Dynamic risk control mobile payment method and system based on reinforcement learning

Abstract

The invention provides a dynamic risk control mobile payment method and system based on reinforcement learning, and the method comprises the steps: obtaining a real-time payment transaction request of a user, so as to construct a transaction state vector, inputting a probability model based on pre-training, and calculating a behavior entropy representing the risk level of the current payment transaction request; inputting to a pre-trained reinforcement learning strategy model, and generating an encryption control parameter; according to the encryption control parameters, encrypting a plurality of fragments by using an anti-quantum encryption algorithm, and distributing the fragments according to a specified sequence through a plurality of communication channels; receiving the encrypted sub-key fragment, generating a restored original payment instruction, and calculating a structural integrity score of the restored original payment instruction after recombination; and performing consistency matching judgment according to the behavior entropy value and the structural integrity score so as to determine whether to execute the payment transaction or not. According to the method, the transmission surface unpredictability is improved through fragment-level disturbance enhancement and channel sequence disturbance, and meanwhile transparent compatibility of code scanning and a near-field interface is kept.

Description

Technical Field

[0001] This invention belongs to the field of dynamic risk control mobile payment, and particularly relates to a dynamic risk control mobile payment method and system based on reinforcement learning. Background Technology

[0002] The existing mobile payment system is often fragmented in its "risk identification—encryption execution—transmission reassembly—transaction approval" chain: the risk control end often evaluates transactions using static rules or fixed models, and the output is not structurally coupled with the encryption layer; the encryption end usually encapsulates the entire instruction at once, lacking segmentation and channel routing control based on risk status; the transmission end sends data in a single channel and in a fixed order, making it easy to infer from traffic analysis and replay; the server-side reassembly and verification are mostly limited to message-level legality checks, making it difficult to quantify the impact of fragment-level distortion or loss on overall security. These problems are amplified in multi-terminal, heterogeneous network, and high-concurrency scenarios, leading to the excessive interception of low-risk transactions and the possibility that high-risk transactions may pass through under the guise of "structural integrity." At the same time, classical public-key cryptography exposes long-term security vulnerabilities under quantum threats. Although some work has attempted to use quantum-resistant algorithms for payment data encryption, it mostly requires high-computing-power environments and treats quantum-resistant encryption as an independent black box, failing to form a verifiable closed loop with behavioral risks, transmission paths, and server-side reassembly strategies.

[0003] Furthermore, many solutions introduce a large number of historical or sensitive features to improve detection accuracy, which increases the pressure of data collection and compliance, as well as the complexity of deployment, and conflicts with the constraints of mobile terminal computing power, power consumption and latency.

[0004] The industry lacks an engineering solution that tightly binds lightweight risk measurement with quantum-resistant structured encryption, channel-level disturbances and quantifiable recombinable integrity without modifying existing QR code scanning and near-field interfaces or adding new dedicated hardware. This solution must satisfy both the low coupling and auditability of each link in the link, and maintain a dynamic balance between availability and security under real-world network conditions such as packet loss, out-of-order delivery and asynchronous arrival. Summary of the Invention

[0005] The purpose of this invention is to propose a mobile payment method and system for dynamic risk control based on reinforcement learning, thereby solving the aforementioned problems.

[0006] To achieve the above objectives, a first aspect of the present invention provides a mobile payment method for dynamic risk control based on reinforcement learning, the method comprising the following steps: The system acquires real-time payment transaction requests from users, extracts and preprocesses the real-time features of these requests to construct a transaction state vector, and inputs the transaction state vector into a pre-trained probability model to calculate the behavioral entropy value that represents the risk level of the current payment transaction request. The transaction state vector and the behavior entropy value are input into a pre-trained reinforcement learning policy model to generate encryption control parameters; wherein, the encryption control parameters include the number of payment instruction segments, the highest strength perturbation parameter, the QR code scanning-NFC-QR code scanning path allocation, and the segment distribution order; According to the encryption control parameters, the payment instruction is fragmented to generate several fragments. The quantum-resistant encryption algorithm is used to encrypt the fragments to generate encryption subkey fragments, which are then distributed in a specified order through multiple communication channels. The server receives the encrypted subkey fragment, decrypts and reassembles it to generate a restored original payment instruction, and calculates the structural integrity score of the reassembled restored original payment instruction. Consistency matching is performed based on the behavioral entropy value and the structural integrity score to determine whether to execute the current payment transaction request.

[0007] Furthermore, the real-time features include transaction timestamp, transaction amount, device identifier, and payment method code; The preprocessing specifically includes: The transaction timestamp is determined to have a periodic characteristic within a day; the transaction amount is logarithmically compressed; and the device identifier is converted using a hash mapping. Range values; normalization of the payment method encoding.

