An identity authentication method and system for enhancing behavioral fingerprint data

Abstract

The present invention discloses a method and system for identity authentication with enhanced behavioral fingerprint data. The method includes the following steps: performing data preprocessing on historical and anonymous behavioral log records to obtain a non-anonymous data set and an anonymous data set; extracting joint extended behavioral features according to the non-anonymous data set; extracting the feature vectors of users in the non-anonymous data set and performing data enhancement; constructing and training a user recognition model; calculating the recognition accuracy of different metric formulas as the weights of the corresponding methods; merging the training set and the validation set to retrain the network, and inputting the test set into the trained model to obtain the similarity scores between each anonymous user and each known user; merging multiple validation sets to obtain a template library, and calculating the similarity scores according to the metric formula; and fusing multiple classifiers to identify the identity of anonymous users according to the similarity score matrix and the weights of each classifier. The present invention achieves better identity recognition effects and provides a new perspective for identifying user identities using frequency features.

Description

Technical Field

[0001] The present invention relates to the technical field of user identity recognition, and in particular to an identity authentication method and system enhanced with behavioral fingerprint data. Background Art

[0002] Identity recognition has important applications in a variety of fields, including security systems, surveillance, and fraud prevention. Currently, the most commonly used biometric identification methods, such as face and fingerprint, are typically one-time, require active user participation, and face challenges introduced by new technologies such as generative adversarial networks. However, behavioral identification methods, such as keystrokes, gait, and logging, offer implicit and continuous authentication and can be used as components of multimodal biometric identification systems to enhance system security, attracting widespread attention from researchers.

[0003] Counting the frequency of user behaviors to form a user interest histogram as a behavioral feature and using this to identify users is a simple and easy method with high recognition accuracy. Directly using histograms as behavioral features ignores the sequential information of user behaviors. Some researchers have addressed this issue by treating frequent, continuous user behaviors as a single new behavior. However, this approach only considers common features across users. While saving computing resources, it loses more personalized user expressions. Furthermore, the use of string sets to represent sequential features significantly reduces recognition efficiency. Another issue that needs to be addressed is that extracting frequency features from raw data requires a certain amount of behavioral accumulation, resulting in a small number of data samples, making it difficult to achieve effective machine learning results. Summary of the Invention

[0004] In order to overcome the defects and shortcomings of the existing technology, the present invention provides an identity authentication method enhanced by behavioral fingerprint data. The present invention achieves better identity recognition effect and provides a new perspective for identifying user identities using frequency characteristics.

[0005] A second object of the present invention is to provide an identity authentication system enhanced by behavioral fingerprint data.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions:

[0007] The present invention provides an identity authentication method enhanced with behavioral fingerprint data, comprising the following steps:

[0008] Obtain historical behavior log records and anonymous behavior log records, perform data preprocessing on the historical behavior log records and the anonymous behavior log records to obtain non-anonymous data sets and anonymous data sets, and divide the non-anonymous data sets by time to obtain multiple non-anonymous data subsets;

[0009] Based on a non-anonymous dataset, according to the feature extraction algorithm SURE + Obtain the joint extended behavior features;

[0010] Extract the feature vectors of users in multiple non-anonymous data subsets according to the joint extended behavior features, select one as the validation set, and the rest as the training set, and perform preliminary data augmentation on the training set. Extract the feature vectors of users in the anonymous dataset according to the joint extended behavior features to obtain the test set;

[0011] Construct a user recognition model based on a convolutional neural network, including a normalization layer, a convolutional layer, residual connections, and the splicing of multiple channel results. Input the training set after preliminary data augmentation into the user recognition model for data augmentation training to obtain the trained user recognition model;

[0012] Randomly erase the records of a certain time interval in the validation set multiple times to obtain the augmented validation set. Input the augmented validation set into the trained user recognition model to obtain the recognition accuracy as the weight of this classifier of the user recognition model based on the convolutional neural network. At the same time, each measurement method also corresponds to a classifier, and calculate the recognition accuracy of the augmented validation set based on different measurement formulas as the weight value of the corresponding classifier

[0013] Merge the training set and the validation set and input them into the user recognition model to retrain the user recognition model. Input the test set into the retrained user recognition model to obtain the similarity scores between each anonymous user and each known user. Merge multiple validation sets to obtain a template library, and calculate the similarity scores between the samples in the test set and the template library according to the measurement formula;

[0014] Construct a similarity score matrix based on multiple similarity scores, and based on the similarity matrix and the weights of each classifier, perform weighted summation to obtain the final similarity matrix, and fuse multiple classifiers to identify the identities of anonymous users.

[0015] As a preferred technical solution, perform data preprocessing on historical behavior log records and anonymous behavior log records, specifically including:

[0016] The data preprocessing includes noise data processing and unified data format. Use continuous integer encoding for behaviors, and divide the historical behavior records according to preset segmentation time points to obtain multiple non-anonymous datasets.

