Sensitive information encryption transmission and analysis method in financial risk control scene

Abstract

The application belongs to the field of information security, and discloses a sensitive information encryption transmission and analysis method in a financial risk control scene, which comprises the following steps: step 1, performing field set division on original fields, respectively encrypting each field based on the field set, and obtaining an encrypted field set; step 2, calculating a risk embedding representation based on the encrypted field set; step 3, constructing an original graph based on the risk embedding representation, performing perturbation processing on the original graph, and obtaining a perturbation graph; step 4, extracting a time behavior feature representation in a time dimension based on the perturbation graph; step 5, calculating a risk score based on the risk embedding representation and the time behavior feature representation, and executing a risk control strategy based on the risk score. The application can continuously track dynamic risk behaviors and calculate risk scores. The overall scheme effectively improves the perception ability of a risk control model to complex attack behaviors while maintaining a high level of privacy protection, and is particularly suitable for financial risk control scenes with multi-party cooperation, cross-domain modeling, high regulatory compliance requirements.

Description

Technical Field

[0001] This invention relates to the field of information security, and in particular to a method for encrypted transmission and analysis of sensitive information in financial risk control scenarios. Background Technology

[0002] In modern financial service systems, risk control systems serve as core infrastructure for ensuring transaction security, preventing fraud, and combating money laundering. Their development level directly impacts the business stability and customer trust of financial institutions. However, to improve the accuracy of risk identification, an increasing number of financial risk control systems rely on machine learning and deep learning models to jointly model and analyze multi-dimensional data, including but not limited to user identity information, transaction behavior trajectories, account relationship structures, device characteristics, geographical location, and time patterns. Because this data contains a large amount of highly sensitive personal privacy information, such as ID numbers, bank accounts, mobile phone numbers, transaction times and amounts, and even behavioral sequences and network paths, modeling and analysis must simultaneously satisfy two difficult-to-reconcile goals: on the one hand, the model needs to extensively collect and integrate this information to ensure identification performance; on the other hand, it must strictly protect user privacy and data security, avoiding any form of leakage risk, especially in cross-institutional and multi-party collaborative scenarios.

[0003] Currently, the mainstream solutions can be roughly categorized into three types: First, all sensitive information is encrypted before transmission or computation, such as using homomorphic encryption and secure multi-party computation (SMPC). However, these technologies have extremely high computational costs and impose many limitations on model structures, making them difficult to deploy flexibly. Second, federated learning is used to keep data locally in each institution, sharing only model parameters to avoid data leakage. However, federated learning has insufficient modeling capabilities when dealing with complex relational data (such as transaction graphs and user behavior chains), and it is difficult to uniformly schedule under heterogeneous data and heterogeneous encryption mechanisms. Third, modeling is performed after reducing the sensitivity of data through pseudo-anonymization, desensitization, and other data processing methods. However, these methods are still inherently vulnerable to re-identification attacks and do not truly achieve structural protection. Especially in scenarios such as transaction chain tracing, pseudo-IDs can easily be reconstructed into the original path. Furthermore, existing systems generally lack specialized modeling mechanisms for dynamic behavioral patterns such as "chain-like abnormal behavior," "off-peak trading," and "device collusion." This allows attackers to effectively circumvent current static risk control rules through cross-account, time-segmented, and geographically-based behavioral combinations. Therefore, how to complete the comprehensive risk modeling of high-dimensional transaction behavior, graph-structured paths, and dynamic time-series patterns under encrypted data conditions remains a key unresolved technical challenge in the current financial risk control system. Summary of the Invention

[0004] The purpose of this invention is to disclose a method for encrypted transmission and analysis of sensitive information in financial risk control scenarios, thereby solving the technical problems pointed out in the background art.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] This invention provides a method for secure management of product image data, including:

[0007] Step 1: Divide the original fields into a field set, and encrypt each field based on the field set to obtain the encrypted field set;

[0008] Step 2: Calculate the risk embedding representation based on the set of encrypted fields;

[0009] Step 3: Construct the original graph based on the risk embedding representation, and perturb the original graph to obtain the perturbed graph;

[0010] Step 4: Extract temporal behavior feature representations in the time dimension based on the perturbation graph;

[0011] Step 5: Calculate the risk score based on the risk embedding representation and the time behavior feature representation, and execute the risk control strategy based on the risk score;

[0012] This involves dividing the original fields into a field set, including:

[0013] Obtain the privacy sensitivity and modeling contribution of each field separately;

[0014] Priority scores are calculated based on privacy sensitivity and modeling contribution.