[0008] Furthermore, the pre-trained probabilistic model is a Gaussian mixture model, with the input being a transaction state vector and the output being a behavioral entropy value, used to assess the risk level of the current state.

[0009] Further, the step of inputting the transaction state vector and the behavior entropy value into a pre-trained reinforcement learning policy model to generate encryption control parameters specifically involves: The transaction state vector and the behavior entropy value are concatenated to form a joint input feature;

[0010] Based on the joint input features, a resource cost regularization term and a structural risk loss term based on policy parameters are introduced to calculate encryption control parameters; wherein, the resource cost regularization term is used to limit the perturbation complexity of the policy and the consumption of the communication channel; the structural risk loss term is used to measure the degree of adaptation of the encryption policy to the current behavior entropy value. The strategy parameters include the number of subkeys, the disturbance strength level, the communication path sequence, and the key transmission order.

[0011] Furthermore, the structural risk loss term is calculated by comprehensively calculating the total encryption complexity of the number of key splits, the perturbation level, and the path perturbation degree, combined with the behavioral entropy value; The resource cost regularization term is one or more of the following: the encryption computation quantity increases linearly with the number of subkeys, the communication path switching cost, and the disturbance level is related to the encryption and decryption overhead.

[0012] Furthermore, the step of segmenting the payment instruction according to the encryption control parameters to generate several segments, encrypting these segments using a quantum-resistant encryption algorithm to generate encryption subkey fragments, and distributing them in a specified order through multiple communication channels specifically includes: Obtain the number of segments and the payment instruction itself. Logically divide the payment instruction into segments equal to the number of segments, as sub-segments. If the payment instruction cannot be evenly divided, add a padding identifier to the corresponding segment, and the server restores it using offset bits. For each sub-segment, combining the perturbation strength and the current sub-segment risk perception factor, quantum-resistant encryption is performed to generate encrypted subkey fragments; Based on the encrypted subkey fragments, and according to the communication path selection vector and transmission order control vector, all encrypted subkey fragments are subjected to channel scheduling and transmission tasks. A fragment-level path perturbation regularization term is introduced to measure the perturbation strength of encrypted fragments in both spatial path and temporal order dimensions, and serves as an evaluation function for the current allocation strategy. A scheduling record table is constructed by combining the communication path selection vector, the transmission order control vector, and the encryption subkey fragment.

[0013] Furthermore, the quantum-resistant encryption uses a standard ring-based error encryption function; The channels include a QR code scanning channel and an NFC channel, which are marked as 0 and 1 respectively; The channel scheduler calls different interfaces according to the communication path selection vector and sends the corresponding encrypted subkey fragments in the order set by the transmission order control vector.

[0014] Further, the server receives the encrypted subkey fragment, decrypts and reassembles it to generate a restored original payment instruction, and calculates the structural integrity score of the reassembled restored original payment instruction, specifically including: The server sorts the scheduling record table in ascending order according to the segment sending order in the communication path selection vector, and generates a decryption queue. The standard deviation value is obtained by looking up the table based on the highest intensity perturbation parameter, the sorted ciphertext fragment is generated, and the RLWE decryptor is initialized. After all the encrypted subkey fragments are decrypted, the server concatenates the successfully decrypted fragments in order to obtain the payment restoration instruction; Based on the encrypted subkey fragments, a structural integrity scoring function is designed to calculate the structural integrity score of each encrypted subkey fragment, which is used to measure the usability of the entire key chain after decryption and reconstruction. If the structural integrity score is lower than the preset threshold, the transaction will be terminated directly and the anomaly will be recorded to avoid misunderstanding of secrets or fraud involving partial information.

[0015] Furthermore, the step of performing a consistency matching judgment based on the behavioral entropy value and the structural integrity score to determine whether to execute the payment transaction specifically includes: A minimum tolerable structural integrity threshold is set for each of the aforementioned behavioral entropy values. If the current structural integrity score is less than the minimum structural integrity threshold, it indicates that the current transaction structural strength does not meet the security strength required by its behavioral risk level, and the transaction must be terminated.

