Personalized short message verification code pushing method based on multi-task learning

Abstract

The invention discloses a personalized short message verification code pushing method based on multi-task learning, and the method comprises the following steps: collecting original user data of a user, and carrying out the preprocessing; inputting to an improved AdaTT network, and carrying out multi-modal feature fusion and weight distribution; calculating a risk probability value after executing dynamic gating fusion with the task proprietary features, and comparing the risk probability value with a preset threshold value; performing cross-task correction and selecting an optimal channel after dynamic gating fusion is performed on the channel task features; after dynamic gating fusion processing is executed, cross-task interaction is carried out, probability distribution is output, and an optimal time window is selected; performing joint judgment, constructing a joint loss function and updating an initial short message verification code pushing strategy; and executing the pushing operation of the short message verification code based on the personalized short message verification code pushing strategy. According to the method, the improved AdaTT network is adopted, personalized short message verification code intelligent pushing is achieved, and the method has the advantages of being high in safety, flexible in strategy and excellent in user experience.

Description

A Personalized SMS Verification Code Push Method Based on Multi-Task Learning Technical Field

[0001] This invention relates to the field of communication security and intelligent push, and in particular to a personalized SMS verification code push method based on multi-task learning. Background Technology

[0002] Existing SMS verification code push methods typically rely on fixed rules or a single risk assessment mechanism, such as using a few features like device identification or geographic location to make a threshold judgment to decide whether to send a verification code to the user. While these methods are simple to implement, they often suffer from insufficient accuracy when facing complex and ever-changing user behavior scenarios. They struggle to effectively differentiate risk levels among different users and in different business scenarios, easily leading to conflicts between security and user experience.

[0003] With the widespread adoption of mobile internet applications, user login and transaction scenarios are becoming increasingly diverse. Traditional single-task modeling methods struggle to meet multi-dimensional needs, resulting in inflexible push channels and rigid timing of message delivery. For example, some users are still forced to receive verification codes in low-risk scenarios, leading to a poor user experience; while in high-risk scenarios, overly lenient judgments may pose security risks.

[0004] Therefore, how to provide a personalized SMS verification code push method based on multi-task learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a personalized SMS verification code push method based on multi-task learning. This invention uses an improved AdaTT network to achieve intelligent push of personalized SMS verification codes, which has the advantages of high security, flexible strategy and excellent user experience.

[0006] A personalized SMS verification code push method based on multi-task learning according to an embodiment of the present invention includes the following steps:

[0007] Collect raw user data and preprocess it to generate a standardized user feature dataset;

[0008] The standardized user feature dataset is input into the multimodal user feature fusion module of the improved AdaTT network to perform multimodal feature fusion and weight allocation processing, generating a global user feature representation;

[0009] Dynamic gating fusion processing is performed on global user feature representation and task-specific features to calculate risk probability values ​​and compare them with preset thresholds to generate risk assessment results.

[0010] Dynamic gating fusion processing is performed on global user feature representation and channel task features, cross-task correction is performed in combination with risk assessment results, and the optimal channel is selected to generate channel selection results;

[0011] Dynamic gating fusion processing is performed based on global user characteristics. Cross-task interaction is carried out by combining risk assessment results and channel selection results. Probability distribution is output and the optimal time window is selected to generate timing prediction results.

[0012] The risk assessment results, channel selection results, and timing prediction results are jointly determined to construct a joint loss function and update the improved AdaTT network parameters to generate a personalized SMS verification code push strategy.

[0013] Based on a personalized SMS verification code push strategy, the SMS verification code is pushed out through the target SMS sending channel within the sending time window.

[0014] Optionally, the original user data includes historical login records, device identification information, geographical location information, usage time period information, and operation behavior sequence. The preprocessing includes performing data cleaning, outlier removal, and normalization.

[0015] Optionally, the generation of the global user feature representation specifically includes:

[0016] The standardized user feature dataset is input into the multimodal user feature fusion module of the improved AdaTT network. The embedding mapping unit performs matrix multiplication operations on various types of input data with the corresponding embedding matrix to obtain the embedding vector representation of each type of data. The dimensionality compression processing is performed on the embedding vector representation to obtain the corresponding low-dimensional dense vector set.

[0017] The low-dimensional dense vector set is input into the interactive attention computing unit to perform vector dot product operation on the low-dimensional dense vector set to obtain the corresponding dot product result. The dot product result is divided by the normalization factor root to obtain the scaled dot product result, and then exponential operation and normalization processing are performed to obtain the attention weight set.

[0018] The attention weights are multiplied element-wise by the corresponding low-dimensional dense vectors through the weighted vector generation unit to obtain weighted vectors. The above operation is repeated for each target low-dimensional dense vector in the low-dimensional dense vector set. All weighted vectors are arranged in the order of the target low-dimensional dense vectors to obtain a set of weighted feature vectors.

[0019] The weighted feature vector set is input into the feature aggregation unit, and a weighted summation process is performed on the weighted feature vector set to obtain the aggregated feature representation. This aggregated feature representation is then input into the shared feature extraction layer, where multi-layer convolution operations are performed to extract local feature vectors. Matrix multiplication and non-linear activation processing are then performed on the local feature vectors to generate a global user feature representation.