[0015] The set of fields to which a field belongs is determined based on priority scores.

[0016] Furthermore, a priority score is calculated based on privacy sensitivity and modeling contribution, including:

[0017] Use respectively and Representation field Privacy sensitivity and modeling contribution;

[0018] based on and Calculate priority scores.

[0019] Furthermore, the set of fields to which a field belongs, based on priority scoring, includes:

[0020] Set the first threshold Second threshold ;

[0021] For fields ,like If it is greater than or equal to the first threshold, then... Store a set of fields with a high level of encryption. ;

[0022] like If it is greater than or equal to the second threshold and less than the first threshold, then... The set of fields with the encryption level is stored. ;

[0023] like If it is less than the second threshold, then Store a set of fields with a low encryption level. .

[0024] Furthermore, each field is encrypted separately based on the field set, including:

[0025] right The fields in the code are encrypted field by field using a homomorphic encryption library;

[0026] right The fields in the code are encrypted by adding Laplace noise;

[0027] right The fields in the data are encrypted using either pseudo-anonymization or plaintext processing.

[0028] Furthermore, obtain the set of encrypted fields, including:

[0029] They are respectively for , and Encrypt the fields in the data to obtain a collection. , and ;

[0030] based on , and Get the collection of encrypted fields.

[0031] Furthermore, step 2 includes:

[0032] Depend on The elements in form a vector , ;

[0033] Depend on The elements in form a vector , ;

[0034] Depend on The elements in form a vector , ;

[0035] , and They represent , and The dimension;

[0036] against Set sub-model structure ;

[0037] against Set sub-model structure ;

[0038] against Set sub-model structure ;

[0039] Will , and They are respectively represented as , and ;

[0040] based on , and Compute the risk embedding representation z.

[0041] Furthermore, step 3 includes:

[0042] Construct the original graph G:

[0043] Let V represent the set of nodes, and each node... This represents a financial account;

[0044] Let Z represent the set of node features;

[0045] Construct a set E of edges based on business logs;

[0046] The original graph G is then represented as ;

[0047] Perturb the original graph G:

[0048] The node ID of each node in the original graph G is perturbed to obtain the set of perturbed nodes. ;

[0049] Perturb each edge in the original graph G to obtain the set of perturbed edges. ;

[0050] For Z, either keep the node features in Z unchanged or perform dimensionality reduction on the features in Z to obtain the set. ;

[0051] based on , and Constructing a perturbation graph .

[0052] Furthermore, step 4 includes:

[0053] For each node ,extract Timestamps of all adjacent edges Construct transaction time series ;

[0054] based on Construct each node Time difference series ;

[0055] Using the Temporal Difference Convolutional Network (TDCN) Calculate and obtain temporal behavior feature representation .

[0056] Furthermore, risk scores are calculated based on risk embedding representations and temporal behavioral feature representations, including:

[0057] Constructing a joint risk input vector ;

[0058] based on Calculate risk score .

[0059] Furthermore, risk control strategies are implemented based on risk scores, including:

[0060] like If so, a high-risk risk control strategy will be implemented;

[0061] like If so, then implement a medium-risk risk control strategy;

[0062] like If so, a low-risk risk control strategy will be implemented;

[0063] and These are the first risk scoring threshold and the second risk scoring threshold, respectively.

[0064] Beneficial effects:

[0065] This invention addresses the systemic contradiction in existing technologies regarding sensitive information protection and complex behavior analysis, proposing a holistic solution driven by scenarios, centered on structural decoupling, and aimed at behavioral robustness. First, based on the dual assessment results of the business sensitivity and risk value of input features, this invention establishes a fine-grained feature-layered encryption mechanism to control data privacy protection strategies at the source. Through differentiated model structure design, it ensures that various encrypted features can participate in effective modeling under suitable paths. Second, to address the inability of traditional models to handle chain relationships and graph structure leakage, this invention constructs a graph representation mechanism with structural perturbation protection capabilities, enabling the risk control system to identify abnormal interaction patterns between potential accounts without exposing the real network structure. Furthermore, to address the problem of attackers evading detection through transaction time-series perturbations or behavioral staggering, this invention introduces a time-dimensional behavioral robustness analysis strategy, supporting non-explicit modeling of transaction time series under encrypted conditions, thereby achieving continuous tracking and risk scoring of dynamic risk behavior. While maintaining a high level of privacy protection, the overall solution effectively enhances the risk control model's ability to detect complex attack behaviors, making it particularly suitable for financial risk control scenarios involving multi-party collaboration, cross-domain modeling, and high regulatory compliance requirements. Attached Figure Description

[0066] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0067] Figure 1 This is a schematic diagram of the image data security management method for the product of the present invention. Detailed Implementation

[0068] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0069] like Figure 1As shown in one embodiment, the present invention provides a method for secure management of product captured image data, including:

[0070] Step 1: Divide the original fields into a field set, and encrypt each field based on the field set to obtain the encrypted field set;

[0071] Step 2: Calculate the risk embedding representation based on the set of encrypted fields;

[0072] Step 3: Construct the original graph based on the risk embedding representation, and perturb the original graph to obtain the perturbed graph;

[0073] Step 4: Extract temporal behavior feature representations in the time dimension based on the perturbation graph;

[0074] Step 5: Calculate the risk score based on the risk embedding representation and the time behavior feature representation, and execute the risk control strategy based on the risk score;

[0075] This involves dividing the original fields into a field set, including:

[0076] Obtain the privacy sensitivity and modeling contribution of each field separately;

[0077] Priority scores are calculated based on privacy sensitivity and modeling contribution.

[0078] The set of fields to which a field belongs is determined based on priority scores.

[0079] Step 1 involves systematically evaluating the raw data fields in the financial risk control system, considering both privacy sensitivity and modeling value, determining the field set to which each field belongs, and performing encryption operations to form an encrypted feature set for subsequent model analysis. .

[0080] Furthermore, for the sample set as Each sample ,Include There are 1, N fields, where v ∈ [1, N]. These fields are divided into five categories based on their business origin, and the Chinese meaning and risk control function of each category are as follows:

[0081] : Identity field, used to uniquely identify users, involving highly sensitive identity information.

[0082] User ID, a unique identifier generated by the platform;

[0083] : Phone number hash value, used for de-identified storage and reverse lookup;

[0084] Bank card number or its encrypted form.

[0085] : Transaction-related fields used to describe the characteristics of a user's transaction behavior.

[0086] Transaction amount;

[0087] : The time when the transaction occurred;

[0088] Merchant ID or payment channel identifier.

[0089] : Device field, which records the terminal or network environment used by the user.

[0090] Unique device identifier (such as IMEI or IDFV);

[0091] Operating system type (e.g., Android, iOS);

[0092] : The IP address used when logging in or making a transaction.

[0093] Geographic fields reflect the spatial location information of transactions or devices.

[0094] GPS coordinates, city code, or address summary;

[0095] : Region / Province / Grid Number.

[0096] : Environment field, used to express contextual information about external environmental factors.

[0097] : Whether to mark holidays;

[0098] Weather code (e.g., sunny / rainy / snowy).

[0099] Furthermore, the privacy sensitivity and modeling contribution of each field are obtained separately, including:

[0100] The process of obtaining the privacy sensitivity of fields includes:

[0101] For each field Define its privacy sensitivity This is set by safety regulations. For example:

[0102] ;

[0103] ;

[0104] ;

[0105] The process of obtaining modeling contribution includes:

[0106] Modeling contribution It is composed of information gain and volatility stability weighting:

[0107] ;

[0108] in:

[0109] Representation field Information gain during GBDT model training;

[0110] Representation field The normalized volatility, i.e., the standard deviation / mean of the historical sample;

[0111] This represents the penalty factor, with a value range of [value range missing]. This is used to control the GBDT model's suppression of unstable features, and is preferred. It can be 0.3.

[0112] Furthermore, a priority score is calculated based on privacy sensitivity and modeling contribution, including:

[0113] Use respectively and Representation field Privacy sensitivity and modeling contribution;

[0114] based on and To calculate the priority score, the field... Priority rating The calculation formula is:

[0115] ;

[0116] Indicates privacy-priority parameters; This indicates a penalty item for the identity field; The definition is: if the field If it belongs to the identity category field, the value of this item is 1; otherwise, it is 0.