[0016] A second aspect of the invention provides a mobile payment system for dynamic risk control based on reinforcement learning, the system comprising: The transaction request acquisition unit is used to acquire the user's real-time payment transaction request, extract the real-time features of the payment transaction request and preprocess them to construct a transaction state vector, and input the transaction state vector into a pre-trained probability model to calculate the behavioral entropy value representing the risk level of the current payment transaction request. The request analysis unit is used to input the transaction state vector and the behavior entropy value into a pre-trained reinforcement learning policy model to generate encryption control parameters; wherein, the encryption control parameters include the number of payment instruction segments, the highest strength perturbation parameter, the QR code scanning-NFC-QR code scanning path allocation, and the segment distribution order; A quantum-resistant encryption unit is used to segment the payment instruction according to the encryption control parameters, generate several segments, encrypt the several segments using a quantum-resistant encryption algorithm to generate encryption subkey fragments, and distribute them in a specified order through multiple communication channels. The request decryption unit is used to indicate that the server receives the encrypted subkey fragment, decrypts and reassembles it to generate a restored original payment instruction, and calculates the structural integrity score of the reassembled restored original payment instruction; The transaction execution unit is used to perform consistency matching judgment based on the behavior entropy value and the structural integrity score to determine whether to execute the current payment transaction request.

[0017] The beneficial technical effects of the present invention are at least as follows: This invention proposes an end-to-end dynamic risk control and quantum-resistant structured encryption execution path. Using lightweight state vectors and behavioral entropy as the sole risk input, it drives a pre-trained strategy to generate a directly deployable policy package and instruction payload handle. This decouples risk perception and encryption execution into an auditable chain process without additional historical or privacy data collection. Subsequently, according to the fragmentation ratio, perturbation level, channel allocation, and transmission order given in the policy package, fragment-level encryption based on a quantum-resistant system is implemented on the payment instructions. Perturbation enhancements tied to risk levels are introduced at the fragment level. The required order and decryption parameters are uniformly encapsulated in a ciphertext packet using metadata, enabling the server to complete out-of-order reordering and consistent configuration decryption without relying on side channels. The server performs fragment-level decryption and instruction reassembly based on the metadata inside the ciphertext packet and outputs a structural integrity score. This score is then used to establish a minimum structural requirement function at the execution layer that corresponds one-to-one with the discrete tags of the upstream risk level. This achieves a bidirectional consistency judgment that "high risk requires high integrity, and high integrity can relax risk." Ultimately, only the combination of this score and the risk level is used as the release switch, forming an engineering judgment outside the traceable, parameterizable, and compliant model. The key to this invention lies in using the strategy packet and ciphertext packet as the central objects to strictly converge the input and output of each link to the product of the previous step, avoiding cross-step dependencies and variable ambiguity; enhancing the unpredictability of the transmission plane with fragment-level perturbation and channel sequence perturbation, while maintaining transparent compatibility with barcode scanning and near-field interfaces; and quantifying the reassembly quality with structural integrity scoring, no longer relying on "successful decryption equals security", thereby achieving a low-modification, strong closed-loop, and easily auditable anti-quantum dynamic payment security control system under the constraints of mobile computing power and latency. Attached Figure Description

[0018] The present invention will be further described with reference to the accompanying drawings, but the embodiments in the drawings do not constitute any limitation on the present invention. For those skilled in the art, other drawings can be obtained based on the following drawings without creative effort.

[0019] Figure 1 This is a flowchart of the mobile payment method for dynamic risk control based on reinforcement learning according to the present invention. Detailed Implementation

[0020] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0021] like Figure 1 As shown in the figure, the mobile payment method for dynamic risk control based on reinforcement learning provided in this embodiment of the invention includes: S1. Obtain the user's real-time payment transaction request, extract the real-time features of the payment transaction request and preprocess them to construct a transaction state vector, and input the transaction state vector into a pre-trained probability model to calculate the behavioral entropy value representing the risk level of the current payment transaction request.

[0022] Specifically, this step extracts structured state information from a payment transaction request and generates a risk quantification indicator based on this state information, which can be used to control subsequent encryption strategies. This risk indicator should be real-time, deployable, reproducible, and independent of long-term user profiles. To this end, we adopt a lightweight state construction scheme, abstracting each transaction into a standardized state vector and combining it with a trained probability model to calculate the risk entropy value of the current state, which serves as input to the downstream strategy decision-making module.

[0023] The specific input information is as follows: Transaction timestamp The data is obtained through the terminal device's operating system and recorded in milliseconds or seconds, usually through API calls such as System.currentTimeMillis() (Android) or NSDate() (iOS).

[0024] Transaction amount The input is provided by the transaction initiator (usually a user-side app or a QR code scanning device), and the system defaults to two decimal places.

[0025] Device Identifier It is obtained through a device-level registration mechanism, such as the IMEI, UUID of the terminal device, or a virtual device number assigned by the payment platform. It is generated by the payment client app during installation and activation and bound to the background.

[0026] Payment method code The system gateway module determines whether the current transaction was initiated through channels such as QR code scanning, NFC, or Bluetooth, which is usually automatically injected by the system during interface calls.

