are considered as nodes of a graph, and the similarity between pixels is used as the weights of the edges, constructing a weighted undirected graph representation of the image. By defining an energy function on the graph, the image segmentation problem can be transformed into solving the optimal partitioning problem. The energy function is the objective function describing the image segmentation problem, used to measure the quality of the segmentation result. The energy function for image segmentation typically includes two terms: a region term and a boundary term. The region term measures the similarity of pixels within a segmented region, encouraging consistent pixel features within the region; the boundary term measures the smoothness of the segmented region boundary, encouraging the boundary to coincide with image edges. By minimizing the energy function, a globally optimal image segmentation result can be obtained. The maximum flow minimum cut algorithm is a classic algorithm for solving the minimum cut problem of a graph. In graph theory, the maximum flow problem refers to finding the maximum flow between the source and sink vertices; the minimum cut problem refers to finding the minimum cut set that partitions the graph into two parts. The maximum flow minimum cut theorem proves the equivalence of the two problems. In image segmentation, a flow network is constructed for the image, and the global optimal segmentation is solved using the maximum flow minimum cut algorithm, dividing the image into foreground and background regions.

[0023] Furthermore, the Gabor wavelet transform method is used to extract iris texture features from the segmented iris image, including: constructing a Gabor wavelet basis function, where the parameters of the Gabor wavelet include scale, orientation, center frequency, and bandwidth; performing radial normalization on the segmented iris image to convert the circular iris image into a rectangular image; performing convolution operation on the obtained rectangular image using the constructed Gabor wavelet basis function to obtain Gabor wavelet transform coefficients; taking the modulus of the Gabor wavelet transform coefficients to obtain a Gabor wavelet energy image; dividing the Gabor wavelet energy image into blocks, calculating the mean and variance of each block to obtain block Gabor features; and combining the block Gabor features according to scale and orientation to obtain the iris texture feature vector.

[0024] Radial normalization refers to the process of unfolding a circular iris image into a rectangular image. Because the iris image is affected by factors such as pupil dilation and camera distance, the iris radius may vary. Radial normalization maps the iris image to a fixed-size rectangular image by interpolating or resampling it radially, achieving a normalized representation of the iris size and shape, facilitating subsequent feature extraction. Convolution is a fundamental operation in signal processing used to extract local features of a signal. For two-dimensional image signals, convolution involves weighted summation of the image using a sliding convolution kernel to obtain a convolution response map. In Gabor wavelet transform, convolution is used to calculate the similarity between an image and Gabor wavelet basis functions at different scales and directions, extracting multi-scale, multi-directional texture features of the image. The Gabor wavelet energy image is the modulus representation of the Gabor wavelet transform coefficients, reflecting the energy distribution of the image at different scales and directions. Since the Gabor wavelet transform coefficients are complex numbers, the modulus operation converts them to real numbers, resulting in the Gabor wavelet energy image. Gabor wavelet energy images characterize the intensity and directionality of local texture in an image, making them an effective method for representing texture features.

[0025] Furthermore, a Zernike moment-based method is used to extract iris shape features from the segmented iris image, including: normalizing the coordinates of the segmented iris image by converting the Cartesian coordinate system to the polar coordinate system to obtain a normalized iris image; representing the normalized iris image as a radial distance function in the polar coordinate system; wherein the radial distance function represents the distance from the iris boundary to the iris center as a function of the angle; expanding the radial distance function using Zernike polynomial basis functions to obtain Zernike moment coefficients; wherein the Zernike polynomial basis functions are a set of orthogonal basis functions used to describe the radial distance function; and selecting preset low-order Zernike moment coefficients as the iris shape feature vector; wherein the preset low-order coefficients correspond to the global features of the iris shape.

[0026] The radial distance function is a mathematical representation of the shape of the iris boundary. In polar coordinates, it represents the distance from the iris boundary to the iris center as a function of angle. By transforming the iris image from Cartesian to polar coordinates, the radial distance function of the iris boundary can be obtained. The radial distance function characterizes the shape features of the iris boundary and is fundamental to iris shape modeling. Zernike moments are an image moment representation method based on orthogonal Zernike polynomial basis functions. Through Zernike polynomial expansion, the radial distance function can be represented as a series of Zernike moment coefficients. Lower-order Zernike moments correspond to global shape features of the image, such as translation, scaling, and rotation, while higher-order moments correspond to local shape details. Presetting lower-order Zernike moment coefficients means selecting a set of Zernike moments of a fixed order (e.g., the first 10 orders) as iris shape features, which effectively represents the iris shape while reducing feature dimensionality.

[0027] Furthermore, a Markov random field-based method is used to extract iris texture statistical features from the segmented iris image, including: dividing the segmented iris image into multiple overlapping local image patches; establishing a Markov random field model for each local image patch; wherein pixels in the image patch are field points, and the correlation between pixels is the interaction between field points; defining the energy function of the Markov random field model using the Gibbs distribution; wherein the energy function consists of pixel gray-level distribution and inter-pixel correlation; estimating the parameters of the energy function using the maximum likelihood estimation method to obtain the Markov random field model parameters; statistically analyzing the Markov random field model parameters to obtain the texture statistical features of the local image patches; wherein the texture statistical features include mean, variance, and entropy; and concatenating the texture statistical features of all local image patches to obtain the iris texture statistical feature vector.