[0020] Optionally, compared to the original AdaTT network, the improved AdaTT network introduces a multimodal user feature fusion module at the input end to perform embedding mapping and interactive computation processing on the standardized user feature dataset to generate a global user feature representation. Task adaptive fusion modules are set between each subtask branch and the global user feature representation to perform dynamic gating fusion processing on the global user feature representation and its corresponding proprietary features, generating risk prediction fusion feature representation, channel selection fusion feature representation, and timing prediction fusion feature representation, respectively. A cross-task attention mechanism is introduced at the subtask output stage to perform interactive feature correction processing on the risk assessment results, channel selection results, and timing prediction results, generating a corrected score vector and a corrected timing feature representation. During the training stage, a joint loss function is constructed to perform weighted combination processing on the risk prediction subtask loss, channel selection subtask loss, and timing prediction subtask loss, generating the joint loss function and updating the parameters of the improved AdaTT network. Finally, the initial SMS verification code push strategy is optimized to generate a personalized SMS verification code push strategy.

[0021] The multimodal user feature fusion module includes an embedding mapping unit, an interaction attention calculation unit, a weighted vector generation unit, and a feature aggregation unit;

[0022] The task adaptive fusion module includes a task feature transformation unit, a gating coefficient generation unit, and a fusion calculation unit.

[0023] Optionally, the generation of the risk assessment results specifically includes:

[0024] The global user feature representation is input into the task adaptive fusion module. The task feature transformation unit performs matrix multiplication on the global user feature representation and the feature task weight matrix. After adding the feature task bias vector, it is input into the nonlinear activation function to perform element-wise activation operation to obtain task-specific features.

[0025] The gating coefficient generation unit concatenates the global user feature representation with the task-specific features into a combined vector, performs matrix multiplication on the combined vector and the gating weight matrix, adds the gating bias vector, and inputs it into the Sigmoid activation function for nonlinear processing to obtain the dynamic gating coefficients.

[0026] The fusion operation unit performs element-wise multiplication of the global user feature representation with the dynamic gating coefficient, performs element-wise multiplication of the task-specific feature with the difference between one and the dynamic gating coefficient, and performs element-wise summation of the results of the two multiplication operations to obtain the fused feature representation.

[0027] The fusion feature representation and risk prediction weight matrix are multiplied by matrix multiplication, and the risk prediction bias vector is added to obtain the linear transformation result.

[0028] The linear transformation result is input into the Sigmoid activation function to perform element-wise operations to obtain the risk probability value. The risk probability value is compared with a preset threshold. When the risk probability value is greater than the threshold, a high-risk label is output. When the risk probability value is less than or equal to the threshold, a low-risk label is output. The label is used as the output to generate the risk assessment result.

[0029] Optionally, the generation of the channel selection result specifically includes:

[0030] The global user feature representation is input into the task adaptive fusion module. The task feature transformation unit performs matrix multiplication with the channel task weight matrix, adds the channel task bias vector, and then inputs it into the nonlinear activation function to perform element-wise operation to obtain the channel task-specific features.

[0031] The channel dynamic gating coefficient is obtained by combining the global user feature representation with the channel task-specific features and performing nonlinear processing through the gating coefficient generation unit.

[0032] The fusion operation unit performs element-wise multiplication of the global user feature representation with the channel dynamic gating coefficient, performs element-wise multiplication of the channel task-specific feature with the difference between one and the channel dynamic gating coefficient, and sums the results of the two multiplication operations element-wise to obtain the channel fusion feature representation.

[0033] Perform matrix multiplication between the channel fusion feature representation and the candidate channel weight matrix, and add the candidate channel bias vector to obtain the candidate score vector;

[0034] A cross-task attention mechanism is adopted to concatenate the candidate score vector with the risk assessment result into an interactive input vector, and perform matrix multiplication with the cross-task weight matrix. After adding the cross-task bias vector, the input is fed into a non-linear activation function to obtain the interactive attention weight. The interactive attention weight and the candidate score vector are then multiplied element-wise and normalized to obtain the corrected score vector.

[0035] Perform a Softmax operation on the corrected score vector to obtain the probability value of each candidate SMS sending channel. Combine all candidate SMS sending channels with their corresponding probability values ​​to obtain a set of candidate SMS sending channels. Select the channel with the highest probability value from the set of candidate SMS sending channels as the output to generate the channel selection result.

[0036] Optionally, the generation of the timing prediction result specifically includes:

[0037] The global user feature representation is input into the task adaptive fusion module, and after processing by the task feature transformation unit, the gating coefficient generation unit and the fusion operation unit, the timing fusion feature representation is obtained.

[0038] A cross-task attention mechanism is adopted to concatenate the timing fusion feature representation, risk assessment results and channel selection results into an interactive input vector, and perform matrix multiplication with the cross-task weight matrix. After adding the cross-task bias vector, the vector is input into the Sigmoid activation function for normalization processing to obtain the timing interactive attention weight.

[0039] The timing interaction attention weights and timing fusion feature representations are multiplied element-wise, and the multiplication result is normalized to obtain the modified timing feature representation.

[0040] Matrix multiplication and nonlinear activation operations are performed on the modified timing feature representation to obtain the transmission timing probability distribution. The time window with the highest probability value in the transmission timing probability distribution is selected as the output to generate the timing prediction result.