[0027] For example, a transaction occurred on August 14, 2025 at 13:36:08, completed by scanning a QR code with an Android phone, amounting to ¥78.50, with the terminal device ID "XYZ123456". The system collected and generated the following input: , , , ; Furthermore, to ensure the comparability and tractability of various input variables in the computational structure, it is necessary to first standardize the input variables and construct a transaction state vector. The format is as follows:

[0028] in: This transforms timestamps into periodic features within a day. Represents the total number of seconds in a day, usually set to . Logarithmic compression of transaction amounts reduces the impact of large transactions and avoids data skewness. The device identifier is converted through a hash mapping. Range values, irreversible, ensuring privacy. Normalize the payment method numbering. This indicates the maximum supported payment method number (such as QR code, NFC, Bluetooth, etc.) to stabilize the input space.

[0029] Furthermore, the completed transaction state vector is constructed. The data is then fed into a Gaussian Mixture Model (GMM). We use a pre-trained GMM to evaluate the risk level of the current state and define a behavioral entropy value. As an estimate of the "abnormality" of behavior, the calculation method is as follows:

[0030] in: The number of components in a mixed distribution is typically 1. ; It is the first The weights of the Gaussian components are obtained through offline training using the EM algorithm; It is the first The mean vector of each component; It is the first The covariance matrix of each component; express Density function values ​​under this distribution; all parameters The density is fixed during the deployment phase, not trained locally, and used only for inference; if the density value is too low, a minimum value can be set to truncate it to avoid numerical instability.

[0031] Numerical results of this formula This can serve as a quantitative representation of whether the system considers the transaction's behavior abnormal. A higher value indicates a lower fit to known trading patterns, suggesting greater potential risk. Unlike traditional rule-based systems, this entropy value does not rely on explicit judgment logic but is derived naturally from the distribution of trading behavior, and all features originate from the transaction itself, without involving historical behavioral trajectories.

[0032] S2. Input the transaction state vector and the behavior entropy value into a pre-trained reinforcement learning policy model to generate encryption control parameters; wherein, the encryption control parameters include the number of payment instruction segments, the highest strength perturbation parameter, the QR code scanning-NFC-QR code scanning path allocation, and the segment distribution order.

[0033] Specifically, this step plays the role of the "decision center" in the overall structure of the invention, and its core task is to transform the state vector calculated in step one. and risk entropy The mapping is translated into an operable set of structured cryptographic control parameters that determine how payment instructions are encrypted and distributed. Unlike traditional static rule systems, this step constructs a lightweight policy decision module that can be deployed on terminals or edge servers, possessing transaction personalization and risk response adaptive capabilities. This is a prerequisite for realizing the dynamic quantum-resistant structure encryption in this invention.

[0034] The input includes the two variables output from step one: The state vector of the current transaction is: 3D normalized real vector; This represents the risk entropy value of the current transaction under the behavioral distribution model. To improve the interpretability and stability of the strategy output, this step will... and Concatenated into joint input features And introduce a set of strategy parameters , respectively, represent the number of subkeys, the level of disturbance intensity, the communication path sequence, and the order in which keys are sent.

[0035] Furthermore, traditional reinforcement learning methods mostly consider only a single objective, such as maximizing reward or minimizing loss. However, in the scenario of this invention, policy generation must simultaneously consider the consistency of encryption strength, key path perturbation, computational cost, and risk level. Therefore, we propose a risk-guided structure-controlled policy generation function, which introduces a structural risk matching term and a perturbation complexity regularization term in a target-weighted manner, as follows:

[0036] in: For the strategy model, Indicates model weights; It is the output parameter set; This is a structural risk loss term used to measure the degree to which the encryption strategy adapts to the current behavioral entropy value. This is a resource cost regularization term used to limit the perturbation complexity of the strategy and the consumption of communication channels; and This refers to the task importance weight, which can be adjusted based on the terminal's capabilities during deployment; all actions All are enumerated or sampled in a finite discrete space.

[0037] Furthermore, in order to better reflect the coupling logic between risk levels and encryption structures in this invention, The construction is not only for Instead of a weighted response, it introduces "the mismatch between structural disturbance capability and risk level" as the core of the loss, namely:

[0038] in: The policy strength estimation function comprehensively calculates the total encryption complexity by considering the number of key splits, perturbation level, and path perturbation degree. This function can be deployed as a lookup table structure or a lightweight regression network to ensure deployment efficiency. The loss term is... Using the target benchmark, if the actual strategy's encryption strength is insufficient or excessive, a bias penalty will be introduced.

[0039] Meanwhile, to limit the degradation of system performance under high-risk conditions, a resource consumption regularization term is introduced. For example, it can be defined as: (The computational complexity of encryption increases linearly with the number of subkeys.) (Communication path switching cost); (The level of disturbance is related to the encryption / decryption overhead). These cost functions All of these are preset hardware capability sensitivity coefficients, set based on device stress test results before deployment.