[0017] As a preferred technical solution, based on a non-anonymous dataset, according to the feature extraction algorithm SURE + Obtain the joint extended behavior features, specifically including:

[0018] Set the parameters of the extended behavior features and calculate the placeholders of the behavior encoding;

[0019] For users in the non-anonymous dataset, the behavior sequence is extracted and the behavior feature hierarchy extracted from the user is expressed as:

[0020] n u =min(n,L u )

[0021] Among them, u represents the user, L u Indicates the length of the behavior sequence;

[0022] The behavior sequence is expressed as:

[0023] Among them, S u Indicates a behavior sequence, ph indicates a placeholder for behavior coding

[0024] The number of times each element in the behavior sequence appears and the behavior number are recorded as The behavior number and frequency of all layers are recorded as That is, the extended behavioral characteristics of user u;

[0025] The extended behavior features obtained for each user are sorted from high to low according to their frequency, and the feature number is retained and recorded as Set the selection ratio, select the features of each user within the selection ratio range, and obtain the final joint extended behavior features, which can be expressed as:

[0026]

[0027] Among them, F n represents the final joint extended behavioral features, J represents the number of non-anonymous data sets, r top Indicates the selection ratio.

[0028] As a preferred technical solution, feature vectors of users from multiple non-anonymous data subsets are extracted based on the joint extended behavioral features. One of them is selected as a validation set, and the rest are used as training sets. Preliminary data augmentation is performed on the training set. Feature vectors of users from the anonymous dataset are extracted based on the joint extended behavioral features to obtain a test set. Specifically, the following steps are performed:

[0029] For the non-anonymous data subset, it is evenly divided into multiple intervals based on time, and the frequency of occurrence of the joint extended behavior feature elements of the user in each interval is counted to obtain the feature matrix Where p = len(F n ), sum each column of the feature matrix as the validation set corresponding to other training sets;

[0030] Randomly erased feature matrix For the records in it, sum each column to obtain training samples, perform data augmentation based on multiple random erasures, and stop the augmentation based on a set threshold of the number of samples in the augmented training set to obtain the augmented training set;

[0031] Count the occurrence frequency of each user's combined expansion behavior features in the anonymous dataset, that is, the feature vector of each user, to obtain the test set.

[0032] As a preferred technical solution, construct a user recognition model based on a convolutional neural network, including a normalization layer, a convolutional layer, a residual connection, and the splicing of multiple channel results. Input the training set after preliminary data augmentation into the user recognition model for data augmentation training to obtain the trained user recognition model, specifically including:

[0033] Normalize the input training set and calculate the probability that the input training set is each known user, specifically expressed as:

[0034]

[0035]

[0036] w i =σ(W*FC input +b);

[0037] Among them, Conv represents a one-dimensional convolutional layer, is the activation function, Flatten represents expanding the results of multiple channels of the convolution into a one-dimensional vector, Dropout represents that neurons are randomly deactivated with a set probability, W is the weight matrix of the fully connected layer, b is the bias of the fully connected layer, and σ(·) is the Softmax activation function;

[0038] Input the training set after preliminary data augmentation into the user recognition model, and train and update the user recognition model with the cross-entropy with L2 regularization as the loss function. The loss function is expressed as:

[0039]

[0040] Among them, y i represents the one-hot encoding of the true label of the i-th sample, θ is all the parameters of the model, λ is the regularization coefficient, and w ij represents the probability that the i-th sample is recognized as the known user u j That is, the similarity between the two;

[0041] Iterate and loop multiple times until the loss function converges or reaches the number of iterations to obtain the trained user recognition model.

[0042] As a preferred technical solution, the training set after preliminary data augmentation is augmented using the Mixup algorithm, which is specifically expressed as follows:

[0043] Randomly linearly combine two sample pairs (X k , y k ), (X l , y l ) in a batch multiple times to form new training sample pairs, which are specifically expressed as follows:

[0044]

[0045] Among them, γ ∈ [0, 1] represents a parameter randomly drawn from Beta(α, α), and α represents a preset parameter.

[0046] As a preferred technical solution, the metric formula includes: a similarity or distance metric formula considering global features, a similarity or distance metric formula considering specific values of behavioral features, and a similarity or distance formula considering the preference ranking of behaviors.

[0047] As a preferred technical solution, it further includes a step of normalizing the similarity score matrix, and normalizing the similarity score matrix constructed based on multiple similarity scores.

[0048] As a preferred technical solution, multiple classifiers are fused to identify the identity of anonymous users. For the identification of a single anonymous user, the one with the maximum similarity to the known user is the identification result. For the identification tasks of multiple anonymous users known to be different users, the maximum sum of the similarities of all identification results is the final identification result.