[0041] Optionally, the generation of the personalized SMS verification code push strategy specifically includes:

[0042] The risk assessment results, channel selection results, and timing prediction results are input into the joint decision module, and vector concatenation operation is performed to obtain the joint input vector.

[0043] Perform matrix multiplication on the joint input vector and the decision weight matrix, and add the decision bias vector to obtain the linear transformation result;

[0044] The linear transformation result is input into the Softmax activation function to perform element-wise operations to obtain the joint probability distribution. The judgment category with the highest probability value is selected from the joint probability distribution. Based on the mapping relationship between the judgment category and the preset push strategy, the corresponding push method, push channel and push timing are output to generate the initial SMS verification code push strategy.

[0045] During the training phase, the losses for the risk prediction subtask, the channel selection subtask, and the timing prediction subtask are calculated separately. The three types of losses are then multiplied element-wise by the corresponding training weight factors and summed to obtain the joint loss function.

[0046] Backpropagation is performed on the joint loss function to update the parameters of the improved AdaTT network until the joint loss function reaches the maximum number of training rounds. Based on the updated model, the initial SMS verification code push strategy is optimized to generate a personalized SMS verification code push strategy.

[0047] The beneficial effects of this invention are:

[0048] This invention introduces a multimodal user feature fusion module into the improved AdaTT network, enabling unified processing of multi-source heterogeneous data such as user historical login records, device identification information, geographical location information, usage time period information, and operation behavior sequences to generate a global user feature representation. This improvement allows the model to comprehensively consider the user's multi-dimensional behavioral characteristics, rather than relying on a single-dimensional risk assessment, thereby enhancing the accuracy and stability of risk prediction and overcoming the problems of single input features and insufficient judgment accuracy in existing technologies.

[0049] Building upon this foundation, this invention establishes task-adaptive fusion modules between each subtask branch and the global user feature representation. Through a dynamic gating fusion mechanism, adaptive weighting is achieved between globally shared information and task-specific features, generating fusion feature representations for risk prediction, channel selection, and timing prediction, respectively. This structure avoids information interference and negative transfer problems in multi-task modeling, enabling the three tasks—risk prediction, channel selection, and timing prediction—to maintain their independence and specificity while sharing information. This significantly improves the synergy of multi-task learning and the robustness of the model.

[0050] Furthermore, this invention introduces a cross-task attention mechanism in the sub-task output stage. The risk assessment result is used as a reference input to correct channel selection, and then the risk assessment result and channel selection result are jointly used to correct timing prediction. Finally, a joint decision module maps the results of the three types of sub-tasks into a complete strategy consisting of push method, push channel, and sending time window. This mechanism establishes directed dependencies between tasks, ensuring logical consistency in the final generated personalized SMS verification code push strategy, and dynamically adjusts the push plan according to different users and risk scenarios. Combined with the joint loss function constructed during the training phase, the model can achieve a trade-off between security and user experience, and optimize the initial push strategy through continuous training and updates, thus solving the problems of insufficient personalization, rigid strategies, and lack of dynamic balance in existing technologies. Therefore, this invention significantly improves the security of the SMS verification code push system while greatly optimizing the user experience, demonstrating significant practical value and promotional significance. Attached Figure Description

[0051] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0052] Figure 1 is a flowchart of a personalized SMS verification code push method based on multi-task learning proposed in this invention;

[0053] Figure 2 is a schematic diagram of the improved AdaTT network structure of a personalized SMS verification code push method based on multi-task learning proposed in this invention. Detailed Implementation

[0054] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0055] Referring to Figure 1-2, a personalized SMS verification code push method based on multi-task learning includes the following steps:

[0056] Collect raw user data and preprocess it to generate a standardized user feature dataset;

[0057] The standardized user feature dataset is input into the multimodal user feature fusion module of the improved AdaTT network to perform multimodal feature fusion and weight allocation processing, generating a global user feature representation;

[0058] Dynamic gating fusion processing is performed on global user feature representation and task-specific features to calculate risk probability values ​​and compare them with preset thresholds to generate risk assessment results.

[0059] Dynamic gating fusion processing is performed on global user feature representation and channel task features, cross-task correction is performed in combination with risk assessment results, and the optimal channel is selected to generate channel selection results;

[0060] Dynamic gating fusion processing is performed based on global user characteristics. Cross-task interaction is carried out by combining risk assessment results and channel selection results. Probability distribution is output and the optimal time window is selected to generate timing prediction results.

[0061] The risk assessment results, channel selection results, and timing prediction results are jointly determined to construct a joint loss function and update the improved AdaTT network parameters to generate a personalized SMS verification code push strategy.

[0062] Based on a personalized SMS verification code push strategy, the SMS verification code is pushed out through the target SMS sending channel within the sending time window.

[0063] In this embodiment, the original user data includes historical login records, device identification information, geographical location information, usage time period information, and operation behavior sequence. The preprocessing includes performing data cleaning, outlier removal, and normalization.