[0040] Before deployment, the strategy model is trained in a simulated environment using a reinforcement learning framework (such as Proximal Strategy Optimization, PPO). The training data consists of synthetic trading records and real risk labels. After training, the parameters are fixed. Deployed on the terminal APP or gateway side, it generates a set of control parameters through a single forward inference process. It requires no online fine-tuning and does not involve the learning process of personal data.

[0041] For example, when the model input is: ,

[0042] Policy model output: : Payment instruction split segment, indicating that the payment instruction will be split into 3 segments; : Indicates the use of the highest intensity perturbation parameters; : Scan-NFC-Scan path allocation; The paragraph distribution order refers to the order in which the second paragraph is published first, the first paragraph is published later, and the third paragraph is published last.

[0043] The results indicate that the model identifies the current transaction behavior as having a high risk. Therefore, the strategy uses a combination of high-perturbation encryption, high path obfuscation, and non-sequential transmission to fully demonstrate the system's structural adaptability under risk conditions.

[0044] S3. According to the encryption control parameters, the payment instruction is fragmented to generate several fragments. The quantum-resistant encryption algorithm is used to encrypt the several fragments to generate encryption subkey fragments, which are then distributed in a specified order through multiple communication channels.

[0045] Specifically, this step is a crucial link in the overall invention structure, transitioning from "strategy generation" to "quantum-resistant encryption behavior implementation." Its core objective is to utilize the structured strategy parameters output in step two. Payment instructions By performing structured partitioning, parameterized perturbation encryption, and multi-channel scheduling and distribution, a key fragmentation structure with transport layer perturbation resistance is formed under quantum security. Its goal is not merely to achieve fragmentation encryption, but to incorporate transaction risk states. and structural disturbance control parameters , We will design a mechanism that can be used to disrupt attackers' decryption models, path recognition models, and fragment reconstruction models, thereby truly realizing a quantum-resistant secure transmission chain in mobile payment scenarios.

[0046] First, the execution module follows Instructions Logical division into Joke Segmentation ,like If it cannot be divided equally, then on the... Padding markers are added to the segment, and the server restores it using offset bits. In this step, to enhance risk controllability, [the following steps are taken]. Participants determine the granularity of the split, i.e., based on... Dynamically adjusting the proportion of each data segment makes it more likely that high-risk transactions will place sensitive fields (such as amount and signature) in separate segments. For example, in At that time, the segmentation was performed using a non-uniform ratio. Instead This splitting scheme uses a pre-defined consistent mapping table between the server and the client, eliminating the need for dynamic synchronization.

[0047] Then, for each sub-shard It performs quantum-resistant encryption. The system is based on the RLWE encryption principle, combined with perturbation strength. And introduce the current sub-sharding risk perception factor. The encryption process is then executed. Depend on In conjunction with the fragment content structure, the perturbation strength is dynamically adjusted to address the non-uniform risk distribution in payment transactions. We innovatively propose the following encryption structure:

[0048] in: For encryption subkey fragments; This is an error-based encryption function on a standard ring, implemented using NTT-optimized transformation during deployment. According to The standard deviation of the disturbance obtained from the table; , which represents the fragment for our proposed encryption perturbation enhancement factor. The relative risk level is obtained by combining the overall risk of the transaction with the data entropy of the segment itself. The randomly generated items are encrypted locally, ensuring that each encryption is unique.

[0049] Here The introduction of this method is one of the key innovations of this step, as it addresses global risk. Compared with local data complexity (e.g.) By combining the compression entropy or hash feature distribution, the encryption perturbation strength can be dynamically adjusted in a personalized manner at the fragment dimension, avoiding the problem that the globally unified perturbation strategy fails under highly sensitive fragments.

[0050] Furthermore, in the encrypted subkey fragment After generation, select a vector based on the communication path. and transmission order control vector The system performs channel scheduling and transmission tasks on all encrypted subkey fragments. Two main communication channels are preset in the system: a barcode scanning channel and an NFC channel, marked as 0 and 1 respectively. The channel scheduling module... Call different interfaces (such as NFCtransmitApdu() in Android systems or postToQRServer() in QR code scanning channels), and strictly follow the instructions. Send in the set order .

[0051] Furthermore, to quantify the effectiveness of the "structural perturbation" in this step, we propose a fragment-level path perturbation regularization term. This is used to measure the perturbation strength of encrypted fragments in both spatial path and temporal order dimensions, and serves as an evaluation function in system policy training. It is defined as follows:

[0052] in: This is an indicator function that is 1 if the condition is true and 0 otherwise; the first term measures whether the communication path between segments has changed; the second term measures whether the sending order has been scrambled (not in accordance with the natural order); coefficients The perturbation dimension weighting factor is set by the system according to the attack modeling strategy.