[0049] To achieve the above second object, the present invention adopts the following technical solutions:

[0050] The present invention provides an identity authentication system for behavior fingerprint data augmentation, including: a data preprocessing module, a feature extraction module, a data augmentation module, a user identification model construction module, a user identification model training module, a classifier weight value calculation module, a similarity score calculation module, and an anonymous user identity identification module;

[0051] The data preprocessing module is used to obtain historical behavior log records and anonymous behavior log records, perform data preprocessing on the historical behavior log records and anonymous behavior log records to obtain a non-anonymous data set and an anonymous data set, and divide the non-anonymous data set by time to obtain multiple non-anonymous data subsets;

[0052] The feature extraction module is used to obtain joint extended behavior features based on the non-anonymous data set according to the feature extraction algorithm SURE + ;

[0053] The data augmentation module is used to extract the feature vectors of users in multiple non-anonymous data subsets according to the combined expansion behavior characteristics, select one as the validation set, and the rest as the training set, and perform preliminary data augmentation on the training set. The feature vectors of users in the anonymous data subset are extracted according to the combined expansion behavior characteristics to obtain the test set;

[0054] The user identification model construction module is used to construct a user identification model based on a convolutional neural network, including a normalization layer, a convolutional layer, a residual connection, and the splicing of multiple channel results;

[0055] The user identification model training module is used to input the training set after preliminary data augmentation into the user identification model for data augmentation training to obtain the trained user identification model, input the validation set into the trained user identification model, adjust the hyperparameters and update the weight values of the user identification model;

[0056] The classifier weight value calculation module is used to calculate the weight value of the classifier. The records in a certain time interval of the validation set are randomly erased multiple times to obtain the augmented validation set. The augmented validation set is input into the trained user identification model to obtain the recognition accuracy as the weight of this classifier of the user identification model based on the convolutional neural network. At the same time, each measurement method also corresponds to a classifier, and the recognition accuracy of the augmented validation set is calculated based on different measurement formulas as the weight value of the corresponding classifier;

[0057] The similarity score calculation module is used to calculate the similarity scores of the test set samples. The training set and the validation set are combined and input into the user identification model to retrain the user identification model. The test set is input into the retrained user identification model to obtain the similarity scores between each anonymous user and each known user. Multiple validation sets are combined to obtain a template library, and the similarity scores between the test set samples and the template library are calculated according to the measurement formula;

[0058] The anonymous user identity identification module is used to output the anonymous user identity identification result. A similarity score matrix is constructed based on multiple similarity scores. According to the similarity matrix and the weights of each classifier, the weighted sum is obtained to get the final similarity matrix, and multiple classifiers are fused to identify the anonymous user identity.

[0059] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0060] The present invention improves the efficiency of extracting extended features by performing integer encoding on continuous behaviors. The combined extended features enable certain similarity calculations between users to be accelerated by means of matrix operations. At the same time, more individual expressions of users are retained. By erasing and linear combination, the training data is enhanced without increasing the cost of data collection and storage, effectively improving the recognition accuracy, supporting the integration of traditional metric-based recognition methods, and obtaining more reliable and interpretable recognition results. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 It is a schematic flowchart of the identity authentication method for behavior fingerprint data enhancement according to the present invention;

[0062] Figure 2 It is a flowchart of constructing combined extended behavior features for anonymous user identity authentication according to the present invention;

[0063] Figure 3 It is a one-dimensional convolutional neural network structure diagram for authenticating anonymous user identities based on combined extended behavior features according to the present invention;

[0064] Figure 4 It is a framework diagram for integrating metric-based and neural network-based anonymous identity authentication according to the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0065] In order to make the objectives, technical solutions and advantages of the present invention clearer and more understandable, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0066] Embodiment 1

[0067] As Figure 1 shown, this embodiment provides an identity authentication method for behavior fingerprint data enhancement, including the following steps:

[0068] S1: Perform data preprocessing on historical behavior logs and anonymous behavior logs and divide them to obtain a non-anonymous data set and an anonymous data set;

[0069] In this embodiment, performing data preprocessing on historical behavior records and anonymous behavior records includes noise data processing and unifying data formats, encoding behaviors with consecutive integers starting from 1, and dividing historical behavior records by time to obtain multiple non-anonymous data sets.

[0070] In this embodiment, the dataset used is the network TV playback behavior records provided by a certain company. The original data format is <userID, CurChannel, NextChannel, time, duration>. The noise records with a viewing duration (duration) less than 5s are removed, sorted in ascending order by [userID, time], and time is converted into date. The final data format retained is <userID, CurChannel, date>. Among them, userID is the user number; CurChannel is the current viewing channel, with a range of 1 - 157; the range of date is 0 - 31; the segmentation time points are set to 0, 10, 11, 20, that is, the records with date in 0 - 10 constitute the non-anonymous dataset 1, and the user set is denoted as The records with date in 11 - 20 constitute the non-anonymous dataset 2, and the user set is denoted as The remaining records during 21 - 31 constitute the anonymous dataset;

[0071] S2: According to the feature extraction algorithm SURE + , determine the joint extended behavior features with the help of the non-anonymous dataset;

[0072] As Figure 2 shown, determine the feature extraction algorithm SURE for the joint extended behavior features used for identity recognition + Specifically, it includes the following steps:

[0073] S2.1. Weigh the recognition accuracy and the consumption of computing resources, determine the parameter n of the extended behavior features according to actual needs, and calculate the placeholder of the behavior encoding In this embodiment, preferably n = 2, B = 157,

[0074] S2.2. For user u in the non-anonymous dataset, extract the behavior sequence S u , denote the length of S u as L u , then the level of behavior features that can be extracted for user u is n u = min(n, L u );

[0075] The behavior sequence of the Kth layer of user u is: Count the number of occurrences of each element and the behavior number and record it as Among them, the behavior sequence of the first layer of user u is S u [1:L u , count the number of occurrences of each element and the behavior number and record it as The behavior sequence of the second layer is S u[1:(L u -1)]+S u [2:L u ×10 ph , count the occurrences of each element and record the behavior number as Continue until the n u th layer, and the behavior sequence is Count the occurrences of each element and record the behavior number as Combine the behavior numbers and counts of all layers and record as That is, the extended behavior feature of user u;

[0076] S2.4. Sort the extended behavior features obtained for each user in descending order of their frequencies, and only retain the feature numbers, denoted as Take the first r top proportion of features for each user, and use them as part of the final combined extended behavior feature F n , that is J represents that a total of J non-anonymous data sets are divided. In this embodiment, the r top preferred value is 0.2, that is, take the first 20% proportion of features for each user. In this embodiment, the non-anonymous data set Combined extended behavior feature F n is expressed as

[0077] S3: Extract the feature vectors of users in multiple non-anonymous data subsets according to the combined extended behavior feature, select one as the validation set, and the rest as the training set, and perform preliminary data augmentation on the training set. Extract the feature vectors of users in the anonymous data set according to the combined extended behavior feature to obtain the test set;

[0078] In this embodiment, according to the combined extended behavior feature, extract the feature vectors of users in multiple non-anonymous data subsets, select one as the validation set, and the rest as the training set, and perform preliminary data augmentation on the training set. Extract the feature vectors of users in the anonymous data set according to the combined extended behavior feature to obtain the test set, which specifically includes:

[0079] S3.1. For non-anonymous data set j, evenly divide the time into m j intervals, and respectively count the frequencies of the elements in F n for each interval of user u to obtain the feature matrix where p = len(F n 0. Sum the columns of the feature matrix as a sample of the corresponding validation set of other training sets. For the non-anonymous data set of this embodiment Evenly divide the time into m1 = 11, m2 = 10 intervals to obtain the feature matrix

[0080] S3.2, Randomly erase the records of a certain row in the feature matrix, sum each column to obtain a training sample, with the label being this user. At most, m samples can be obtained. For 0 ≤ m ≤ m - 1, randomly erase the records of two rows in the feature matrix, sum each column to obtain a training sample, with the label being this user. At most, j n samples can be obtained. Then, through this method, the training set samples can be expanded from 1 to at most p samples. In this embodiment, the number of training set samples amplified for each user is controlled by the parameter ω, that is, when the number of training set samples amplified for a user is greater than ω, the amplification stops to avoid noise caused by too short a time interval for statistical frequency, thereby obtaining a training set. In this embodiment, ω is preferably ω = 10; S3.3, Count the occurrence frequency of each user in the anonymous dataset with respect to F

[0081] to obtain test set samples; n In this embodiment, the feature vectors of users in the three datasets are extracted according to the combined extended behavior features, and data augmentation is performed to obtain the first training set, the first validation set, the second training set, the second validation set, and the test set;

[0082] S4: As

[0083] shown, construct a user recognition model based on a convolutional neural network, including a normalization layer, a convolutional layer, a residual connection, and the splicing of multiple channel results. Input the preliminarily amplified training set to train the model, and perform further data augmentation training based on the preliminarily amplified training set; Figure 3 In this embodiment, the specific process of constructing a user recognition model based on a convolutional neural network and data augmentation training is as follows:

[0084] S4.1, Normalize the input sample

[0085] to accelerate convergence, which can be specifically expressed as: where E(X ) represents the mean, D(X i ) represents the variance, and ε is to prevent the denominator from being 0; i S4.2, Calculate the probability of the input sample for each known user through a convolutional layer, a residual connection, a fully connected layer, etc. Its mathematical formula is described as follows:

[0086]

[0087]

[0088] ​

[0089] w i = σ(W * FC input + b);

[0090] Among them, Conv represents a one-dimensional convolutional layer, is the activation function, Flatten means expanding the results of multiple channels of convolution into a one-dimensional vector, Dropout means that neurons are randomly deactivated with a certain probability, W is the weight matrix of the fully connected layer, b is the bias of the fully connected layer, and σ(·) is the Softmax activation function;

[0091] In this embodiment, the convolutional kernel size of the one-dimensional convolutional layer Conv is 3, the stride is 1, and zero-padding operation is adopted to keep the dimensions unchanged before and after the convolutional operation. The activation function can be selected from GELU or RELU activation functions, the number of channels is 4, and neurons are randomly deactivated with a probability of 0.9;