[0064] In this embodiment, the generation of the global user feature representation specifically includes:

[0065] The standardized user feature dataset is input into the multimodal user feature fusion module of the improved AdaTT network. The embedding mapping unit performs matrix multiplication operations on various types of input data with the corresponding embedding matrix to obtain the embedding vector representation of each type of data. The embedding vector representation is subjected to dimensionality compression processing to obtain the corresponding low-dimensional dense vector set. The embedding matrix is ​​generated according to the preset dimensional parameters during the initialization phase and is obtained through parameter updates during the training process.

[0066] The low-dimensional dense vector set is input into the interactive attention computing unit to perform vector dot product operation on the low-dimensional dense vector set to obtain the corresponding dot product result. The dot product result is divided by the normalization factor root to obtain the scaled dot product result, and then exponential operation and normalization processing are performed to obtain the attention weight set. The specific process of obtaining the dot product result includes: sequentially selecting each low-dimensional dense vector from the low-dimensional dense vector set as the target low-dimensional dense vector, performing element-wise multiplication and summation on the target low-dimensional dense vector and all low-dimensional dense vectors in the low-dimensional dense vector set respectively to obtain the corresponding dot product result. The normalization factor is the square of the dimension of the low-dimensional dense vector.

[0067] The attention weights are multiplied element-wise by the corresponding low-dimensional dense vectors through the weighted vector generation unit to obtain weighted vectors. The above operation is repeated for each target low-dimensional dense vector in the low-dimensional dense vector set. All weighted vectors are arranged in the order of the target low-dimensional dense vectors to obtain a set of weighted feature vectors.

[0068] The weighted feature vector set is input into the feature aggregation unit, and a weighted summation process is performed on the weighted feature vector set to obtain the aggregated feature representation. This aggregated feature representation is then input into the shared feature extraction layer, where multi-layer convolution operations are performed to extract local feature vectors. Matrix multiplication and non-linear activation processing are then performed on the local feature vectors to generate a global user feature representation.

[0069] In this embodiment, compared with the original AdaTT network, the improved AdaTT network introduces a multimodal user feature fusion module at the input end to perform embedding mapping and interactive computation processing on the standardized user feature dataset to generate a global user feature representation. Task adaptive fusion modules are set between each subtask branch and the global user feature representation to perform dynamic gating fusion processing on the global user feature representation and the corresponding proprietary features, generating risk prediction fusion feature representation, channel selection fusion feature representation, and timing prediction fusion feature representation, respectively. A cross-task attention mechanism is introduced at the subtask output stage to perform interactive feature correction processing on the risk assessment results, channel selection results, and timing prediction results, generating a corrected score vector and a corrected timing feature representation. During the training stage, a joint loss function is constructed to perform weighted combination processing on the risk prediction subtask loss, channel selection subtask loss, and timing prediction subtask loss, generating a joint loss function and updating the parameters of the improved AdaTT network. Finally, the initial SMS verification code push strategy is optimized to generate a personalized SMS verification code push strategy.

[0070] The multimodal user feature fusion module includes an embedding mapping unit, an interaction attention calculation unit, a weighted vector generation unit, and a feature aggregation unit;

[0071] The task adaptive fusion module includes a task feature transformation unit, a gating coefficient generation unit, and a fusion calculation unit.

[0072] In this embodiment, the generation of the risk assessment result specifically includes:

[0073] The global user feature representation is input into the task adaptive fusion module. The task feature transformation unit performs matrix multiplication on the global user feature representation and the feature task weight matrix. After adding the feature task bias vector, it is input into the nonlinear activation function to perform element-wise activation operation to obtain task-specific features.

[0074] The gating coefficient generation unit concatenates the global user feature representation with the task-specific features into a combined vector, performs matrix multiplication on the combined vector and the gating weight matrix, adds the gating bias vector, and inputs it into the Sigmoid activation function for nonlinear processing to obtain the dynamic gating coefficients.

[0075] The fusion operation unit performs element-wise multiplication of the global user feature representation with the dynamic gating coefficient, performs element-wise multiplication of the task-specific feature with the difference between one and the dynamic gating coefficient, and performs element-wise summation of the results of the two multiplication operations to obtain the fused feature representation.

[0076] The fusion feature representation and risk prediction weight matrix are multiplied by matrix multiplication, and the risk prediction bias vector is added to obtain the linear transformation result.

[0077] The linear transformation result is input into a nonlinear activation function to perform element-wise operations to obtain a risk probability value. The risk probability value is compared with a preset threshold. When the risk probability value is greater than the threshold, a high-risk label is output. When the risk probability value is less than or equal to the threshold, a low-risk label is output. The label is used as the output to generate a risk assessment result. The risk probability value is equal to one divided by one plus the negative first power of the natural exponential function. The input of the natural exponential function is the linear transformation result.

[0078] In this embodiment, the generation of the channel selection result specifically includes:

[0079] The global user feature representation is input into the task adaptive fusion module. The task feature transformation unit performs matrix multiplication on the global user feature representation and the channel task weight matrix. After adding the channel task bias vector, it is input into the nonlinear activation function to perform element-wise operations to obtain the channel task-specific features.

[0080] The global user feature representation and channel task-specific features are combined and nonlinearly processed by the gating coefficient generation unit to obtain the channel dynamic gating coefficient. The specific process of obtaining the channel dynamic gating coefficient includes: concatenating the global user feature representation and channel task-specific features into a combined vector, performing matrix multiplication on the combined vector and the gating weight matrix, adding the gating bias vector, and then inputting it into the Sigmoid activation function for nonlinear operation.