[0053] This regularization term does not directly participate in the deployment inference process, but it serves as one of the reward terms during the policy network training phase, guiding the model to generate a scheduling output structure with higher "path unpredictability" and enhancing the anti-analysis capability of key fragments at the physical transport layer.

[0054] S4. The server receives the encrypted subkey fragment, decrypts and reassembles it to generate a restored original payment instruction, and calculates the structural integrity score of the reassembled restored original payment instruction.

[0055] Specifically, this step is used on the server side to receive and reconstruct the encrypted subkey fragment sent in the previous stage. And based on the scheduling record table The fragments are reordered, decrypted, and concatenated to ultimately reconstruct the original payment instruction. The process involves assessing the structural integrity of the restructured result. This step does not involve determining whether the transaction should be executed; it only verifies the reliability and integrity of the keychain restructure. The result, along with the behavioral risk level, will be used in the next step to determine whether to execute the payment.

[0056] Furthermore, the server first based on The order in which the fragments are sent, for Sort in ascending order to generate a decryption queue. For example, if The received ciphertext is in the following order: The server will proceed according to... The logical order is restored.

[0057] The system is based on the parameters of the highest intensity disturbance. Find the standard deviation from the table Initialize the RLWE decryptor. Perform the following operations on each ciphertext fragment:

[0058] in: For the sorted number A ciphertext fragment; The RLWE decryptor deployed in the server-side TEE uses the same modulus and polynomial dimensions as the client. If decryption fails due to excessive disturbance, missing data, or misaligned random factors, then Marked as The decryption failure handling process will then begin.

[0059] all After decryption, the server reassembles the successfully decrypted fragments in sequence to obtain the restored payment instruction. If a segment fails, then Empty spaces are reserved and marked for use by subsequent judgment modules.

[0060] Furthermore, to quantify the integrity of the restored structure, a structural integrity scoring function is proposed. This is used to measure the usability of the entire key chain after decryption and reassembly.

[0061] in: For indicator functions, representing fragments Was the decryption successful? The closer the score is to 1, the higher the reconstruction success rate and the better the system structural integrity; this score does not include behavioral entropy values. It focuses on the physical recoverability of the structured execution chain.

[0062] For example: If The system successfully decrypted the fragment as follows: , , ,in ,but The next step will use this score, along with the risk level, to determine whether the transaction can proceed.

[0063] In addition, the system can be configured with a structural integrity score. As a minimum threshold for structural integrity, for example If the reconstruction score falls below this value, the transaction will be terminated immediately and the anomaly recorded to prevent misunderstandings or fraudulent information.

[0064] S5. Based on the behavioral entropy value and the structural integrity score, a consistency matching judgment is performed to determine whether to execute the current payment transaction request.

[0065] Specifically, this step is the final step in the entire invention system used to complete transaction execution control, and its goal is to convert the behavioral entropy value obtained in step one into a specific value. Compared with the structural integrity score obtained in step four A joint risk-structure consistency matching judgment is performed to determine whether the transaction request can be executed. This judgment mechanism is the system's "dynamic security gate," and its logical design directly reflects the closed-loop strategy of "guiding the encrypted structure with behavioral risks and feeding back risk decisions with structural results" in the present invention.

[0066] Furthermore, to achieve joint judgment, we design a risk-structure consistency mapping function. Its core idea is: the system has an entropy value for each behavior. A minimum tolerable structural integrity threshold should be set. If the current structure score If the value falls below this threshold, it indicates that the current transaction structure does not meet the safety requirements of its behavioral risk level, and the transaction must be terminated. We define the following execution control formula:

[0067] in: This is a control signal; 1 indicates that execution is allowed, and 0 indicates that execution is denied. It is a dynamic threshold function that defines the minimum structural integrity required under the current risk level; It is an indicator function that outputs 1 if the condition is true, and 0 otherwise.

[0068] Furthermore, to further demonstrate the feasibility of this solution in practical deployment in the mobile payment field, we propose... One implementation is a piecewise linear function, for example: like ,but ; like ,but ; like ,but .

[0069] This function can be set through system configuration files or generated adaptively through data backtesting, reflecting the system's logic for higher structural requirements of high-risk transactions.

[0070] For example, if the entropy of a transaction... This is a high-risk transaction, as set by the system. The current success rate of decrypting the fragment is... ,but The system will suspend the execution of the transaction and report it to the risk control system. Conversely, if ,and Even if the restructuring fails, the system can still allow the transaction according to a lenient strategy to avoid mistakenly killing normal low-risk behavior.