[0028] The Gibbs distribution is a probability distribution used in statistical mechanics to describe the equilibrium state of a particle system, and it is also commonly used to define the energy function of a Markov random field. In the Markov random field model, the Gibbs distribution defines the joint probability distribution of field configurations, and its probability is exponentially related to the system energy. For a field configuration x, its Gibbs probability distribution is: Where E(x) is the energy function, T is the temperature parameter, and Z is the partition function. The Gibbs distribution has the local Markov property, meaning that the state of a field point depends only on the state of its neighboring field points. By defining a suitable energy function, the Gibbs distribution can characterize the correlation and constraint relationships between field points, and can be used for image texture modeling. Maximum likelihood estimation is a commonly used parameter estimation method used to estimate the parameters of a probabilistic model based on observed data. Assuming that the observed data follows a certain probability distribution, maximum likelihood estimation solves for the estimated values ​​of the probability distribution parameters by maximizing the likelihood function of the observed data. For Markov random field models, maximum likelihood estimation is used to estimate the energy function parameters of the Gibbs distribution. By constructing the likelihood function of local image patches, the energy function parameters are solved using optimization algorithms such as gradient descent and iterative reweighted least squares, thus obtaining the parameter estimates of the Markov random field model. Maximum likelihood estimation has excellent properties of asymptotic consistency and asymptotic efficiency. Feature concatenation refers to the process of connecting multiple feature vectors into a high-dimensional feature vector in a certain order. In iris texture statistical feature extraction, the iris image is divided into multiple local image patches, and the texture statistical features of each patch, such as mean, variance, and entropy, are extracted separately. To obtain the texture feature representation of the entire iris image, the feature vectors of all local image patches need to be concatenated to form a feature vector of length n*d, where n is the number of image patches and d is the feature dimension of each image patch. Feature concatenation integrates local texture features into global features, improving the robustness and discriminativeness of the feature representation. The concatenated high-dimensional feature vector can be used for subsequent feature selection, dimensionality reduction, and classifier design.

[0029] Compared to existing technologies, the advantages of this application are:

[0030] By employing facial feature extraction and similarity calculation based on convolutional neural networks, the effects of variations in facial pose and lighting can be overcome, achieving robust and efficient facial identity verification. When facial verification fails, dual biometric verification using voiceprint and iris scans is introduced, forming multimodal biometric authentication, which significantly improves the security of identity verification.

[0031] In voiceprint recognition, deep neural networks are used for speech denoising, which can effectively remove environmental noise. By extracting cepstral coefficient features through linear predictive analysis and using formant estimation algorithms to obtain vocal tract frequency features, the individual differences in speech can be accurately characterized. By fusing cepstral coefficient and formant frequency features, a more comprehensive and fine-grained voiceprint feature representation can be obtained.

[0032] In iris recognition, convolutional neural networks are used for iris localization, which can accurately obtain the iris region; iris segmentation is achieved through graph cut algorithm, which can overcome the local optimum problem of traditional methods and obtain the global optimum segmentation result; iris texture, shape and statistical features are extracted by Gabor wavelet, Zernike moment and Markov random field, which can characterize the uniqueness of the iris from multiple levels and angles and obtain a highly recognizable iris feature representation.

[0033] Using Bayesian networks to fuse the similarity between voiceprint recognition and iris recognition can fully leverage the complementarity of the two biometric features, overcome the shortcomings of single features, and improve the reliability and accuracy of identity verification. Attached Figure Description

[0034] This specification will be further described by way of exemplary embodiments, which will be described in detail with reference to the accompanying drawings. These embodiments are not limiting; in these embodiments, the same reference numerals denote the same structures, wherein:

[0035] Figure 1 This is an exemplary flowchart of a cloud-based digital e-commerce platform according to some embodiments of this specification;

[0036] Figure 2 This is an exemplary flowchart of second authentication as shown in some embodiments of this specification;

[0037] Figure 3 This is an exemplary flowchart illustrating the acquisition of speech similarity according to some embodiments of this specification;

[0038] Figure 4 This is an exemplary flowchart illustrating the acquisition of linear prediction cepstral coefficient features according to some embodiments of this specification;

[0039] Figure 5 This is an exemplary flowchart illustrating the acquisition of formant frequency characteristics according to some embodiments of this specification;

[0040] Figure 6 This is an exemplary flowchart illustrating the acquisition of iris recognition similarity according to some embodiments of this specification;

[0041] Figure 7 This is an exemplary flowchart illustrating the acquisition of segmented iris images according to some embodiments of this specification;

[0042] Figure 8 This is an exemplary flowchart illustrating the acquisition of iris texture feature vectors according to some embodiments of this specification;

[0043] Figure 9This is an exemplary flowchart illustrating the acquisition of iris shape feature vectors according to some embodiments of this specification;

[0044] Figure 10 This is an exemplary flowchart illustrating the acquisition of texture statistical features according to some embodiments of this specification. Detailed Implementation

[0045] The methods and systems provided in the embodiments of this specification will now be described in detail with reference to the accompanying drawings.

[0046] Figure 1 This is an exemplary flowchart of a cloud-based digital e-commerce platform according to some embodiments of this specification. The cloud-based digital e-commerce platform includes: acquiring multiple frames of facial images; preprocessing the acquired multiple frames of facial images; extracting features from the preprocessed multiple frames of facial images using a pre-trained convolutional neural network to obtain feature vectors; calculating the similarity between the obtained feature vectors and feature vectors in a pre-stored authorized personnel facial feature database to obtain a similarity score; if the calculated similarity score is greater than a threshold, authentication is successful, allowing the user to access the platform; if the calculated similarity score is less than or equal to the threshold, authentication fails; when authentication fails, authentication is performed according to a pre-set second authentication method.