[0117] In the above formula, the first term is the convex penalty form of the privacy score, preventing highly sensitive fields from being underestimated; the second term is the model value; and the third term is the structure regularization. At that time, an encryption bias is applied to identity-related fields.

[0118] As a preferred option .

[0119] As a preferred option, It is 0.15.

[0120] Furthermore, the set of fields to which a field belongs, based on priority scoring, includes:

[0121] Set the first threshold Second threshold ;

[0122] For fields ,like If it is greater than or equal to the first threshold, then... Store a set of fields with a high level of encryption. ;

[0123] like If it is greater than or equal to the second threshold and less than the first threshold, then... The set of fields with the encryption level is stored. ;

[0124] like If it is less than the second threshold, then Store a set of fields with a low encryption level. .

[0125] In this invention, the first threshold Second threshold They can be 0.8 and 0.5 respectively.

[0126] Furthermore, each field is encrypted separately based on the field set, including:

[0127] right The fields in the code are encrypted field by field using a homomorphic encryption library (such as SEAL / BFV);

[0128] right The fields in the code are encrypted by adding Laplace noise;

[0129] right The fields in the data are encrypted using either pseudo-anonymization (hash mapping, numerical bucketing, etc.) or plaintext processing.

[0130] Furthermore, obtain the set of encrypted fields, including:

[0131] They are respectively for , and Encrypt the fields in the data to obtain a collection. , and ;

[0132] based on , and Retrieve the set of encrypted fields, including:

[0133] ;

[0134] This represents a set of encrypted fields.

[0135] Should Prepare data for the input of each branch model in the next step, and retain the field-policy mapping relationship in the form of labels for later use.

[0136] Because the three types of sets mentioned above differ in encoding methods, fidelity, and computational constraints, directly concatenating them for modeling will lead to training failures, information leakage, or performance degradation. Therefore, this step proposes a branch-aware neural network structure and combines it with a dynamic feature fusion mechanism for privacy risk awareness to complete the risk embedding representation. The construction of this model provides a unified input foundation for subsequent graph structure analysis and behavior recognition.

[0137] Furthermore, step 2 includes:

[0138] Depend on The elements in form a vector , It can only perform linear modeling using addition and multiplication, and cannot perform nonlinear transformations such as ReLU and softmax.

[0139] Depend on The elements in form a vector , It contains noisy disturbances, and there are value offsets, but the structure is usable;

[0140] Depend on The elements in form a vector , Strong modeling capabilities but carries the risk of sensitive information leakage;

[0141] , and They represent , and The dimension;

[0142] against Set sub-model structure :

[0143] The model structure uses weighted summation plus a gating function for activation, avoiding any nonlinear operations, ensuring that it adapts to the computational restrictions of homomorphic encryption, which only allows addition and multiplication. At the same time, the gating mechanism controls feature sparsity, thereby improving representation capabilities.

[0144] To avoid the computational complexity and gradient propagation difficulties of homomorphic encryption, a linear weighted structure combined with gated sparse activation is adopted.

[0145] ;

[0146] , This is the sparse trainable weight matrix under the HE path, initially set as an identity matrix; the HE path refers to the path the model takes when processing highly sensitive features processed by homomorphic encryption. Input consisting of ) The dedicated computing channel or branch used at that time, all operations under this path must meet the computing restrictions of homomorphic encryption, and nonlinear functions or complex operations that cannot be encrypted or decrypted are not allowed;

[0147] For example, through a gated function. Suppress excessively small weights;

[0148] Use sigmoid or square activation to satisfy HE computational constraints;

[0149] To select a polynomial approximation form of the square activation function or the sigmoid function to satisfy the HE computation constraint, the HE computation constraint means that when performing model computation in a homomorphic encryption environment, in order to ensure that all computations can be executed in the ciphertext state, only basic arithmetic operations supported by the homomorphic encryption scheme are allowed in the model structure, usually addition and multiplication.

[0150] This indicates element-wise multiplication, ensuring that the computation is supported by HE;

[0151] This structural design automatically adjusts the sparsity of homomorphic data, thereby improving the model's expressive power.