[0081] The fusion operation unit performs element-wise multiplication of the global user feature representation with the channel dynamic gating coefficient, performs element-wise multiplication of the channel task-specific feature with the difference between one and the channel dynamic gating coefficient, and sums the results of the two multiplication operations element-wise to obtain the channel fusion feature representation.

[0082] Perform matrix multiplication between the channel fusion feature representation and the candidate channel weight matrix, and add the candidate channel bias vector to obtain the candidate score vector;

[0083] A cross-task attention mechanism is adopted, which concatenates the candidate score vector with the risk assessment result into an interactive input vector, performs matrix multiplication with the cross-task weight matrix, adds the cross-task bias vector, and inputs it into a nonlinear activation function to obtain interactive attention weights. The interactive attention weights and candidate score vectors are then multiplied element-wise and normalized. The normalized weighted candidate score vector is output as the cross-task corrected candidate score vector to obtain the corrected score vector. The cross-task weight matrix is ​​randomly generated according to the dimension parameter of the interactive input vector, and the parameter value is continuously adjusted during the training phase by combining the backpropagation algorithm with the gradient update of the loss function.

[0084] A Softmax operation is performed on the corrected score vector to obtain the probability value of each candidate SMS sending channel. All candidate SMS sending channels and their corresponding probability values ​​are combined to obtain a set of candidate SMS sending channels. The channel with the highest probability value in the set of candidate SMS sending channels is selected as the output to generate a channel selection result. The channel selection result includes the target SMS sending channel, the maximum probability value of the target SMS sending channel in the set of candidate SMS sending channels, and the identification information of the target SMS sending channel.

[0085] In this embodiment, the generation of the timing prediction result specifically includes:

[0086] The global user feature representation is input into the task adaptive fusion module, and after processing by the task feature transformation unit, the gating coefficient generation unit and the fusion operation unit, the timing fusion feature representation is obtained.

[0087] A cross-task attention mechanism is adopted to concatenate the timing fusion feature representation, risk assessment result, and channel selection result into an interactive input vector. Then, matrix multiplication is performed with the cross-task weight matrix, and after adding the cross-task bias vector, the input is fed into the Sigmoid activation function for normalization processing to obtain the timing interactive attention weight. The concatenation specifically includes: arranging and combining the timing fusion feature representation, risk assessment result, and channel selection result according to a preset concatenation order, and connecting each input vector in parallel under the same batch dimension according to the feature dimension to obtain the interactive input vector.

[0088] The timing interaction attention weights and timing fusion feature representations are multiplied element-wise, and the multiplication result is normalized to obtain the modified timing feature representation.

[0089] Matrix multiplication and nonlinear activation operations are performed on the modified timing feature representation to obtain the transmission timing probability distribution. The time window with the highest probability value in the transmission timing probability distribution is selected as the output to generate the timing prediction result.

[0090] In this embodiment, the generation of the personalized SMS verification code push strategy specifically includes:

[0091] The risk assessment results, channel selection results, and timing prediction results are input into the joint decision module, and vector concatenation operation is performed to obtain the joint input vector.

[0092] Perform matrix multiplication on the joint input vector and the decision weight matrix, and add the decision bias vector to obtain the linear transformation result;

[0093] The linear transformation result is input into the Softmax activation function to perform element-wise operations to obtain a joint probability distribution. The judgment category with the highest probability value is selected from the joint probability distribution. Based on the mapping relationship between the judgment category and the preset push strategy, the corresponding push method, push channel and push timing are output to generate an initial SMS verification code push strategy. The judgment categories include normal push category, delayed push category, multi-channel push category and reject push category, which correspond to the personalized SMS verification code push strategies generated under low risk, medium risk, high risk and high risk conditions, respectively. The preset push strategy mapping relationship is the correspondence between the judgment category and the push method, push channel and push timing. It is pre-configured in the system design stage according to the security control strategy and user experience requirements, and stored in the form of rule table or parameterization.

[0094] During the training phase, the losses for the risk prediction subtask, channel selection subtask, and timing prediction subtask are calculated separately. The three types of losses are multiplied element-wise with their corresponding training weight factors and then summed to obtain a joint loss function. The risk prediction subtask loss is calculated by inputting the risk probability value and the corresponding real risk label into the cross-entropy loss function. The channel selection subtask loss is calculated by inputting the candidate SMS sending channel probability distribution and the corresponding real channel label into the cross-entropy loss function. The timing prediction subtask loss is calculated by inputting the sending timing probability distribution and the corresponding real timing label into the cross-entropy loss function.

[0095] Backpropagation is performed on the joint loss function to update the parameters of the improved AdaTT network until the joint loss function reaches the maximum number of training rounds. Based on the updated model, the initial SMS verification code push strategy is optimized to generate a personalized SMS verification code push strategy.