[0047] When a user visits an e-commerce platform, the system captures multiple frames of the user's face and transmits them to a cloud server. When a user accesses the platform via a PC, mobile device, or other terminal, the system automatically uses the front-end device's camera to capture consecutive multiple frames of the user's face. To ensure the quality of the captured facial images, the system can guide the user to adjust their head posture so that their face is directly facing the camera and control the lighting conditions to meet preset requirements. The capture process lasts 2 to 3 seconds, with the camera continuously capturing 60 to 90 frames of facial images at a rate of 30 frames per second. The system encodes and compresses the captured raw image frames and securely transmits them to the cloud authentication server via network protocols such as HTTP and TCP. Simultaneously, the original captured images are deleted from the local terminal to protect user privacy.

[0048] The cloud server preprocesses the received multi-frame face images and extracts facial feature vectors using a pre-trained convolutional neural network model. After receiving the multi-frame face images, the cloud server first performs a quality assessment, discarding frames with insufficient lighting, skewed images, or blurry images. Then, the selected acceptable image frames undergo uniform size normalization, scaling the images to 256×256 pixels. Next, an adaptive histogram equalization algorithm is used to normalize the lighting, reducing the impact of lighting variations on feature extraction. Finally, the preprocessed images are input into a pre-trained deep convolutional neural network for feature extraction. This neural network uses the ResNet-18 architecture, containing 18 convolutional and fully connected layers, and can extract 4096-dimensional facial feature vectors. The network model parameters are pre-trained using a large-scale face dataset, demonstrating strong feature representation capabilities. For each frame of the image, the network outputs a 4096-dimensional feature vector. The feature vectors of multiple frames are fused together by mean pooling to form a final 4096-dimensional face feature vector, which serves as the feature representation for user authentication.

[0049] The cloud server calculates the similarity score by comparing the extracted facial feature vector with a pre-stored database of authorized user facial features. Specifically, it calculates a similarity score by comparing the extracted 4096-dimensional facial feature vector with the pre-stored database of authorized user facial features. The database stores the facial feature vectors of all registered users, with each user corresponding to a 4096-dimensional feature vector. The database uses relational databases such as MySQL for efficient and secure storage and management. The similarity calculation uses cosine similarity, which measures the cosine similarity between the facial feature vector to be verified and the feature vector of each user in the database. A higher cosine similarity indicates a closer similarity between the two feature vectors, and a higher probability of a match. Let the facial feature vector to be verified be A, and the feature vector of a user in the database be B. The formula for calculating the cosine similarity between the two vectors is: Where A⊙B represents the inner product of vectors A and B, and ||A|| and ||B|| represent the L2 norms of vectors A and B, respectively. To improve feature retrieval efficiency and reduce computational complexity under massive user scale, the system adopts a feature dimensionality reduction method based on PCA (Principal Component Analysis) to reduce the 4096-dimensional feature vector to 128 dimensions, thereby accelerating subsequent similarity calculation. After calculating the cosine similarity between the features of the face to be verified and the features of all registered users, the maximum similarity value is taken as the identity similarity score of the current face, and the corresponding user ID is recorded. Let the maximum similarity score be max_score, and the corresponding user ID be max_user_id, then the current face authentication result can be represented as a binary tuple of (max_user_id, max_score). The system returns this authentication result to the e-commerce platform for subsequent identity verification and authorization judgment. To balance security and user experience, the system presets a similarity score threshold. If `max_score` > `threshold`, authentication is considered successful; otherwise, if `max_score` ≤ `threshold`, authentication fails, and a secondary verification process using voiceprint and iris scans must be initiated. The `threshold` can be flexibly set according to actual business security needs. A higher threshold results in higher authentication accuracy but may reduce user pass rate; a lower threshold results in a higher user pass rate but presents certain security risks.

[0050] Figure 2This is an exemplary flowchart of the second authentication process shown in some embodiments of this specification. If the similarity score exceeds a preset threshold, authentication is successful, and the user can access the platform normally; otherwise, authentication fails, triggering the second authentication process. The cloud server obtains the user's authentication similarity score `max_score` and the preset threshold `threshold`. The server determines the authentication result based on the relationship between the two: if `max_score > threshold`, the user is considered to have successfully authenticated. The cloud server signs and encrypts the successful authentication result (including user ID, timestamp, etc.) and returns it to the e-commerce platform. After receiving the successful authentication result, the e-commerce platform allows the user to access platform resources and services normally, including but not limited to browsing products, adding items to the shopping cart, placing orders, and making payments. At the same time, the e-commerce platform records the user's login log for subsequent auditing and security analysis. If `max_score ≤ threshold`, the user is considered to have failed authentication. This means that the facial features submitted by the user have low similarity to all user features in the registration feature library, and the user's identity cannot be confirmed. At this time, the cloud server returns the verification failure result to the e-commerce platform and triggers the second authentication process of voiceprint + iris recognition. After receiving the verification failure result, the e-commerce platform rejects the user's access request and pops up a secondary identity verification prompt interface to guide the user to complete the subsequent voiceprint and iris collection.