[0152] This is the bias term for the branch linear transformation under the HE path;

[0153] h represents the output feature dimension of the HE branch model;

[0154] The intermediate output vector under the HE path represents the vector based on... Risk representation of features;

[0155] against Set sub-model structure :

[0156] The model uses a simple one-layer neural network structure with a ReLU activation function and a specially designed training regularization to resist differential privacy noise, so that the model can still extract stable and effective feature information even in the presence of perturbations.

[0157] Considering the presence of noise in features under differential privacy, a local noise robustness regularization term is introduced to enhance training:

[0158] Perturbation resampling is added during the training phase to construct local consistency constraints. :

[0159] ;

[0160] in , ;

[0161] n1 is the number of samples used to calculate the local consistency loss;

[0162] and These represent the model's inputs to the original differential privacy settings. The output representation and the model's differential privacy input after perturbation;

[0163] This represents a noise vector sampled from a Laplace distribution centered at 0 with a scale parameter of b.

[0164] This is used to combat uncontrollable noise interference in DP and ensure that the model output is stable under disturbances.

[0165] against Set sub-model structure :

[0166] The model structure adopts a multilayer perceptron structure, which contains three fully connected neural networks. Each layer includes a linear transformation, a nonlinear activation function (ReLU), and a batch normalization operation. This structure can fully model the complex nonlinear relationships in low-sensitivity plaintext features and has strong expressive power. It is suitable for feature processing scenarios with no encryption or slight pseudo-anonymity.

[0167] Introducing "information content balance constraints" to prevent low-sensitivity data from "suppressing" high-sensitivity paths at the information level:

[0168] Use KL distance constraints to keep its output distribution interpretable:

[0169] ;

[0170] in It is a set of prior distributions, which can be set as uniform or business risk preference priors;

[0171] This represents the Kullback-Leibler divergence, an asymmetric distance metric used to measure the difference between two probability distributions. In this scenario, it measures the degree of difference between the model's predicted distribution and the prior distribution; the greater the difference, the stronger the penalty.

[0172] This indicates that it is specifically designed for processing low-sensitivity pseudo-anonymized feature vectors. The sub-model structure takes plaintext features that have been desensitized or lightly masked as input and outputs a vector representation. It can be converted into a probability distribution form by the softmax function for subsequent KL distance regularization constraints. This model has complete expressive power and is usually composed of multi-layer fully connected neural networks. It is suitable for fine-grained modeling of unencrypted and low-sensitivity features.

[0173] Indicates the use of constraints The regularization term for the difference between the model output distribution and the prior distribution π;

[0174] Will , and They are respectively represented as , and ;

[0175] based on , and The risk embedding representation z is computed, including:

[0176] Use a prior risk perception network (which can be a set of trainable parameters) to output a risk weight matrix. ,satisfy ;

[0177] ;

[0178] It can be adjusted based on the type of task (such as whether anti-money laundering behavior prediction is performed);

[0179] This fusion method can activate encrypted paths on demand, increase the contribution of high-sensitivity paths, and solve the drawbacks of plaintext paths dominating in traditional fusion.

[0180] Furthermore, step 3 includes:

[0181] Construct the original graph G:

[0182] Let V represent the set of nodes. Each node This represents a financial account;

[0183] Let Z represent the set of node features. ;

[0184] The elements in Z are the risk embedding representations calculated in step 2;

[0185] Construct a set E of edges based on business logs;

[0186] The original graph G is then represented as ;

[0187] Perturb the original graph G:

[0188] The node ID of each node in the original graph G is perturbed to obtain the set of perturbed nodes. ;

[0189] Perturb each edge in the original graph G to obtain the set of perturbed edges. ;

[0190] For Z, either keep the node features in Z unchanged or perform dimensionality reduction on the features in Z to obtain the set. ;

[0191] based on , and Constructing a perturbation graph Disturbance diagram Represented as:

[0192] .

[0193] Furthermore, the set E of edges constructed based on business logs includes:

[0194] The edges in the original graph are constructed according to the following three behaviors:

[0195] Transfer transaction (primary): If account Towards Initiate a transfer and record the transaction. , and includes attributes such as amount;

[0196] Shared login device (auxiliary): If and Log in to the same device within 24 hours, and also build ;

[0197] Network co-domain behavior (times): If and Multiple appearances in the same IP network segment, building .