[0096] Example 1:

[0097] To verify the feasibility of this invention in practice, it was applied to the user authentication system of a large internet service company. This company's online platform boasts a massive registered user base covering the entire country, with an extremely high daily number of logins and transaction requests. For a long time, the company's SMS verification code push mechanism has relied primarily on fixed threshold judgment strategies. For example, when a new device or login from a different location is detected, the system directly triggers an SMS verification code push. While this method provides some degree of security for user accounts, it often leads to two problems due to neglecting individual user characteristics and behavioral differences: firstly, frequent triggering of SMS verification codes in low-risk situations degrades the user experience, causing users to feel cumbersome and inconvenienced during high-frequency use; secondly, in some high-risk situations, due to limited judgment dimensions, there is still a risk that may not be identified in a timely manner. This contradiction between security and user experience has long plagued the company's authentication system.

[0098] In this scenario, the method of this invention is deployed in the background process of an authentication system. When a user initiates a login or critical operation request, the system first collects multi-source data, including the user's historical login records, device identification information, geographical location information, usage time period information, and operation behavior sequences, and preprocesses this data to generate a standardized user feature dataset. This data is then input into an improved AdaTT network, where embedding mapping and feature interaction are completed in the multimodal user feature fusion module, thereby generating a unified global user feature representation. This process enables the system to comprehensively characterize user behavior habits and usage characteristics, rather than relying on single-dimensional data for judgment.

[0099] Subsequently, in the risk prediction branch, the system utilizes a task-adaptive fusion module to dynamically gating and fusing global user feature representations with risk task-specific features. It calculates the risk probability of the user's current action and compares it with a preset threshold to generate a risk assessment result. If the risk probability is high, this result is passed to the channel selection and timing prediction branches as an important basis for subsequent adjustments. During channel selection, the system not only utilizes global user features but also combines the risk assessment result to perform cross-task adjustments, calculating the probability of candidate SMS sending channels and ultimately selecting the optimal target channel. In timing prediction, the system similarly inputs the risk assessment result and channel selection result based on global features. After cross-task interaction and probability distribution calculation, it outputs the most suitable verification code sending time window. Through this multi-task collaboration and layered adjustments, the system can dynamically adjust the push method and timing while ensuring security, forming a complete personalized SMS verification code push strategy.

[0100] To further verify the effectiveness, the company ran the method of this invention in parallel with its existing threshold determination method for a period of time, applying them to different user groups and business lines. During this actual operation, the authentication process using the method of this invention was able to more reasonably distinguish between low-risk and high-risk users. In the login process for low-risk users, this invention reduced unnecessary SMS verification code pushes, and users clearly felt that the authentication process was simpler and smoother when continuously using the service, reducing unnecessary operational steps. In the login or transaction requests of high-risk users, this invention, through risk prediction and cross-task correction using a multi-task learning model, can promptly trigger stricter authentication policies, effectively preventing potential account security issues.

[0101] Table 1. Performance Comparison of the Invention and Traditional Personalized SMS Verification Code Push Methods

[0102] Indicator Category | Traditional Method | Invention Method | Risk Identification Accuracy (%) | 83.6 | 9 | 4.2 | Channel Selection Accuracy (%) | 80.4 | 9 | 2.1 | Timing Prediction Accuracy (%) | 79.3 | 9 | 1.5 | False Alarm Rate (%) | 9.8 | 3.6 | Anomaly Response Latency (ms) | 212 | 9 | 5 | User Experience Satisfaction (%) | 82.7 | 9 | 3.4 | Model Convergence Rounds (epochs) | 5627 surface

[0103] As can be clearly seen from Table 1, the method of the present invention is superior to the traditional method in many indicators.

[0104] In terms of risk identification accuracy, the traditional method achieves 83.6%, while the method of this invention achieves 94.2%, representing a significant improvement. This invention introduces a dynamic gating mechanism into the task adaptive fusion module and combines it with multimodal feature input, enabling the model to more comprehensively characterize the user's true risk features. This improves the risk identification capability and reduces over-triggered detections in low-risk situations and missed detections in high-risk situations.

[0105] Regarding the accuracy of channel selection, traditional methods achieve 80.4%, while the method of this invention reaches 92.1%. This invention introduces a cross-task attention mechanism in the channel selection process, interactively correcting the risk assessment results with the channel selection branch. This allows the model to adaptively select the best channel based on different risk levels, avoiding the selection bias that is prone to occur in traditional methods under a single dimension, thereby significantly improving the rationality and accuracy of channel selection.

[0106] Regarding the accuracy of timing prediction, traditional methods achieve 79.3%, while the method of this invention achieves 91.5%. This invention introduces cross-task interaction into the timing prediction branch, combining risk assessment results and channel selection results as inputs to form a corrected timing feature representation, ensuring a more accurate time window for CAPTCHA push notifications. This allows push notifications to avoid both premature triggering causing interference and delayed triggering affecting efficiency, offering a significant advantage over traditional methods.

[0107] Regarding the false positive rate, the traditional method has a false positive rate of 9.8%, while the method of this invention significantly reduces it to 3.6%. This invention utilizes multimodal user feature fusion and dynamic risk assessment to effectively reduce invalid CAPTCHA pushes, thereby lowering the frequency of user disruption. The reduced false positive rate directly reflects the model's optimization effect on user experience in real-world applications, avoiding the user churn problem caused by high false positives in traditional methods.

