Payment scene-oriented interaction intention recognition and error correction system

Abstract

The invention, which relates to the technical field of payment security, discloses a payment-scene-oriented interaction intention identification and error correction system comprising an input analysis module, an intention simulation module, a dynamic decision module, a biological verification module, an audit evidence storage module, and a cross-scene knowledge migration module. According to the method, multi-modal data such as voice, texts, images and touch tracks are integrated, structured feature vectors are generated through a cross-modal attention network, the problem of incomplete single-modal coverage is solved, cross-modal data consistency verification is achieved based on a unified semantic tag system, and the reliability of input sources is graded by combining equipment fingerprints and geographic positions, so that the reliability of the input sources is improved. A high-risk transaction protection capability is enhanced, a generative adversarial network is utilized to construct a virtual attack sample library, attacks such as tampering with characters similar in shape and AI faking voiceprints are simulated, unknown threats are actively defended through cosine similarity matching, a user historical behavior statistical model is integrated, and known risks such as high-frequency small-amount transfer are passively intercepted. And a closed-loop incremental learning continuous optimization model is supported.

Description

Technical Field

[0001] The present invention relates to the field of payment security technology, and in particular to an interactive intention recognition and error correction system for payment scenarios. Background Art

[0002] In the modern digital payment ecosystem, user needs and transaction scenarios are becoming increasingly complex, covering a variety of scenarios such as daily transfers, cross-border remittances, bill payments, charitable donations, commercial settlements, etc. Users may complete payments through various methods such as voice commands, manual input, code scanning, touch operations, etc. In different scenarios, there are significant differences in the accuracy and security requirements of information input.

[0003] Existing systems are mostly based on a single modality to analyze payment intentions, without integrating multi-source data, resulting in missing key information or logical conflicts, low entity recognition accuracy, lack of a unified semantic labeling system and heterogeneous data alignment mechanism, inability to achieve automatic association and verification of cross-modal data, and inability to defend against new attacks. Strategies such as mandatory multi-factor authentication do not take into account user operating habits and device performance differences, resulting in a significant decline in user experience in high-security scenarios, high complexity of the verification process, lack of dynamic optimization mechanism, long user waiting time, lack of scenario association modeling capabilities, inability to migrate risk features through graph neural networks, inability to achieve data consistency verification in a multi-chain environment, low audit efficiency and insufficient credibility. Summary of the Invention

[0004] The purpose of the present invention is to address the shortcomings of the prior art and to propose an interactive intention recognition and error correction system for payment scenarios.

[0005] To achieve the above objectives, the present invention adopts the following technical solutions: an interactive intention recognition and error correction system for payment scenarios, including an input parsing module, an intention simulation module, a dynamic decision-making module, a biometric verification module, an audit and evidence storage module, and a cross-scenario knowledge transfer module. The input parsing module performs multi-source data collection, modal feature extraction, heterogeneous data alignment, and structured feature fusion to output a payment intention feature vector. The intention simulation module is used to generate virtual attack samples through a generative adversarial network and calculate the cosine similarity between the user's current operation characteristics and the virtual attack samples in combination with a statistical model for user historical behavior analysis. The dynamic decision-making module, based on a deep reinforcement learning framework, combines the payment intention feature vector and risk assessment results to dynamically assign error correction strategy priorities through a three-dimensional reward function. The biometric verification module generates a comprehensive verification confidence by fusing voiceprint, touch pen pressure, and iris dynamic characteristics to adapt to transaction scenarios with different risk levels. The audit and evidence storage module builds a consortium chain network based on blockchain technology, generates semantic fingerprints of payment instructions, and performs compliance verification through smart contracts. The cross-scenario knowledge transfer module models a heterogeneous graph of payment scenarios using graph neural networks and knowledge graph technologies to achieve generalized transmission of risk features and sharing of model parameters.

[0006] As a further description of the above technical solution: The input parsing module collects voice commands through the microphone array sampling rate, combines the environmental noise suppression algorithm to improve the signal-to-noise ratio, converts the voice into text commands through the voice recognition engine, captures user operation behavior through the camera, uses the YOLOv7 algorithm to identify payment credentials, and records the pressure and time curve of the touch trajectory. It achieves multi-device time synchronization based on the NTP protocol, attaches millisecond timestamps and device geographic location information to the input data, and outputs the payment intention feature vector through cascade fusion and cross-modal attention weighted dimensionality reduction of voice, text and image features.