[0198] Set boundary weight This is the normalized value of the transaction amount (if it is not a transfer, the corresponding edge weight is a constant of 0.1).

[0199] Furthermore, the node ID of each node in the original graph G is perturbed to obtain a set of perturbed nodes. ,include:

[0200] Each node The real ID is perturbed by the following function as follows: :

[0201] ;

[0202] in To ensure that the same ID is not duplicated at different times, the system key is changed daily.

[0203] Output The perturbed ID is 64 bytes long and is used for subsequent graph storage.

[0204] It is actually a 256-bit binary hash value, represented in the system as 64 hexadecimal characters.

[0205] Furthermore, each edge in the original graph G is perturbed to obtain a set of perturbed edges. ,include:

[0206] Perturbation method 1: Randomly add edges

[0207] Each real edge With probability Replace with , This ensures that pseudo-edges do not form paths in the original graph. For nodes and The edge between;

[0208] recommend It can be set according to the organization's compliance level;

[0209] The replacement method is: randomly select the target node. Maintaining a constant number of edges enhances structural fuzziness.

[0210] Disturbance Method 2: Edge Weight Noise Injection

[0211] Border rights Add Gaussian noise:

[0212] ;

[0213] recommend Applicable to amount normalization The scenario of intervals;

[0214] Avoid exposing sensitive edges (large amounts, high-frequency transactions) after disturbances.

[0215] This indicates that the edge weights follow a Gaussian distribution (i.e., a normal distribution) with a mean of 0 and a variance of σ1 squared; and this represents random noise applied to the edge weights.

[0216] Indicates adding to the edge The Gaussian noise term is used to simulate unpredictable disturbances, which masks the true weight of the edges, thus preventing attackers from identifying sensitive transaction patterns by the size of the edge weights.

[0217] Two perturbation methods can be used simultaneously, or only one perturbation method can be used.

[0218] Furthermore, step 4 includes:

[0219] For each node ,extract Timestamps of all adjacent edges (outgoing or incoming) Construct transaction time series , k1 is the total number of timestamps in the time series;

[0220] in It is the transaction time after the disturbance, which comes from the edge attribute in step 3; the timestamp can be normalized to the relative time within a day or hour to avoid the exposure of the absolute time.

[0221] based on Construct each node Time difference series :

[0222] ;

[0223] This sequence reflects the dynamic behavioral patterns of the account, such as transaction frequency, intensity, and active periods.

[0224] Using the Temporal Difference Convolutional Network (TDCN) Calculate and obtain temporal behavior feature representation :

[0225] ;

[0226] One-dimensional convolution kernel (e.g., large convolution kernel size) ); This represents a one-dimensional convolution operation; This is a bias term.

[0227] The dimensions can be set to 16~64 for subsequent risk fusion.

[0228] Reasons for choosing convolutional structures:

[0229] Convolution has the ability to capture local patterns, making it suitable for identifying unusual short-term, dense transactions.

[0230] In perturbation-time environments, convolution is more robust than RNN and can avoid amplification of propagation errors;

[0231] It supports batch processing, facilitating deployment and parallel processing.

[0232] Furthermore, risk scores are calculated based on risk embedding representations and temporal behavioral feature representations, including:

[0233] Constructing a joint risk input vector :

[0234] ;

[0235] Risk embedding representation, used to represent nodes Static risk characteristics; This is a temporal behavior feature representation used to represent nodes. Dynamic behavioral characteristics; Indicates a splicing operation; ;

[0236] and They represent and The dimension;

[0237] based on Calculate risk score .

[0238] All features are in encrypted form or in an intermediate representation to ensure that no plaintext information is leaked.

[0239] Furthermore, risk scoring The calculation formula is:

[0240] ;

[0241] Variable description:

[0242] , for The linear transformation matrix;

[0243] , Dynamic behavior vector A linear mapping;

[0244] The time-series bias adjustment factor for individual accounts is calculated by the auxiliary network (gating function).

[0245] The global time-series behavior mean (of all accounts) (mean), which plays a centralizing role;

[0246] This is the bias term used when calculating the final risk score;

[0247] It is the sigmoid activation function;

[0248] The final risk score indicates that the higher the score, the greater the risk.