[0108] Regarding anomaly response latency, the traditional method takes 212 milliseconds, while the method of this invention reduces it to 95 milliseconds, resulting in a significant improvement in response speed. This invention reduces feature computation redundancy by optimizing the network structure and introducing a task-adaptive fusion module. Simultaneously, it utilizes attention weights in cross-task interactions to achieve rapid information aggregation, thereby significantly shortening the response time under abnormal conditions and providing the system with more real-time security protection capabilities in high-concurrency environments.

[0109] In terms of user experience satisfaction, traditional methods achieve 82.7%, while the method of this invention reaches 93.4%. This invention, through dynamic risk assessment, precise channel selection, and reasonable timing prediction, reduces unnecessary CAPTCHA interference for users under low-risk conditions and provides higher security guarantees under high-risk conditions, thereby improving the overall experience. This level of experience optimization is something traditional methods cannot achieve simultaneously.

[0110] Regarding the number of convergence rounds, traditional methods require 56 rounds, while the method of this invention requires only 27 rounds, significantly shortening the training convergence time. This invention utilizes multi-task learning and joint loss function design, enabling different sub-tasks to share information and mutually promote each other. This reduces model oscillations and overfitting during training, accelerates convergence, reduces computational overhead, and improves system deployment efficiency.

[0111] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A personalized SMS verification code push method based on multi-task learning, characterized in that, The process includes the following steps: collecting raw user data and preprocessing it to generate a standardized user feature dataset; inputting the standardized user feature dataset into the multimodal user feature fusion module of the improved AdaTT network to perform multimodal feature fusion and weight allocation processing to generate a global user feature representation; Dynamic gating fusion processing is performed on global user feature representation and task-specific features to calculate risk probability values ​​and compare them with preset thresholds to generate risk assessment results. Dynamic gating fusion processing is performed on global user feature representation and channel task features, cross-task correction is performed in combination with risk assessment results, and the optimal channel is selected to generate channel selection results; Dynamic gating fusion processing is performed based on global user characteristics. Cross-task interaction is carried out by combining risk assessment results and channel selection results to output probability distribution and select the optimal time window to generate timing prediction results. The risk assessment results, channel selection results and timing prediction results are jointly judged to construct a joint loss function and update the improved AdaTT network parameters to generate a personalized SMS verification code push strategy. Based on a personalized SMS verification code push strategy, the SMS verification code is pushed out through the target SMS sending channel within the sending time window.

2. The personalized SMS verification code push method based on multi-task learning according to claim 1, characterized in that, The original user data includes historical login records, device identification information, geographical location information, usage time period information, and operation behavior sequence. The preprocessing includes performing data cleaning, outlier removal, and normalization.

3. The personalized SMS verification code push method based on multi-task learning according to claim 1, characterized in that, The generation of the global user feature representation specifically includes: inputting a standardized user feature dataset into the multimodal user feature fusion module of the improved AdaTT network; performing matrix multiplication operations between various input data and their corresponding embedding matrices through an embedding mapping unit to obtain embedding vector representations of various data types; performing dimensionality compression processing on the embedding vector representations to obtain the corresponding low-dimensional dense vector set; inputting the low-dimensional dense vector set into the interactive attention calculation unit to perform vector dot product operations on the low-dimensional dense vector set to obtain the corresponding dot product result; dividing the dot product result by the normalization factor root to obtain the scaled dot product result; and performing exponential operations and normalization processing to obtain the attention. The weighted vector set is generated by performing element-wise multiplication of the attention weights with the corresponding low-dimensional dense vectors in the weighted vector generation unit. This process is repeated for each target low-dimensional dense vector in the set of low-dimensional dense vectors, and all weighted vectors are arranged in the order of the target low-dimensional dense vectors to obtain a weighted feature vector set. The weighted feature vector set is then input to the feature aggregation unit, where a weighted summation is performed to obtain an aggregated feature representation. This aggregated feature representation is then input to the shared feature extraction layer, where multi-layer convolution operations are performed to extract local feature vectors. Matrix multiplication and non-linear activation are then applied to the local feature vectors to generate a global user feature representation.

4. The personalized SMS verification code push method based on multi-task learning according to claim 3, characterized in that, Compared to the original AdaTT network, the improved AdaTT network introduces a multimodal user feature fusion module at the input end. This module performs embedding mapping and interactive computation on the standardized user feature dataset to generate a global user feature representation. Task-adaptive fusion modules are set between each subtask branch and the global user feature representation to perform dynamic gating fusion processing on the global user feature representation and its corresponding proprietary features, generating risk prediction fusion feature representations, channel selection fusion feature representations, and timing prediction fusion feature representations, respectively. A cross-task attention mechanism is introduced at the subtask output stage to perform interactive feature correction processing on the risk assessment results, channel selection results, and timing prediction results, generating corrected score vectors and corrected timing feature representations. During the training phase, a joint loss function is constructed, and weighted combination processing is performed on the risk prediction subtask loss, channel selection subtask loss, and timing prediction subtask loss to generate the joint loss function and update the parameters of the improved AdaTT network. Finally, the initial SMS verification code push strategy is optimized to generate a personalized SMS verification code push strategy. The multimodal user feature fusion module includes an embedding mapping unit, an interactive attention computation unit, a weighted vector generation unit, and a feature aggregation unit. The task adaptive fusion module includes a task feature transformation unit, a gating coefficient generation unit, and a fusion calculation unit.