[0051] Figure 3This is an exemplary flowchart illustrating the acquisition of voice similarity according to some embodiments of this specification. In the second identity verification process, the system prompts the user to input voice and iris data. The user's voice data and iris image are collected and uploaded to the cloud. When the user's identity verification fails, the e-commerce platform automatically pops up a secondary identity verification page, prompting the user to perform dual verification of voiceprint and iris. For voice data collection, the system generates a short text of 8-10 Chinese characters using a random algorithm and displays the text on the verification page, requiring the user to read the text aloud. Simultaneously, the system calls the device's microphone to record the user's reading at a sampling rate of 44.1kHz and a quantization bit depth of 16 bits, with the recording duration controlled within 5 seconds. After recording, the system encodes and compresses the voice data and uploads it to the cloud via HTTPS protocol, while the local terminal deletes the original recording data. For iris image collection, the system guides the user to align their eyes with the camera through the verification page to capture a high-definition iris image. To ensure the quality of the acquisition, the camera is equipped with an infrared fill light, illuminating the eyes with 850nm near-infrared light to highlight the details of the iris texture. The original iris image captured must have a resolution of at least 1000×1000 pixels. After capture, the iris image is compressed and encoded, then uploaded via HTTPS, and the original image is deleted from the local terminal. The entire capture process should be completed within 5 seconds to avoid discomfort caused by prolonged staring at the infrared light source. After the voice data and iris image are captured, the e-commerce platform sends a secondary verification request to the cloud server, along with metadata such as the user ID and timestamp, requesting the initiation of the voiceprint + iris identity verification process.

[0052] After receiving the voice data uploaded by the user, the cloud server first uses a deep neural network-based speech denoising algorithm to denoise the speech. This algorithm uses a pre-trained convolutional neural network model to identify and suppress noise components by convolving the logarithmic amplitude spectrum of the speech signal in the frequency domain, thereby obtaining a clean speech signal. After denoising, the system performs dual-threshold energy endpoint detection on the speech data, automatically identifying the start and end points of valid speech segments, and removing silent segments before and after the speech to obtain pre-processed speech data.

[0053] Figure 4This is an exemplary flowchart illustrating the acquisition of linear predictive cepstral coefficient features according to some embodiments of this specification. The system uses linear predictive analysis to extract linear predictive cepstral coefficient (LPCC) features from preprocessed speech data. The specific steps are as follows: The preprocessed speech data is segmented into frames with a frame length of 25ms and a frame shift of 10ms, resulting in a speech frame sequence. A Hamming window is applied to each frame of speech data to reduce spectral leakage. The windowed speech frame is sw(n) = s(n) × w(n), where w(n) is the Hamming window function. The autocorrelation function is calculated on the windowed speech frame to obtain P+1 autocorrelation function coefficients R(k), k = 0, 1, ..., P, where P is the linear prediction order, typically 12. Based on the autocorrelation function coefficients, the P-order linear prediction coefficients a_i, i = 1, 2, ..., P, are calculated using the Levinson-Durbin recursive algorithm. The linear prediction coefficients a_i are then subjected to a Fast Fourier Transform (FFT) to obtain the frequency domain prediction coefficients A(ω). The logarithm of the frequency domain prediction coefficients is taken and an inverse Fourier transform is performed to obtain the cepstral coefficients c(n), which are the LPCC features. The 12-dimensional LPCC features of each frame of speech are extracted and concatenated with their first and second order difference parameters to obtain a 36-dimensional LPCC feature vector.

[0054] Figure 5 This is an exemplary flowchart illustrating the acquisition of formant frequency features according to some embodiments of this specification. The system uses a formant estimation algorithm to extract formant frequency features from speech data. The specific steps are as follows: Calculate the linear prediction spectrum H(ω) using linear prediction coefficients a_i. Smooth the linear prediction spectrum using a triangular window function to remove detail fluctuations in the spectral peaks, obtaining a smoothed spectrum H_s(ω). Obtain the local maxima in H_s(ω) as the initial formant frequencies F_i, i = 1, 2, ..., L, where L is the initial number of formants. Perform harmonic detection on F_i to determine if it is an integer multiple of the fundamental frequency, and remove the harmonic peaks. Obtain the optimized formant frequencies F_oi. Select the top 4 formant frequencies with the highest energy in F_oi as the formant frequency features. Perform a 512-point FFT on the formant frequency features to obtain the frequency domain formant features R_FFT. Filter the R_FFT using 24 triangular Mel filters to obtain the 24-dimensional Mel frequency domain formant features R_Mel. Logarithmic transformation and 12-point discrete cosine transform (DCT) are applied to R_Mel to obtain 12-dimensional formant Mel frequency cepstral coefficients (MFCC) features. The 12-dimensional formant MFCC features are concatenated with its first-order and second-order difference parameters to obtain a 36-dimensional formant frequency feature vector.

[0055] The obtained 36-dimensional LPCC features and 36-dimensional formant frequency features are concatenated to form a 72-dimensional voiceprint feature vector. This voiceprint feature vector is then input into a pre-trained deep neural network model, which outputs a similarity score between the current voiceprint and the voiceprints of registered users. This network employs a deep embedding learning architecture, which can map heterogeneous features of different dimensions to a common feature subspace for similarity measurement, thereby improving the accuracy of voiceprint verification. The network is trained using a large-scale multilingual speech dataset and optimized using a contrastive loss function to minimize the distance between the voiceprint feature embeddings of the same speaker in the feature space and maximize the distance between the feature embeddings of different speakers. Finally, the output layer uses the sigmoid function to obtain a voiceprint similarity score between 0 and 1, which serves as the basis for the credibility of voiceprint verification.