[0249] The innovation lies in the introduction of an individual difference adjustment term, which can dynamically capture the degree to which individual behavior deviates from the group mean, and is suitable for detecting "individual deviation-type abnormal behavior" (such as sudden activity or atypical transactions); compared with fixed weighted fusion, this design is more flexible and is particularly suitable for "non-steady-state feature modeling" in risk control; all calculations can be completed in the intermediate state without relying on plaintext or label extrapolation.

[0250] Loss function structure optimization (emphasizing model robustness under encryption):

[0251] To ensure compatibility with training environments under encrypted / perturbed data conditions, a risk score training objective function with a predictive conservatism regularization term is designed:

[0252] ;

[0253] Variable description:

[0254] This represents the standard binary cross-entropy loss;

[0255] This indicates a risk label;

[0256] The second term is the smoothing regularization term of the predicted distribution, which prevents the model output from being concentrated in an excessively extreme manner by minimizing the KL divergence with the mean distribution (common under encrypted features).

[0257] 3 represents the strength of the regularization term (e.g.) );

[0258] This represents the predicted mean of the current batch.

[0259] In encrypted training, due to noise or incomplete representation of features, the model is prone to outputting extreme values ​​(too high or too low). The KL term designed in this invention makes the model prediction more "neutral and conservative", which is a very important stability optimization under encrypted conditions.

[0260] Furthermore, risk control strategies are implemented based on risk scores, including:

[0261] like If so, high-risk risk control strategies will be implemented, including freezing accounts, immediately locking transactions, and generating compliance reports;

[0262] like If so, medium-risk risk control strategies will be implemented, including multi-factor verification, restriction of certain behaviors, and suspension of transactions pending review.

[0263] like If so, a low-risk risk control strategy will be implemented, including normal release and logging only.

[0264] and These are the first risk scoring threshold (e.g., 0.9) and the second risk scoring threshold (e.g., 0.7), respectively.

[0265] Each strategy corresponds to a configurable execution template, which is set by business personnel on the risk control platform.

[0266] For example:

[0267] Assuming account :

[0268] Static vector (64 dimensions);

[0269] Dynamic vector (32 dimensions);

[0270] Obtain the concatenated vector ;

[0271] Model output ;

[0272] satisfy Execute automatically:

[0273] Lock account;

[0274] Send API calls to push emails / SMS messages;

[0275] The risk control console records detailed input vectors and For post-hoc analysis by humans.

[0276] This invention addresses the systemic contradiction in existing technologies regarding sensitive information protection and complex behavior analysis, proposing a holistic solution driven by scenarios, centered on structural decoupling, and aimed at behavioral robustness. First, based on the dual assessment results of the business sensitivity and risk value of input features, this invention establishes a fine-grained feature-layered encryption mechanism to control data privacy protection strategies at the source. Through differentiated model structure design, it ensures that various encrypted features can participate in effective modeling under suitable paths. Second, to address the inability of traditional models to handle chain relationships and graph structure leakage, this invention constructs a graph representation mechanism with structural perturbation protection capabilities, enabling the risk control system to identify abnormal interaction patterns between potential accounts without exposing the real network structure. Furthermore, to address the problem of attackers evading detection through transaction time-series perturbations or behavioral staggering, this invention introduces a time-dimensional behavioral robustness analysis strategy, supporting non-explicit modeling of transaction time series under encrypted conditions, thereby achieving continuous tracking and risk scoring of dynamic risk behavior. While maintaining a high level of privacy protection, the overall solution effectively enhances the risk control model's ability to detect complex attack behaviors, making it particularly suitable for financial risk control scenarios involving multi-party collaboration, cross-domain modeling, and high regulatory compliance requirements.