5. The personalized SMS verification code push method based on multi-task learning according to claim 1, characterized in that, The generation of the risk assessment result specifically includes: inputting the global user feature representation into the task adaptive fusion module; performing matrix multiplication between the global user feature representation and the feature task weight matrix through the task feature transformation unit; adding the feature task bias vector; and inputting the result into a nonlinear activation function for element-wise activation to obtain task-specific features; concatenating the global user feature representation and task-specific features into a combined vector through the gating coefficient generation unit; performing matrix multiplication between the combined vector and the gating weight matrix; adding the gating bias vector; and inputting the result into a Sigmoid activation function for nonlinear processing to obtain dynamic gating coefficients; and then using the fusion operation unit to process the global user feature representation... The system performs element-wise multiplication with the dynamic gating coefficient, multiplying the task-specific feature with the difference between one and the dynamic gating coefficient, and summing the results of the two multiplications element-wise to obtain the fused feature representation. The fused feature representation is then multiplied with the risk prediction weight matrix, and a risk prediction bias vector is added to obtain the linear transformation result. This linear transformation result is input into the Sigmoid activation function for element-wise operation to obtain the risk probability value. This risk probability value is compared with a preset threshold; a high-risk label is output when the risk probability value is greater than the threshold, and a low-risk label is output when the risk probability value is less than or equal to the threshold. These labels are then used as output to generate the risk assessment result.

6. The personalized SMS verification code push method based on multi-task learning according to claim 1, characterized in that, The generation of the channel selection result specifically includes: inputting the global user feature representation into the task adaptive fusion module; performing matrix multiplication with the channel task weight matrix through the task feature transformation unit; adding the channel task bias vector and inputting it into a nonlinear activation function for element-wise operation to obtain channel task-specific features; combining and nonlinearly processing the global user feature representation and channel task-specific features through the gating coefficient generation unit to obtain channel dynamic gating coefficients; performing element-wise multiplication of the global user feature representation and channel dynamic gating coefficients through the fusion operation unit; performing element-wise multiplication of the channel task-specific features with the difference between one and the channel dynamic gating coefficients; and summing the results of the two multiplication operations element-wise to obtain the channel fusion feature representation; and then combining the channel fusion feature representation with... The candidate channel weight matrix is ​​multiplied using matrix multiplication and then supplemented with a candidate channel bias vector to obtain a candidate score vector. A cross-task attention mechanism is employed, concatenating the candidate score vector with the risk assessment result to form an interaction input vector. This vector is then multiplied with the cross-task weight matrix, and after adding the cross-task bias vector, it is input into a non-linear activation function to obtain interaction attention weights. The interaction attention weights and candidate score vectors are then multiplied element-wise and normalized to obtain a corrected score vector. A softmax operation is performed on the corrected score vector to obtain the probability value of each candidate SMS sending channel. All candidate SMS sending channels and their corresponding probability values ​​are combined to obtain a candidate SMS sending channel set. The channel with the highest probability value is selected from this set as the output, generating the channel selection result.

7. The personalized SMS verification code push method based on multi-task learning according to claim 1, characterized in that, The generation of the timing prediction result specifically includes: inputting the global user feature representation into the task adaptive fusion module, which processes it through the task feature transformation unit, the gating coefficient generation unit, and the fusion operation unit to obtain the timing fusion feature representation; employing a cross-task attention mechanism, concatenating the timing fusion feature representation, risk assessment result, and channel selection result into an interactive input vector, performing matrix multiplication with the cross-task weight matrix, adding the cross-task bias vector, and inputting it into the Sigmoid activation function for normalization processing to obtain the timing interactive attention weight; performing element-wise multiplication of the timing interactive attention weight and the timing fusion feature representation, and performing normalization processing on the multiplication result to obtain the modified timing feature representation; performing matrix multiplication and nonlinear activation operations on the modified timing feature representation to obtain the transmission timing probability distribution, and selecting the time window with the highest probability value from the transmission timing probability distribution as the output to generate the timing prediction result.

8. The personalized SMS verification code push method based on multi-task learning according to claim 1, characterized in that, The generation of the personalized SMS verification code push strategy specifically includes: inputting the risk assessment result, channel selection result, and timing prediction result into the joint decision module, performing vector concatenation operation to obtain a joint input vector; performing matrix multiplication operation between the joint input vector and the decision weight matrix, and adding the decision bias vector to obtain a linear transformation result; inputting the linear transformation result into the Softmax activation function to perform element-wise operation to obtain a joint probability distribution, selecting the decision category with the highest probability value from the joint probability distribution, and outputting the corresponding push method, push channel, and push timing according to the mapping relationship between the decision category and the preset push strategy to generate an initial SMS verification code push strategy; calculating the risk prediction subtask loss, channel selection subtask loss, and timing prediction subtask loss respectively during the training phase, multiplying the three types of losses element-wise with the corresponding training weight factors and accumulating them to obtain a joint loss function; performing backpropagation to update the parameters of the improved AdaTT network on the joint loss function until the joint loss function reaches the maximum training rounds, and optimizing the initial SMS verification code push strategy based on the updated model to generate a personalized SMS verification code push strategy.

Images