[0071] The innovation of this judgment logic lies in the fact that it does not follow the principle of "the higher the structural integrity, the better" or "the higher the risk, the more likely it is to be rejected." Instead, it establishes a "response consistency function" relationship between structural integrity and risk level, thereby realizing a true encryption-decryption strategy and a dynamic risk collaborative control mechanism. Furthermore, this method does not rely on deep learning models or user profiling; it completes transaction judgment solely based on real-time input, making it highly deployable and easy to debug and audit.

[0072] This invention also provides a mobile payment system for dynamic risk control based on reinforcement learning, the system comprising: The transaction request acquisition unit is used to acquire the user's real-time payment transaction request, extract the real-time features of the payment transaction request and preprocess them to construct a transaction state vector, and input the transaction state vector into a pre-trained probability model to calculate the behavioral entropy value representing the risk level of the current payment transaction request. The request analysis unit is used to input the transaction state vector and the behavior entropy value into a pre-trained reinforcement learning policy model to generate encryption control parameters; wherein, the encryption control parameters include the number of payment instruction segments, the highest strength perturbation parameter, the QR code scanning-NFC-QR code scanning path allocation, and the segment distribution order; A quantum-resistant encryption unit is used to segment the payment instruction according to the encryption control parameters, generate several segments, encrypt the several segments using a quantum-resistant encryption algorithm to generate encryption subkey fragments, and distribute them in a specified order through multiple communication channels. The request decryption unit is used to indicate that the server receives the encrypted subkey fragment, decrypts and reassembles it to generate a restored original payment instruction, and calculates the structural integrity score of the reassembled restored original payment instruction; The transaction execution unit is used to perform consistency matching judgment based on the behavior entropy value and the structural integrity score to determine whether to execute the current payment transaction request.

[0073] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0074] In the embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection of apparatuses or units may be electrical, mechanical, or other forms.

[0075] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0076] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A mobile payment method for dynamic risk control based on reinforcement learning, characterized in that, The method includes the following steps: The system acquires real-time payment transaction requests from users, extracts and preprocesses the real-time features of these requests to construct a transaction state vector, and inputs the transaction state vector into a pre-trained probability model to calculate the behavioral entropy value that represents the risk level of the current payment transaction request. The transaction state vector and the behavior entropy value are input into a pre-trained reinforcement learning policy model to generate encryption control parameters; wherein, the encryption control parameters include the number of payment instruction segments, the highest strength perturbation parameter, the QR code scanning-NFC-QR code scanning path allocation, and the segment distribution order; According to the encryption control parameters, the payment instruction is fragmented to generate several fragments. The quantum-resistant encryption algorithm is used to encrypt the fragments to generate encryption subkey fragments, which are then distributed in a specified order through multiple communication channels. The server receives the encrypted subkey fragment, decrypts and reassembles it to generate a restored original payment instruction, and calculates the structural integrity score of the reassembled restored original payment instruction. Consistency matching is performed based on the behavioral entropy value and the structural integrity score to determine whether to execute the current payment transaction request.

2. The mobile payment method for dynamic risk control based on reinforcement learning according to claim 1, characterized in that, The real-time features include transaction timestamp, transaction amount, device identifier, and payment method code; The preprocessing specifically includes: The transaction timestamp is determined to have a periodic characteristic within a day; the transaction amount is logarithmically compressed; and the device identifier is converted using a hash mapping. Range values; normalization of the payment method encoding.

3. The mobile payment method for dynamic risk control based on reinforcement learning according to claim 1, characterized in that, The pre-trained probabilistic model is a Gaussian mixture model. The input is a transaction state vector, and the output is a behavioral entropy value, which is used to assess the risk level of the current state.

4. The mobile payment method for dynamic risk control based on reinforcement learning according to claim 1, characterized in that, The step of inputting the transaction state vector and the behavior entropy value into a pre-trained reinforcement learning policy model to generate encryption control parameters specifically involves: The transaction state vector and the behavior entropy value are concatenated to form a joint input feature; Based on the joint input features, a resource cost regularization term and a structural risk loss term based on policy parameters are introduced to calculate encryption control parameters; wherein, the resource cost regularization term is used to limit the perturbation complexity of the policy and the consumption of the communication channel; the structural risk loss term is used to measure the degree of adaptation of the encryption policy to the current behavior entropy value. The strategy parameters include the number of subkeys, the disturbance strength level, the communication path sequence, and the key transmission order.