[0007] As a further description of the above technical solution: The intent simulation module builds an adversarial sample library covering 12 types of payment fraud scenarios. It uses a statistical model built based on the user's historical transaction frequency, device credibility, and geographic location risk to predict periodic transaction behavior and match user operation characteristics with adversarial samples through cosine similarity.

[0008] As a further description of the above technical solution: The dynamic decision module triggers the following verification strategies based on the risk score: for level one risk, voiceprint, iris and touch handwriting pressure authentication is mandatory; for level two risk, a pop-up window is displayed for secondary confirmation and an associated transaction limit reminder is provided; for level three risk, a text message verification code is sent and a correction suggestion is provided if the payee name similarity is ≥90%.

[0009] As a further description of the above technical solution: The biometric verification module extracts the Mel-frequency cepstral coefficients of the voiceprint, detects the coherence of natural pronunciation to distinguish forgery attacks, analyzes the dynamic entropy value of touch handwriting pressure to identify mechanical script attacks, and uses a 3D structured light module to capture the three-dimensional characteristics of iris texture and living body vibration frequency.

[0010] As a further description of the above technical solution: The audit evidence module builds a consortium chain network through a practical Byzantine fault-tolerant consensus mechanism, deploys compliance review contracts, policy execution contracts and cross-chain verification contracts, and performs transaction compliance verification.

[0011] As a further description of the above technical solution: The cross-scenario knowledge transfer module constructs a payment scenario heterogeneous graph based on scenario feature vectors and risk pattern nodes, aggregates high-frequency scenario features to low-frequency scenario nodes through a layered graph attention network, and uses differential privacy technology to synchronize the migrated model parameters to the intention simulation module and the dynamic decision-making module.

[0012] The present invention has the following beneficial effects: 1. The present invention uses multimodal fusion, generative adversarial defense, dynamic strategy optimization, biometric authentication collaboration, privacy-enhanced evidence storage, and cross-scenario migration technologies to achieve accurate analysis of payment intentions, active defense against unknown attacks, a dynamic balance between user experience and security, and full-link trusted auditing. It constructs a heterogeneous graph of payment scenarios, migrates risk features of high-frequency scenarios to low-frequency scenarios through graph neural networks, solves the problem of data sparsity, combines the differential privacy protection migration process, and synchronizes model parameters in real time through a self-healing interface to adapt to the risk distribution of different scenarios.

[0013] 2. The present invention integrates multimodal data such as voice, text, images and touch trajectories, generates structured feature vectors through a cross-modal attention network, solves the problem of incomplete single-modal coverage, implements cross-modal data consistency verification based on a unified semantic labeling system, combines device fingerprints and geographic location to grade the input source credibility, enhances the ability to protect against high-risk transactions, uses a generative adversarial network to build a virtual attack sample library, simulates attacks such as tampering with similar characters and AI-forged voiceprints, actively defends against unknown threats through cosine similarity matching, integrates user historical behavior statistical models, passively intercepts known risks such as high-frequency small-amount transfers, and supports closed-loop incremental learning to continuously optimize the model.

[0014] 3. The present invention balances user experience, security and operation speed based on a three-dimensional reward function, dynamically adjusts the weight of the verification strategy, supports a hot update mechanism to synchronize the fraud feature library in real time, combines with the alliance chain to ensure the timeliness and consistency of strategy execution, minimizes user interference, coordinates voiceprints, touch handwriting pressure and iris liveness features, dynamically allocates verification weights, resists forgery attacks, improves the quality of voiceprint data through environmental noise suppression, combines dynamic entropy analysis to distinguish real users from mechanical scripts, uses zero-knowledge proof and hash algorithm to realize the desensitization and storage of sensitive information, ensures that data cannot be tampered with, and combines the layered alliance chain network with smart contract automated auditing to support cross-chain hash locking to verify the consistency of multi-chain data. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 This is a system architecture diagram of the present invention. DETAILED DESCRIPTION

[0016] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0017] Reference Figure 1 The present invention provides an embodiment of an interactive intention recognition and error correction system for payment scenarios, comprising an input parsing module, an intention simulation module, a dynamic decision-making module, a biometric verification module, an audit and evidence storage module, and a cross-scenario knowledge transfer module. The input parsing module performs multi-source data acquisition, modal feature extraction, heterogeneous data alignment, and structured feature fusion to output a payment intention feature vector. The intention simulation module is used to generate virtual attack samples using a generative adversarial network and, combined with a statistical model for analyzing user historical behavior, calculates the cosine similarity between the user's current operation characteristics and the virtual attack samples. The dynamic decision-making module, based on a deep reinforcement learning framework, combines the payment intention feature vector and risk assessment results to dynamically assign error correction strategy priorities using a three-dimensional reward function. The biometric verification module generates a comprehensive verification confidence by fusing voiceprint, touch pen pressure, and iris dynamic characteristics to adapt to transaction scenarios of different risk levels. The audit and evidence storage module builds a consortium chain network based on blockchain technology, generates semantic fingerprints for payment instructions, and performs compliance verification through smart contracts. The cross-scenario knowledge transfer module models a heterogeneous graph of payment scenarios using graph neural networks and knowledge graph technologies to achieve generalized transmission of risk features and sharing of model parameters.