[0092] S4.3. Input the preliminarily amplified training set into the neural network model, and train and update the classification model with the cross-entropy added with L2 regularization as the loss function. The formula description of the loss function is:

[0093]

[0094] where y i represents the one-hot encoding of the true label of the i-th sample, θ is all the parameters of the model, λ is the regularization coefficient, and w ij represents the probability that the i-th sample is recognized as the known user u j , that is, the similarity between the two;

[0095] S4.4. Perform further data augmentation on the data, and then input it into the neural network model trained in step S4.3 to continue training and updating the model. The further data augmentation specifically refers to Mixup: randomly linearly combining two sample pairs (X k , y k ), (X l , y l ) in a batch to form new training sample pairs. The formula description is as follows:

[0096]

[0097] where γ ∈ [0, 1] is randomly drawn from Beta(α, α), and α is a manually set parameter, which is set to 1 in this embodiment;

[0098] S4.5. Continuously repeat the processes of steps S4.3 and S4.4 until the loss of the true samples converges or reaches the specified number of loops;

[0099] S4.6. Input the validation set into the trained user recognition model, adjust the hyperparameters, and update the weight values of the user recognition model;

[0100] In this embodiment, the validation set is input into the trained model, the loss function image and classification accuracy are observed, and the hyperparameters such as the learning rate and regularization coefficient are adjusted; the finally determined learning rate in this embodiment is 0.1, the momentum is 0.9, the regularization coefficient is 0.001, the batchsize is 1000, and the epoch is 100;

[0101] S5: Randomly erase the records of a certain time interval in the validation set multiple times to obtain an enhanced validation set. Input the enhanced validation set into the trained user recognition model, and obtain the recognition accuracy as the weight of this classifier of the user recognition model based on the convolutional neural network. At the same time, each measurement method also corresponds to a classifier, and the recognition accuracy of the enhanced validation set is calculated based on different measurement formulas as the weight value of the corresponding classifier;

[0102] In this embodiment, for the recognition of a single anonymous user, the one with the maximum similarity to the known user is the recognition result, called NM. For the recognition task of multiple anonymous users known to be different from each other, the maximum sum of the similarities of all recognition results is the final recognition result, called GM. The formulas are described as follows:

[0103] where U x represents the set of anonymous users, that is, the users in the validation set or test set, and U y represents the set of known users, that is, the users in the training set;

[0104] That is, if one of the test samples is recognized as user No. 1, no other samples can be user No. 1;

[0105] In this embodiment, randomly erasing the records of a certain time interval in the validation set multiple times to obtain the enhanced validation set specifically is:

[0106] Denote the matrix composed of the feature vectors of all users in the first validation set as Denote the matrix composed of the feature vectors of all users in the second validation set as The first validation set erases the records of the first day, the second day,... the m2th day respectively, and counts the occurrence frequency of each user with respect to F n to obtain the enhanced first validation set denoted as Similarly, obtain the enhanced second validation set denoted as

[0107] In this embodiment, calculating the weight of the classifier of the user recognition model based on the convolutional neural network according to the enhanced validation set specifically is:

[0108] Taking the first validation set as an example, the enhanced validation set in which u i The feature vector of is input into the trained neural network classifier. The output value of each category is the similarity between the 'anonymous' user and each category. According to NM or GM, the recognition result is By calculating the recognition accuracy of simple matching in the first validation set where is an indicator function. When the value is 1, otherwise it is 0. Similarly, calculate the recognition accuracy of simple matching in the second validation set under the network model trained on the training set Then for the recognition task of a single user, the weight of this classifier is Similarly, the weights of this classifier for the recognition tasks of multiple different users can be obtained as

[0109] In this embodiment, each metric method also corresponds to a classifier respectively. The specific steps for calculating the recognition accuracy of the enhanced validation set based on different metric formulas as the weight values of the corresponding classifier are as follows:

[0110] S5.1. Calculate the similarity between the users in the validation set and the known users based on different metric formulas, specifically including: similarity or distance metric formulas considering global features, such as Jaccard; similarity or distance metric formulas considering the specific values of behavioral features, such as Cosine, KL; similarity or distance formulas considering the preference ranking of behaviors, such as Spearman rank correlation coefficient, Kendall-τ rank correlation coefficient;

[0111] Taking the anonymous user in the first validation set and the known user in the second validation set as an example; the similarity calculation formulas between an anonymous user and all known users are respectively:

[0112] where is the matrix obtained by replicating and expanding the feature vector of user u in the first validation set to When both a and b are not 0, a∧b is 1, otherwise it is 0. When a and b are not both 0, a∨b is 1, otherwise it is 0;

[0113] where represents the matrix obtained by expanding the behavioral probability of user u in the first validation set,

[0114] PX 2 represents the matrix composed of the behavioral probabilities of all users in the second validation set;;