[0056] Figure 6 This is an exemplary flowchart illustrating the acquisition of iris recognition similarity according to some embodiments of this specification. After receiving the iris image uploaded by the user, the cloud server first uses a pre-trained convolutional neural network model to perform iris localization on the image. This network model is based on the U-Net architecture, extracting multi-scale features of the image through an encoder and restoring spatial details through a decoder, thus achieving end-to-end iris localization. Figure 7 This is an exemplary flowchart illustrating the acquisition of segmented iris images according to some embodiments of this specification. The localization result is a compact rectangular region that fully covers the iris texture region. Within the iris region, the system uses an image segmentation algorithm to extract the precise iris region. Specifically, the iris region is represented as a weighted undirected graph G = (V, E), where the node set V is the pixel set and the edge set E represents the connection relationship between adjacent pixels. The edge weight w(i, j) represents the similarity between pixels i and j, calculated using a Gaussian function: w(i, j) = exp(-||I(i)-I(j)||^2 / (2σ^2)), where I(i) is the gray value of pixel i and σ is the scale parameter. The energy function E(f) of the graph is constructed, where f is the pixel label (foreground / background) configuration. The energy function consists of a region term R(f) and a boundary term B(f): E(f) = λR(f) + B(f), where λ is a balance factor, R(f) considers the posterior probability of a pixel belonging to the iris / non-iris region, and B(f) considers the consistency of labels between adjacent pixels. The maximum flow minimum cut algorithm is used to solve the energy function globally for optimality, obtaining the segmentation curve C. This results in minE(f) = min{λR(f) + B(f)}. The iris region is then segmented according to the optimal segmentation curve C, with pixels inside C representing iris pixels and pixels outside C representing non-iris pixels. This yields the segmented binary iris image I_seg.

[0057] Figure 8This is an exemplary flowchart illustrating the acquisition of iris texture feature vectors according to some embodiments of this specification. For iris texture feature extraction, the system uses Gabor wavelet transform to extract iris texture features. Gabor wavelets are a class of bandpass filters that simulate the human visual system and can effectively capture multi-directional, multi-scale texture information. The Gabor wavelet basis function ψ(x, y; λ, θ, ψ, σ, γ) is constructed, where λ is the scale parameter, θ is the direction parameter, ψ is the phase offset, σ is the standard deviation of the Gaussian kernel function, and γ is the aspect ratio. The segmented iris image I_seg is radially normalized, and the circular iris is unfolded into a rectangular image I_norm. The Gabor wavelet basis function is used to convolve I_norm: G(x, y; λ, θ) = I_norm(x, y) * ψ(x, y; λ, θ, ψ, σ, γ), to obtain the Gabor wavelet transform coefficients G. The modulus is then taken to obtain the Gabor wavelet energy image E_mag. The E_mag is divided into blocks, and the mean μ and standard deviation σ of each block are calculated to form a block Gabor feature (μ, σ). The block Gabor features at different scales and orientations are concatenated to obtain the iris texture feature vector f_texture. Eight orientations and five scales are selected, resulting in a feature dimension of 40.

[0058] Figure 9 This is an exemplary flowchart illustrating the acquisition of iris shape feature vectors according to some embodiments of this specification. The system extracts iris shape features using a Zernike moment-based method. Zernike moments are shape descriptors based on orthogonal basis functions, invariant to rotation and scale changes. The segmented iris image I_seg is normalized by mapping Cartesian coordinates (x, y) to polar coordinates (ρ, θ), resulting in a normalized iris image I_polar. In polar coordinates, the iris boundary can be represented by the radial distance function r(θ), i.e., the distance from the boundary point to the iris center O varies with angle θ. The Zernike polynomial basis function V_nm(ρ, θ) is used to expand r(θ) to obtain the Zernike moment coefficients Z_nm: Z_nm=(n+1) / π∫∫_D[V_nm(ρ, θ)]*r(θ)ρdρdθ, where n is the order, m is the repetition number, and * denotes complex conjugation. Low-order Zernike moment coefficients (e.g., order n = 1 to 12) are selected to construct the iris shape feature vector f_shape. The first 30 moments are extracted to obtain a 30-dimensional feature vector.

[0059] Figure 10This is an exemplary flowchart illustrating the acquisition of texture statistical features according to some embodiments of this specification. For iris statistical feature extraction, the system employs a Markov Random Field (MRF)-based method to extract the statistical features of iris texture. The MRF model considers the spatial correlation between pixels and can characterize the local statistical properties of the texture. The segmented iris image I_seg is divided into K overlapping local image blocks, each block being S×S in size. An MRF model is established for each image block, with pixel grayscale values ​​representing the observed state, assumed to follow a Gibbs distribution. The MRF energy function E is: E=Σ_iV_1(y_i)+Σ_iΣ_j∈N_iV_2(y_i,y_j), where V_1 is the univariate potential function of pixel yi, related to the pixel grayscale distribution; V_2 is the binary potential function of pixel yi and its neighboring pixels yj, reflecting the spatial correlation between pixels. Given the observed pixel values, the MRF model parameters θ are solved using maximum likelihood estimation to obtain the probability distribution P(y;θ) of the local texture. Statistical features, including mean, variance, and entropy, are extracted from P(y; θ). The statistical features of K image patches are concatenated to obtain the iris statistical feature vector f_stat, which has a dimension of K×3.