[0018] The input parsing module collects voice commands through the microphone array sampling rate, improves the signal-to-noise ratio by combining the environmental noise suppression algorithm, converts the voice into text commands through the voice recognition engine, captures the user's operation behavior through the camera, uses the YOLOv7 algorithm to identify the payment credentials, and records the pressure and time curve of the touch trajectory. It realizes multi-device time synchronization based on the NTP protocol, adds millisecond timestamps and device geographic location information to the input data, and outputs the payment intention feature vector through cascade fusion and cross-modal attention weighted dimensionality reduction of voice, text and image features. The intention simulation module constructs an adversarial sample library covering 12 types of payment fraud scenarios. The statistical model constructed based on the user's historical transaction frequency, device credibility and geographic location risk predicts periodic transaction behavior and matches user operation features with adversarial samples through cosine similarity. The dynamic decision module triggers the following verification strategies according to the risk score. The first-level risk forces voiceprint, Iris and touch handwriting pressure authentication, secondary risk pop-up window for secondary confirmation and associated transaction limit prompts, third-level risk sends SMS verification code and prompts correction suggestions with the payee name similarity ≥90%. The biometric verification module extracts the Mel-frequency cepstral coefficients of the voiceprint, detects the coherence of natural pronunciation to distinguish forgery attacks, analyzes the dynamic entropy value of touch handwriting pressure, identifies mechanical script attacks, and uses a 3D structured light module to capture the three-dimensional features of the iris texture and the frequency of living body vibrations. The audit evidence module builds a consortium chain network through a practical Byzantine fault-tolerant consensus mechanism, deploys compliance review contracts, policy execution contracts and cross-chain verification contracts, and performs transaction compliance verification. The cross-scenario knowledge transfer module constructs a payment scenario heterogeneous graph based on scenario feature vectors and risk pattern nodes, aggregates high-frequency scenario features to low-frequency scenario nodes through a layered graph attention network, and uses differential privacy technology to synchronize the migrated model parameters to the intention simulation module and dynamic decision-making module.

[0019] The input parsing module serves as the front-end data processing core of the system. It uses multimodal data fusion technology to perform multi-dimensional analysis of the user's payment intention. The workflow is divided into four stages: multi-source data acquisition, modal feature extraction, heterogeneous data alignment, and structured feature fusion. It finally outputs a standardized payment intention feature vector for downstream modules to perform risk analysis and decision-making. Voice commands are collected through a microphone array with a sampling rate of no less than 44.1kHz to ensure audio quality. An environmental noise suppression algorithm is used, combined with spectral subtraction and a deep filtering network to minimize the impact of environmental noise and increase the signal-to-noise ratio to more than 15dB. The end-to-end system based on the Conformer-Transformer hybrid model is then called. The speech recognition engine converts speech into text commands and supports keyboard input and optical character recognition (OCR). For example, when scanning account information from payment receipts, it uses the BERT-INTENT model for intent classification and identifies 17 payment scenarios, including personal transfers, corporate transfers, utility bill payments, communications bill payments, tax payments, refund processing, account top-ups, credit card repayments, investment and wealth management, red envelopes and gift cards, cross-border payments, bill payments, charitable donations, commercial payments, loan services, subscription services, and special payment scenarios. It also uses entity recognition technology to extract key fields, including transaction type, amount, payee, time constraint, currency unit, payment account number, payment purpose, and payee qualifications. The F1 score for entity recognition reaches 0.93. Capture user operation behaviors such as gestures and facial expressions through the camera, and use the YOLOv7 target detection algorithm to identify payment credentials such as QR codes and barcodes. For touch operations, record the touch trajectory at a sampling rate of 100Hz, extract pressure and time curve features for dynamic handwriting analysis, and achieve multi-device time synchronization based on the NTP protocol. Add millisecond-level timestamps to all input data and bind device geographic location information, such as GPS, Beidou or base station positioning, with an accuracy of less than 500 meters. Build a unified semantic labeling system that complies with the ISO20022 standard extension, covering key fields such as transaction type, amount, and payee. For example, the payee information in the voice command "Transfer 500 to Zhang San" needs to be automatically associated with the payee list input by touch to ensure consistency across modal data. Extract Mel-frequency cepstral coefficients and Log-Mel spectrum features from voice data to generate a 128-dimensional feature vector. Extract 768-dimensional embedding vectors from text data through the BERT model to capture deep semantic information. Image data The fully connected layer of the ResNet-50 model extracts 2048-dimensional visual features and combines them with cascaded fusion and attention weighting. Cascaded fusion concatenates the feature vectors of speech, text, and image into a 2944-dimensional mixed vector. Attention weighting calculates the weights of each modality using a cross-modal attention network. Specifically, the concatenated feature vectors are input into a weight matrix. After normalization using the Softmax function, attention weights are assigned to the different modalities, ultimately reducing the dimension to 128 dimensions. The resulting payment intent feature vector is then output. Input integrity verification enforces the integrity of key fields, such as ensuring that the amount cannot be left blank, and detects logical conflicts between multimodal data. For example, if the voice command "transfer 500 yuan" and the manually entered amount "5000 yuan" differ, a conflict alert is triggered to assess data credibility. The legitimacy of the operation source is verified based on device fingerprints, MAC addresses, SIM card ICCIDs, and Bluetooth device lists. Input sources are also graded for credibility: camera > voice > manual input.