[0115] where is the rank of, then is its average. If then R 2 is the matrix formed by the ranks of all users obtained according to X 2 , is the vector formed by the average rank of each user;

[0116] ε is to prevent the denominator from being zero;

[0117] S5.2. The enhanced validation set obtains a similarity score matrix according to the above similarity calculation method, and then performs NM or GM to obtain the recognition result. The recognition accuracy of all users in the first validation set and the recognition accuracy of all users in the second validation set can be obtained, and the average value of the accuracy is used as the weight of each classifier:

[0118] S6: Combine the training set and the validation set, retrain the network according to the determined parameters, input the samples of the test set into the trained neural network to obtain the similarity score between each anonymous user and each known user; combine multiple validation sets to obtain a template library, and calculate the similarity score between the samples of the test set and the template library according to the similarity calculation formula;

[0119] In this embodiment, based on the feature matrix After amplification, the first training set and the second training set are obtained. The first training set has N1 samples, and the second training set has N2. Combine the first training set and the second training set to obtain N1 + N2 samples. Retrain the network according to the determined parameters and structure, input the samples of the test set into the trained neural network to obtain the similarity score between each anonymous user and each known user. The similarity score matrix is denoted as W NN ;

[0120] Combine the first validation set and the second validation set, that is, add the feature vectors of the corresponding users. According to the description of the foregoing similarity calculation formula, calculate the similarity score between each anonymous user in the anonymous dataset and all known users , and are respectively denoted as W Jaccard , W KL , W Spear , W JKL ;

[0121] S7: According to the obtained multiple similarity score matrices and the weights of each classifier, fuse multiple classifiers to further identify the identity of anonymous users.

[0122] Such as Figure 4As shown, based on the obtained multiple similarity score matrices and the weights of each classifier, multiple classifiers are fused to further identify the identity of anonymous users, specifically including:

[0123] S7.1 Normalize the obtained multiple similarity score matrices, while ensuring that the larger the value, the more similar, or the smaller the value, the more similar. Specifically, it is expressed as:

[0124]

[0125] S7.2 According to the normalized similarity matrix and the weights of each classifier, perform weighted summation to obtain the final similarity matrix, and for different identification tasks, use NM or GM to obtain the identification result;

[0126] According to the normalized similarity matrix and the weights of each classifier, perform weighted summation to obtain the final similarity matrix under different identification tasks. Specifically, it is expressed as:

[0127]

[0128]

[0129] For different identification tasks, for the similarity matrix W NM and W GM respectively use NM and GM to obtain the final identification result.

[0130] Embodiment 2

[0131] This embodiment provides an identity authentication system for behavioral fingerprint data enhancement, including: a data preprocessing module, a feature extraction module, a data enhancement module, a user identification model construction module, a user identification model training module, a classifier weight value calculation module, a similarity score calculation module, and an anonymous user identity identification module;

[0132] In this embodiment, the data preprocessing module is used to obtain historical behavior log records and anonymous behavior log records, perform data preprocessing on the historical behavior log records and anonymous behavior log records to obtain a non-anonymous data set and an anonymous data set, and divide the non-anonymous data set by time to obtain multiple non-anonymous data subsets;

[0133] In this embodiment, the feature extraction module is used to obtain joint extended behavior features based on the non-anonymous data set according to the feature extraction algorithm SURE + ;

[0134] In this embodiment, the data augmentation module is used to extract the feature vectors of users in multiple non-anonymous data subsets according to the joint expansion behavior characteristics, select one as the validation set, and use the rest as the training set. Then, it performs preliminary data augmentation on the training set, extracts the feature vectors of users in the anonymous data subset according to the joint expansion behavior characteristics to obtain the test set;

[0135] In this embodiment, the user identification model construction module is used to construct a user identification model based on a convolutional neural network, including a normalization layer, a convolutional layer, a residual connection, and the splicing of multiple channel results;

[0136] In this embodiment, the user identification model training module is used to input the training set after preliminary data augmentation into the user identification model for data augmentation training to obtain the trained user identification model. Then, it inputs the validation set into the trained user identification model to adjust the hyperparameters and update the weight values of the user identification model;

[0137] In this embodiment, the classifier weight value calculation module is used to calculate the weight value of the classifier. It randomly erases the records in a certain time interval of the validation set multiple times to obtain the augmented validation set, inputs the augmented validation set into the trained user identification model, and takes the recognition accuracy as the weight of this classifier of the user identification model based on the convolutional neural network. At the same time, each measurement method also corresponds to a classifier, and the recognition accuracy of the augmented validation set is calculated as the weight value of the corresponding classifier based on different measurement formulas;

[0138] In this embodiment, the similarity score calculation module is used to calculate the similarity scores of the test set samples. It combines the training set and the validation set and inputs them into the user identification model to retrain the user identification model. Then, it inputs the test set into the retrained user identification model to obtain the similarity scores between each anonymous user and each known user. It combines multiple validation sets to obtain a template library and calculates the similarity scores between the test set samples and the template library according to the measurement formula;

[0139] In this embodiment, the anonymous user identity identification module is used to output the anonymous user identity identification result. It constructs a similarity score matrix based on multiple similarity scores, and according to the similarity matrix and the weights of each classifier, performs weighted summation to obtain the final similarity matrix, and fuses multiple classifiers to identify the anonymous user identity.