[0060] The extracted iris texture features f_texture, iris shape features f_shape, and iris statistical features f_stat are concatenated to obtain the final iris feature vector: f_iris = [f_texture, f_shape, f_stat]. Considering that the scale and distribution of different types of features may vary greatly, the feature vector needs to be normalized to balance the contribution of each feature subspace. Commonly used normalization methods include: Min-Max normalization: linearly scaling the features to the interval [0, 1]. Let the i-th dimension of the feature vector be x_i, then the normalized component x_i' is: x_i' = (x_i - min(x)) / (max(x) - min(x)). Z-score normalization: subtracting the mean from the feature component and dividing by the standard deviation to make it follow a standard normal distribution N(0, 1). The normalized component x_i' is: x_i' = (x_i - μ) / σ, where μ is the mean of the feature component and σ is the standard deviation. L2 normalization: Divide the eigenvector by its L2 norm to make it a unit vector. The normalized eigenvector f' is: f'=f / ||f||_2, where ||*||_2 represents the L2 norm, and ||f||_2=sqrt(Σ_ix_i^2).

[0061] After normalization, the final iris feature vector f_iris is obtained. Hamming distance is used as the similarity metric, comparing the iris feature to be verified with iris features in the registry one by one to obtain a similarity score. Let the iris feature to be verified be f_1, and the i-th iris feature in the registry be f_2^i. The Hamming distance d_i between the two features is defined as: d_i = (f_1XOR f_2^i) / N, where XOR is the exclusive OR operation, and N is the dimension of the feature vector. The value of d_i ranges from [0, 1], representing the dissimilarity between the two binary feature vectors; the smaller the distance, the more similar they are. By setting a similarity threshold τ, it can be determined whether the iris to be verified is a registered iris. If min(d_i) < τ, the iris to be verified is considered to match an iris in the registry, and the verification passes; otherwise, the verification fails. The selection of the threshold τ requires a trade-off between the security and convenience of the iris recognition system, and can be optimized on the development set through methods such as cross-validation.

[0062] To further improve the accuracy and efficiency of iris recognition, several optimization strategies can be employed, such as: Quality assessment: Before feature extraction, the acquired iris images are assessed for quality, and low-quality samples such as blurry, occluded, and reflective images are removed to improve recognition performance. Feature selection: Through methods such as feature ranking and subset search, the most discriminative subset of iris features is selected to reduce feature dimensionality and matching complexity. Index acceleration: Data structures such as hash indexes and tree indexes are used for rapid retrieval of high-dimensional iris features, shortening recognition latency. Multimodal fusion: Combining multiple biometric features such as iris, fingerprint, and face for fusion recognition further enhances system security.

[0063] The cloud server inputs voice similarity and iris similarity into a Bayesian network fusion model, combining the verification results of the two biometric features to obtain the user's identity verification result. Based on the static Bayesian network, a time dimension is introduced to construct a Dynamic Bayesian Network (DBN) model. DBN uses time slices to represent random variables that change over time, and time slices are connected by state transition probabilities to form temporal relationships. Each time slice t contains five nodes: a voice recognition similarity node V^t, an iris recognition similarity node I^t, a voice credibility node T_V^t, an iris credibility node T_I^t, and an identity verification result node R^t. The dependencies between nodes are as follows: V^t and I^t are observation nodes, representing the similarity between voice and iris recognition at time t; T_V^t and T_I^t are latent variable nodes, representing the credibility of voice and iris recognition results at time t, with values ​​ranging from {0, 1}; R^t is a latent variable node, representing the true result of identity verification at time t, with values ​​ranging from {0, 1}, where 0 indicates identity verification failure and 1 indicates success; R^t depends on T_V^t and T_I^t, reflecting the impact of voice and iris credibility on the identity verification result; there is a conditional dependency between T_V^t and T_I^t, reflecting the correlation between the two biometric verification processes; there is a state transition dependency between nodes with the same name in adjacent time slices, such as from R^(t-1) to R^t, reflecting the dynamic changes in identity authenticity.

[0064] Model parameter learning is based on user historical verification data D={(V_n,I_n,R_n)}. The Expectation-Maximization (EM) algorithm is used to train the DBN parameters, including: the conditional probability distribution (CPD) of each node, such as P(R^t|T_V^t,T_I^t), which represents the probability distribution of the authentication result given the voice and iris credibility; and the state transition probability distribution, such as P(R^t|R^(t-1)), which represents the transition probability of the authentication result between two consecutive time slices. The EM algorithm estimates parameters iteratively, with each iteration consisting of two steps: E-step: Estimating the posterior probability distribution of the latent variables based on the current parameters, i.e., P(T_V^t, T_I^t, R^t|V_n, I_n); M-step: Updating the model parameters based on the posterior probabilities of the latent variables, maximizing the likelihood function Q(θ|θ^old)=ΣP(T_V^t, T_I^t, R^t|V_n, I_n;θ^old)logP(V_n, I_n, T_V^t, T_I^t, R^t;θ). In the E-step, the forward-backward algorithm is used to calculate the posterior probabilities of the latent variables. This algorithm recursively calculates the forward probability α and the backward probability β to obtain the joint posterior probability γ. In the M-step, the Junction Tree algorithm is used for precise inference, calculating the marginal probabilities of the latent variables through message passing and updating the CPD and transition probabilities.