[0277] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for encrypted transmission and analysis of sensitive information in financial risk control scenarios, characterized in that, include: Step 1: Divide the original fields into a field set, and encrypt each field based on the field set to obtain the encrypted field set; Step 2: Calculate the risk embedding representation based on the set of encrypted fields; Step 3: Construct the original graph based on the risk embedding representation, and perturb the original graph to obtain the perturbed graph; Step 4: Extract temporal behavior feature representations in the time dimension based on the perturbation graph; Step 5: Calculate the risk score based on the risk embedding representation and the time behavior feature representation, and execute the risk control strategy based on the risk score; This involves dividing the original fields into a field set, including: Obtain the privacy sensitivity and modeling contribution of each field separately; Priority scores are calculated based on privacy sensitivity and modeling contribution. Determine the set of fields to which a field belongs based on priority scores; Step 3 includes: Construct the original graph G: Let V represent the set of nodes, and each node... This represents a financial account; Let Z represent the set of node features; A set E of edges is constructed based on business logs; The original graph G is then represented as ; Perturb the original graph G: The node ID of each node in the original graph G is perturbed to obtain the set of perturbed nodes. ; Perturb each edge in the original graph G to obtain the set of perturbed edges. ; For Z, either keep the node features in Z unchanged or perform dimensionality reduction on the features in Z to obtain the set. ; based on , and Constructing a perturbation graph ; Step 4 includes: For each node ,extract Timestamps of all adjacent edges Construct transaction time series ; based on Construct each node Time difference series ; Using the Temporal Difference Convolutional Network (TDCN) Calculate and obtain temporal behavior feature representation .

2. The method for encrypted transmission and analysis of sensitive information in a financial risk control scenario according to claim 1, characterized in that, Priority scores are calculated based on privacy sensitivity and modeling contribution, including: Use respectively and Representation field Privacy sensitivity and modeling contribution; based on and Calculate priority scores.

3. The method for encrypted transmission and analysis of sensitive information in a financial risk control scenario according to claim 2, characterized in that, The set of fields to which a field belongs, determined by priority scoring, includes: Set the first threshold Second threshold ; For fields ,like If it is greater than or equal to the first threshold, then... Store a set of fields with a high level of encryption. ; like If it is greater than or equal to the second threshold and less than the first threshold, then... The set of fields with the encryption level stored ; like If it is less than the second threshold, then Store a set of fields with a low encryption level. ; Fields Priority rating The calculation formula is: ； Indicates privacy-priority parameters; This indicates a penalty item for the identity field; The definition is: if the field If it belongs to the identity category field, the value of this item is 1; otherwise, it is 0.

4. The method for encrypted transmission and analysis of sensitive information in a financial risk control scenario according to claim 3, characterized in that, Encrypt each field separately based on the field set, including: right The fields in the code are encrypted field by field using a homomorphic encryption library; right The fields in the code are encrypted by adding Laplace noise; right The fields in the data are encrypted using either pseudo-anonymization or plaintext processing.

5. The method for encrypted transmission and analysis of sensitive information in a financial risk control scenario according to claim 4, characterized in that, Retrieve the set of encrypted fields, including: They are respectively for , and Encrypt the fields in the data to obtain a collection. , and ; based on , and Get the collection of encrypted fields.

6. The method for encrypted transmission and analysis of sensitive information in a financial risk control scenario according to claim 5, characterized in that, Step 2 includes: Depend on The elements in form a vector , ; Depend on The elements in form a vector , ; Depend on The elements in form a vector , ; , and They represent , and The dimension; against Set the first sub-model structure The first sub-model structure uses weighted summation plus a gating function for activation, while controlling feature sparsity through a gating mechanism; against Set the second sub-model structure The second sub-model structure includes a one-layer neural network structure and the ReLU activation function. against Set the third sub-model structure The third sub-model structure adopts a multilayer perceptron structure, which contains three fully connected neural networks. Each layer includes linear transformation, nonlinear activation function and batch normalization operation. Will , and They are respectively represented as , and ; based on , and Compute the risk embedding representation z.

7. The method for encrypted transmission and analysis of sensitive information in a financial risk control scenario according to claim 1, characterized in that, Risk scores are calculated based on risk embedding representations and temporal behavioral feature representations, including: Constructing a joint risk input vector ; based on Calculate risk score .

8. The method for encrypted transmission and analysis of sensitive information in a financial risk control scenario according to claim 7, characterized in that, Risk control strategies are implemented based on risk scores, including: like If so, a high-risk risk control strategy will be implemented; like If so, then implement a medium-risk risk control strategy; like If so, a low-risk risk control strategy will be implemented; and These are the first risk scoring threshold and the second risk scoring threshold, respectively.

Images