5. The mobile payment method for dynamic risk control based on reinforcement learning according to claim 4, characterized in that, The structural risk loss term is calculated by comprehensively calculating the total encryption complexity of the number of key splits, the level of perturbation, and the path perturbation degree, combined with the behavioral entropy value. The resource cost regularization term is one or more of the following: the encryption computation quantity increases linearly with the number of subkeys, the communication path switching cost, and the disturbance level is related to the encryption and decryption overhead.

6. The mobile payment method for dynamic risk control based on reinforcement learning according to claim 1, characterized in that, The process of segmenting the payment instruction according to the encryption control parameters to generate several segments, encrypting these segments using a quantum-resistant encryption algorithm to generate encryption subkey fragments, and distributing them in a specified order through multiple communication channels specifically includes: Obtain the number of segments and the payment instruction itself. Logically divide the payment instruction into segments equal to the number of segments, as sub-segments. If the payment instruction cannot be evenly divided, add a padding identifier to the corresponding segment, and the server restores it using offset bits. For each sub-segment, combining the perturbation strength and the current sub-segment risk perception factor, quantum-resistant encryption is performed to generate encrypted subkey fragments; Based on the encrypted subkey fragments, and according to the communication path selection vector and transmission order control vector, all encrypted subkey fragments are subjected to channel scheduling and transmission tasks. A fragment-level path perturbation regularization term is introduced to measure the perturbation strength of encrypted fragments in both spatial path and temporal order dimensions, and serves as an evaluation function for the current allocation strategy. A scheduling record table is constructed by combining the communication path selection vector, the transmission order control vector, and the encryption subkey fragment.

7. The mobile payment method for dynamic risk control based on reinforcement learning according to claim 6, characterized in that, The quantum-resistant encryption uses a standard ring-based error encryption function; The channels include a QR code scanning channel and an NFC channel, which are marked as 0 and 1 respectively; The channel scheduler calls different interfaces according to the communication path selection vector and sends the corresponding encrypted subkey fragments in the order set by the transmission order control vector.

8. The mobile payment method for dynamic risk control based on reinforcement learning according to claim 6, characterized in that, The server receives the encrypted subkey fragment, decrypts and reassembles it to generate a restored original payment instruction, and calculates the structural integrity score of the reassembled restored original payment instruction, specifically including: The server sorts the scheduling record table in ascending order according to the segment sending order in the communication path selection vector, and generates a decryption queue. The standard deviation value is obtained by looking up the table based on the highest intensity perturbation parameter, the sorted ciphertext fragment is generated, and the RLWE decryptor is initialized. After all the encrypted subkey fragments are decrypted, the server concatenates the successfully decrypted fragments in order to obtain the payment restoration instruction; Based on the encrypted subkey fragments, a structural integrity scoring function is designed to calculate the structural integrity score of each encrypted subkey fragment, which is used to measure the usability of the entire key chain after decryption and reconstruction. If the structural integrity score is lower than the preset threshold, the transaction will be terminated directly and the anomaly will be recorded to avoid misunderstanding of secrets or fraud involving partial information.

9. The mobile payment method for dynamic risk control based on reinforcement learning according to claim 1, characterized in that, The step of performing a consistency matching judgment based on the behavioral entropy value and the structural integrity score to determine whether to execute the payment transaction specifically includes: A minimum tolerable structural integrity threshold is set for each of the aforementioned behavioral entropy values. If the current structural integrity score is less than the minimum structural integrity threshold, it indicates that the current transaction structural strength does not meet the security strength required by its behavioral risk level, and the transaction must be terminated.

10. A mobile payment system for dynamic risk control based on reinforcement learning, characterized in that, The system includes: The transaction request acquisition unit is used to acquire the user's real-time payment transaction request, extract the real-time features of the payment transaction request and preprocess them to construct a transaction state vector, and input the transaction state vector into a pre-trained probability model to calculate the behavioral entropy value representing the risk level of the current payment transaction request. The request analysis unit is used to input the transaction state vector and the behavior entropy value into a pre-trained reinforcement learning policy model to generate encryption control parameters; wherein, the encryption control parameters include the number of payment instruction segments, the highest strength perturbation parameter, the QR code scanning-NFC-QR code scanning path allocation, and the segment distribution order; A quantum-resistant encryption unit is used to segment the payment instruction according to the encryption control parameters, generate several segments, encrypt the several segments using a quantum-resistant encryption algorithm to generate encryption subkey fragments, and distribute them in a specified order through multiple communication channels. The request decryption unit is used to indicate that the server receives the encrypted subkey fragment, decrypts and reassembles it to generate a restored original payment instruction, and calculates the structural integrity score of the reassembled restored original payment instruction; The transaction execution unit is used to perform consistency matching judgment based on the behavior entropy value and the structural integrity score to determine whether to execute the current payment transaction request.

Images