[0065] In the authentication reasoning phase, the speech recognition similarity {V_1, ..., V_K} and iris recognition similarity {I_1, ..., I_K} of the user's K consecutive interactions are obtained, constructing an observation sequence O = {(V_1, I_1), ..., (V_K, I_K)}. The Viterbi algorithm is used to reason on the DBN to find the most likely hidden state sequence H* = {(T_V^1*, T_I^1*, R^1*), ..., (T_V^K*, T_I^K*, R^K*)}. The Viterbi algorithm employs dynamic programming to recursively calculate the optimal hidden state for each time slice, ultimately obtaining the globally optimal path. Based on the authentication result R^K* of the last time slice, combined with the global threshold τ and decision rules, it is determined whether the user's identity has been verified: if P(R^K* = 1|O) ≥ τ, the user's authentication is considered successful; otherwise, authentication is considered failed, and the user's access request is rejected.

[0066] Decision optimization and model updates are performed after obtaining the authentication result, with corresponding processing for different situations: Authentication successful (S): The current authentication data (V_K, I_K, R_K) is added to the user's historical dataset D, and the parameters of DBN are updated using an incremental learning algorithm. Incremental learning improves authentication accuracy by locally adjusting CPD and transition probabilities to adapt the model to changes in user behavior patterns. Authentication failed: Suspicious behavior is recorded, including interaction time, device information, IP address, etc., for subsequent security audit analysis. Simultaneously, the global threshold τ and decision rules are dynamically adjusted to improve the ability to identify suspicious users. During decision optimization, factors such as historical authentication results and the degree of user behavior anomalies are comprehensively considered, and a weighted voting mechanism is used to dynamically adjust the authentication pass standard. An adaptive weight allocation strategy is designed to assign different credibility weights to authentication results from different sources, enhancing the robustness of the system.

[0067] Based on the identity verification result, the cloud server returns the identity verification result to the e-commerce platform. If the identity verification is successful, the user is allowed to continue accessing the platform services; otherwise, the user's request is rejected, and the preset exception handling process is initiated.

Claims

1. A cloud-based digital e-commerce platform, including: Capture multiple frames of facial images; Preprocess the acquired multi-frame face images; A pre-trained convolutional neural network is used to extract features from pre-processed multi-frame face images to obtain feature vectors; The obtained feature vector is compared with the feature vector in the pre-stored authorized personnel face feature database to calculate the similarity score. If the calculated similarity score is greater than the threshold, the authentication is successful, and the user is allowed to access the platform. If the calculated similarity score is less than or equal to the threshold, authentication fails. When authentication fails, the system will verify the identity using a pre-configured second authentication method. Secondary authentication methods include: Extract user voice data, use a deep neural network model to perform voiceprint recognition, and obtain voice recognition similarity; The process involves acquiring user iris images, extracting iris features, calculating the similarity with pre-stored iris features, and obtaining iris recognition similarity. A dynamic Bayesian network model is constructed, introducing time-series factors into the static Bayesian network. Using multiple consecutive verifications of speech recognition and iris recognition similarity as input, the dynamic process of identity verification is modeled. Each time slice of the dynamic Bayesian network includes a speech recognition similarity node, an iris recognition similarity node, a speech credibility node, an iris credibility node, and an identity verification result node. Time slices are connected by state transition probabilities. Based on historical user verification data, the expected value algorithm is used to learn the parameters of the dynamic Bayesian network, including the conditional probability distribution and state transition probability distribution of each node. Specifically, the expected step in the expected value algorithm is calculated using a forward-backward algorithm, utilizing Junction. The Tree algorithm is used for inference calculations. During identity verification, the similarity of speech recognition and iris recognition from multiple consecutive user interactions is obtained to construct an observation sequence. Then, the Viterbi algorithm is used to perform inference on a dynamic Bayesian network to find the most likely hidden state sequence, i.e., the dynamic change process of identity authenticity. Based on the identity authenticity probability at the final moment, combined with a global threshold and decision rules, it is determined whether the user's identity has passed verification. The decision rules comprehensively consider the global threshold, historical verification results, and suspicious user behavior factors, and adopt a weighted voting mechanism to adaptively adjust the standard for identity verification. If the identity verification is successful, the current verification data is added to the user's historical dataset, and the parameters of the dynamic Bayesian network are updated using an incremental learning algorithm to achieve online adaptive optimization of the model. Otherwise, the user access is denied, and an early warning mechanism is activated to record suspicious behavior.

2. The cloud-based digital e-commerce platform according to claim 1, characterized in that: The speech recognition similarity is obtained, including: Collect user voice data; A speech denoising algorithm based on deep neural networks is used to denoise the collected user speech data; Endpoint detection and silence removal are performed on the denoised speech data to obtain preprocessed speech data. Linear predictive analysis was used to extract linear predictive cepstral coefficient features from the preprocessed speech data. A formant estimation algorithm is used to extract the formant frequency features of the preprocessed speech data; The extracted linear prediction cepstral system features and formant frequency features are concatenated to obtain the acoustic signature feature vector; The obtained voiceprint feature vector is used as input, and a pre-trained deep neural network model is used to obtain speech similarity.

3. The cloud-based digital e-commerce platform according to claim 2, characterized in that: The linear predictive analysis method is used to extract the linear predictive cepstral coefficient features of the preprocessed speech data, including: The preprocessed speech data is segmented into frames to obtain a speech frame sequence. The Hamming window function is used to window the speech frame sequence to obtain windowed speech frames; The autocorrelation function coefficients are obtained by calculating the autocorrelation function of the windowed speech frames. Based on the autocorrelation function coefficients, the Levinson-Durbin recursive algorithm is used to calculate the linear prediction coefficients; Perform a fast Fourier transform on the linear prediction coefficients to obtain the frequency domain prediction coefficients; Taking the logarithm and inverse Fourier transform of the frequency domain prediction coefficients yields the cepstral coefficients; Multidimensional cepstral coefficients are extracted as features for linear prediction of cepstral coefficients.