[0140] The above embodiments are the preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications made without departing from the spirit and principle of the present invention shall be equivalent replacement methods and shall be included in the protection scope of the present invention.

Claims

1. An identity authentication method for enhancing behavioral fingerprint data, characterized in that, It includes the following steps: Obtain historical behavior log records and anonymous behavior log records, perform data preprocessing on the historical behavior log records and anonymous behavior log records to obtain a non-anonymous data set and an anonymous data set, and divide the non-anonymous data set by time to obtain multiple non-anonymous data subsets; Based on a non-anonymous dataset, according to the feature extraction algorithm SURE + Obtain the joint extended behavior features; Extract the feature vectors of users in multiple non-anonymous data subsets according to the joint extended behavior features, select one as the validation set, and the rest as the training set, and perform preliminary data augmentation on the training set. Extract the feature vectors of users in the anonymous data set according to the joint extended behavior features to obtain the test set; Construct a user recognition model based on a convolutional neural network, including a normalization layer, a convolutional layer, a residual connection, and the splicing of multiple channel results. Input the training set after preliminary data augmentation into the user recognition model for data augmentation training to obtain the trained user recognition model. Input the validation set into the trained user recognition model, adjust the hyperparameters and update the weight values of the user recognition model; Randomly erase the records of a certain time interval in the validation set multiple times to obtain the enhanced validation set. Input the enhanced validation set into the trained user recognition model to obtain the recognition accuracy as the weight of this classifier of the user recognition model based on the convolutional neural network. At the same time, each measurement method also corresponds to a classifier, and calculate the recognition accuracy of the enhanced validation set based on different measurement formulas as the weight value of the corresponding classifier; Merge the training set and the validation set and input them into the user recognition model to retrain the user recognition model. Input the test set into the retrained user recognition model to obtain the similarity scores between each anonymous user and each known user. Merge multiple validation sets to obtain a template library, and calculate the similarity scores between the samples in the test set and the template library according to the measurement formula; Construct a similarity score matrix based on multiple similarity scores, and perform weighted summation according to the similarity matrix and the weights of each classifier to obtain the final similarity matrix, and fuse multiple classifiers to identify the identity of anonymous users.

2. The identity authentication method for enhancing behavioral fingerprint data according to claim 1, characterized in that, Perform data preprocessing on the historical behavior log records and anonymous behavior log records, specifically including: The data preprocessing includes noise data processing and unified data format. The behavior is encoded with continuous integers, and the historical behavior records are divided into multiple non-anonymous data sets according to the time length of the anonymous data set.

3. The identity authentication method for enhancing behavioral fingerprint data according to claim 1, wherein Based on a non-anonymous dataset, according to the feature extraction algorithm SURE + Obtain the joint extended behavior features, specifically including: Set the parameters of the extended behavior features and calculate the placeholders of the behavior encoding; For the users in the non-anonymous data set, extract the behavior sequence, and hierarchically represent the behavior features extracted by the users as: n u = min(n, L u ) Among them, u represents the user, and L u represents the length of the behavior sequence; The behavior sequence is represented as: Among them, S u represents the behavior sequence, and ph represents the placeholder for behavior encoding; Record the number of occurrences of each element in the statistical behavior sequence and the behavior number as Record the combined behavior numbers and occurrences of all layers as That is, the extended behavior feature of user u; Sort the extended behavior features obtained for each user in descending order of their frequencies, retain the feature numbers, denoted as Set the selection ratio, and select the features of each user within the selection ratio range to obtain the final combined extended behavior features.

4. The identity authentication method for enhancing behavioral fingerprint data according to claim 3, wherein Extract the feature vectors of users in multiple non-anonymous data subsets according to the joint extended behavior features, select one as the validation set, and the rest as the training set, and perform preliminary data augmentation on the training set. Extract the feature vectors of users in the anonymous data set according to the joint extended behavior features to obtain the test set, specifically including: For the non-anonymous data subset, it is evenly divided into multiple intervals based on time. The frequency of the occurrence of the combined extended behavior feature elements for each user in each interval is counted respectively to obtain the feature matrix where p = len(F n ), and the sum of each column of the feature matrix is used as the validation set corresponding to other training sets; Random erasure feature matrix The records in are summed for each column to obtain training samples. Data augmentation is performed based on multiple random erasures, and the augmentation stops based on a set threshold for the number of samples in the augmented training set to obtain the augmented training set; Count the occurrence frequency of each user's joint extended behavior features in the anonymous data set, that is, the feature vector of each user, to obtain the test set.