4. The cloud-based digital e-commerce platform according to claim 3, characterized in that: The formant frequency features of the preprocessed speech data are extracted using a formant estimation algorithm, including: Calculate the linear prediction spectrum based on the obtained linear prediction coefficients; A triangular window function is used to smooth the linear prediction spectrum, resulting in a smoothed linear prediction spectrum. The local maxima of the smooth linear prediction spectrum are obtained as the initial formant frequencies; Harmonic detection is performed on the initial resonant frequency, and harmonic peaks are removed to obtain the optimized resonant frequency. The top N highest-energy resonant frequencies in the optimized resonant frequencies are used as resonant frequency features. The frequency characteristics of the formant peaks are obtained by performing a discrete Fourier transform on the formant peak frequency characteristics. The frequency domain formant features are filtered by a triangular Mel filter bank to obtain the Mel frequency domain formant features; Logarithmic and discrete cosine transforms were applied to the Mel frequency domain formant characteristics to obtain the MFCC characteristics of the formants. The MFCC characteristics of the resonance peak are concatenated with the corresponding first-order and second-order difference parameters to obtain the final resonance peak frequency characteristics.

5. The cloud-based digital e-commerce platform according to claim 1, characterized in that: Obtain iris recognition similarity, including: An iris detection algorithm based on a convolutional neural network is used to locate the iris in the iris image to obtain the iris region; An image segmentation algorithm is used to segment the iris region to obtain a segmented iris image. The image segmentation algorithm obtains the globally optimal segmentation curve by constructing an energy function of the iris region. The Gabor wavelet transform method was used to extract iris texture features from the iris segmentation image; The iris shape features of the segmented iris image were extracted using a method based on Zernike moments. The statistical features of iris texture in the segmented iris image were extracted using a Markov random field-based method. The extracted iris texture features, iris shape features, and iris texture statistical features are fused to obtain the iris feature vector; Using the iris feature vector as input, the Hamming distance metric is used to calculate the iris recognition similarity.

6. The cloud-based digital e-commerce platform according to claim 5, characterized in that: An image segmentation algorithm is used to segment the iris region, resulting in a segmented iris image. The image segmentation algorithm constructs an energy function for the iris region to obtain the globally optimal segmentation curve, and performs iris segmentation as follows: The iris region is represented as a weighted undirected graph, where the nodes of the graph are the pixels of the iris region, and the weights of the edges are the similarity between adjacent pixels. Construct an energy function for a weighted undirected graph. The energy function includes a region term and a boundary term. The region term represents the probability that a pixel belongs to the iris region or a non-iris region, and the boundary term represents the probability that adjacent pixels have consistent labels. The energy function is minimized using the maximum flow minimum cut algorithm to obtain the globally optimal segmentation curve; The iris region is segmented based on the globally optimal segmentation curve to obtain the segmented iris image; where the area inside the segmentation curve is the iris region and the area outside is the non-iris region.

7. The cloud-based digital e-commerce platform according to claim 6, characterized in that: The Gabor wavelet transform method was used to extract iris texture features from the segmented iris image, including: Construct Gabor wavelet basis functions, where the parameters of the Gabor wavelet include scale, direction, center frequency, and bandwidth; Radial normalization is performed on the segmented iris image to convert the circular iris image into a rectangular image; The Gabor wavelet transform coefficients are obtained by convolving the constructed Gabor wavelet basis functions onto the rectangular image. The Gabor wavelet transform coefficients are moduloed to obtain the Gabor wavelet energy image; The Gabor wavelet energy image is divided into blocks, and the mean and variance of each block are calculated to obtain the block Gabor features; The Gabor features in blocks are combined according to scale and orientation to obtain the iris texture feature vector.

8. The cloud-based digital e-commerce platform according to claim 7, characterized in that: The iris shape features of the segmented iris image are extracted using a Zernike moment-based method, including: The segmented iris image is normalized by converting the Cartesian coordinate system to the polar coordinate system to obtain the normalized iris image. In polar coordinates, the normalized iris image is represented as a radial distance function; where the radial distance function represents the distance from the iris boundary to the iris center as a function of the angle. The radial distance function is expanded using Zernike polynomial basis functions to obtain the Zernike moment coefficients; where the Zernike polynomial basis functions are a set of orthogonal basis functions used to describe the radial distance function; Select preset low-order Zernike moment coefficients as the iris shape feature vector; where the preset low-order coefficients correspond to the global features of the iris shape.

9. The cloud-based digital e-commerce platform according to claim 8, characterized in that: The statistical features of iris texture in the segmented iris image were extracted using a Markov random field-based method, including: The segmented iris image is divided into multiple overlapping local image blocks; A Markov random field model is established for each local image patch; where the pixels in the image patch are field points, and the correlation between pixels is the interaction between field points. The energy function of the Markov random field model is defined using the Gibbs distribution; the energy function consists of the pixel gray-level distribution and the correlation between pixels. The parameters of the energy function are estimated using the maximum likelihood estimation method, thus obtaining the parameters of the Markov random field model. By statistically analyzing the parameters of the Markov random field model, the texture statistical features of local image patches are obtained; among them, the texture statistical features include mean, variance and entropy. The texture statistical features of all local image patches are concatenated to obtain the iris texture statistical feature vector.

Images