5. The identity authentication method for enhancing behavioral fingerprint data according to claim 1, wherein Construct a user recognition model based on a convolutional neural network, including a normalization layer, a convolutional layer, a residual connection, and the splicing of multiple channel results. Input the training set after preliminary data augmentation into the user recognition model for data augmentation training to obtain the trained user recognition model, which specifically includes: Normalize the input training set and calculate the probability of the input training set for each known user, which is specifically expressed as: w i = σ(W * FC input + b); Among them, Conv represents a one-dimensional convolutional layer, is the activation function, Flatten means expanding the results of multiple channels of convolution into a one-dimensional vector, Dropout means that neurons are randomly deactivated with a set probability, W is the weight matrix of the fully connected layer, b is the bias of the fully connected layer, and σ(·) is the Softmax activation function; Input the training set after preliminary data augmentation into the user recognition model, and train and update the user recognition model with the cross-entropy with L2 regularization as the loss function; Iterate and loop multiple times until the loss function converges or the iteration times are reached to obtain the trained user recognition model.

6. The method for identity authentication with enhanced behavioral fingerprint data according to claim 1 or 5, characterized in that The training set after preliminary data augmentation is augmented using the Mixup algorithm, which is specifically expressed as: Randomly linearly combine two sample pairs (X k , y k ), (X l , y l ) in a batch multiple times to form new training sample pairs, which are specifically expressed as: Among them, γ∈[0,1] represents the parameter randomly drawn from Beta(α,α), and α represents the preset parameter.

7. The identity authentication method for enhancing behavioral fingerprint data according to claim 1, wherein The metric formula includes: a similarity or distance metric formula considering global features, a similarity or distance metric formula considering the specific values of behavioral features, and a similarity or distance formula considering the preference ranking of behaviors.

8. The identity authentication method for enhancing behavioral fingerprint data according to claim 1, wherein It also includes a step of normalizing the similarity score matrix, and normalizing the similarity score matrix constructed based on multiple similarity scores.

9. The identity authentication method for enhancing behavioral fingerprint data according to claim 1, wherein Fuse multiple classifiers to identify the identity of anonymous users. For the identification of a single anonymous user, the one with the greatest similarity to the known users is the identification result. For the identification tasks of multiple anonymous users known to be different users, the maximum sum of the similarities of all identification results is the final identification result.

10. An identity authentication system for enhancing behavioral fingerprint data, characterized in that, An identity authentication method for implementing the behavior fingerprint data augmentation described in any one of claims 1-9, including: a data preprocessing module, a feature extraction module, a data augmentation module, a user recognition model construction module, a user recognition model training module, a classifier weight value calculation module, a similarity score calculation module, and an anonymous user identity recognition module; The data preprocessing module is used to obtain historical behavior log records and anonymous behavior log records, perform data preprocessing on the historical behavior log records and anonymous behavior log records to obtain a non-anonymous data set and an anonymous data set, and divide the non-anonymous data set by time to obtain multiple non-anonymous data subsets; The feature extraction module is used to obtain joint extended behavior features based on the non-anonymous dataset according to the feature extraction algorithm SURE + ​ The data augmentation module is used to extract the feature vectors of users in multiple non-anonymous data subsets according to the jointly extended behavioral features, select one as the validation set, and use the rest as the training set, and perform preliminary data augmentation on the training set, and extract the feature vectors of users in the anonymous data set according to the jointly extended behavioral features to obtain the test set; The user recognition model construction module is used to construct a user recognition model based on a convolutional neural network, including a normalization layer, a convolutional layer, a residual connection, and the splicing of multiple channel results; The user recognition model training module is used to input the training set after preliminary data augmentation into the user recognition model for data augmentation training to obtain the trained user recognition model, input the validation set into the trained user recognition model, and adjust the hyperparameters and update the weight values of the user recognition model; The classifier weight value calculation module is used to calculate the weight value of the classifier. Records in a certain time interval of the validation set are randomly erased multiple times to obtain an enhanced validation set. The enhanced validation set is input into the trained user recognition model, and the recognition accuracy is obtained as the weight of this classifier of the user recognition model based on the convolutional neural network. At the same time, each measurement method also corresponds to a classifier respectively, and the recognition accuracy of the enhanced validation set is calculated as the weight value of the corresponding classifier based on different measurement formulas; The similarity score calculation module is used to calculate the similarity score of the test set samples. The training set and the validation set are combined and input into the user recognition model to retrain the user recognition model. The test set is input into the retrained user recognition model to obtain the similarity score between each anonymous user and each known user. Multiple validation sets are combined to obtain a template library, and the similarity score between the test set samples and the template library is calculated according to the measurement formula; The anonymous user identity recognition module is used to output the anonymous user identity recognition result. A similarity score matrix is constructed based on multiple similarity scores. According to the similarity matrix and the weights of each classifier, the weighted sum is used to obtain the final similarity matrix, and multiple classifiers are fused to recognize the anonymous user identity.

Images