[0020] The intention simulation module is the core protection unit of the payment security system. Through the synergy of generative adversarial networks and historical behavior analysis, it realizes passive and active defense against payment risks, creates a virtual attack pattern library and a high-risk transaction pattern library. The high-risk transaction pattern library stores known attack patterns. The intention simulation module collects user voice data, text data, image / video data, touch trajectory data, device fingerprint data, geographic location data, transaction record data and biometric data from the user's payment behavior, and extracts core features from them. The extracted features are organized into a standard format and stored in the database to construct a feature library. The virtual attack pattern library outputs virtual transaction instructions with semantic confusion and fraud features through the generator, which is the adversarial sample. For example, For example, if "transfer 5,000 yuan to Zhang San" is tampered with to "transfer five thousand to Zhang San", the discriminator distinguishes between real transaction data and generated adversarial samples, and continuously optimizes the generator's fraud feature generation ability through adversarial training. At the same time, it improves the discriminator's ability to identify unknown attack patterns and predict new attacks. The simulated attack types cover 12 types of payment fraud scenarios, including high-frequency small-amount transfer attacks that simulate money laundering, cross-border currency unit tampering attacks for exchange rate fraud, and phishing attacks that confuse similar names of payees with similar characters. The user's current operation features are compared with the adversarial samples by cosine similarity. The cosine similarity is the cosine value of the angle between two feature vectors in space, which measures their directional consistency. The value range is between 0 and 1. The closer the value is to 1, the higher the similarity. The formula is , is the feature vector of user operation and the feature vector of virtual attack sample, It is the modulus length of the vector. The semantic feature vector of the user's current operation is extracted, including the instruction text, amount, and payee attributes, and cosine similarity matching is performed with the virtual attack sample. When the similarity between the user's operation feature and the attack sample exceeds the cosine similarity threshold, it is marked as a high-risk operation. A real-time update mechanism is set to add new attack types and adjust the feature cosine similarity threshold in real time, such as "AI synthetic voiceprint phishing attack". The corresponding sample is automatically generated, and the cosine similarity threshold is dynamically adjusted according to the feedback of the system's interception effect. The system's risk resistance is improved through "attack and defense" closed-loop training. The high-risk transaction pattern library builds a statistical model based on the user's historical behavior. The user's historical behavior includes transfer time, transaction amount, geographic location, and historical transaction frequency. , common payees, device credibility, payment scenario preferences, transaction cycle patterns, touch operation characteristics, biometric authentication habits and risk operation records, so as to predict users' periodic transaction behaviors and passively defend against known abnormal patterns. The statistical models include time series analysis model, spatial behavior model, amount threshold model and user behavior baseline model. The time series analysis model adopts LSTM-ARIMA hybrid model to analyze users' periodic transaction behaviors. The spatial behavior model clusters users' common transaction locations based on DBSCAN algorithm, establishes a geographical fence with an accuracy of 500 meters, and connects to China Mobile's base station positioning API. The amount threshold model combines regional consumption level and user income portrait to dynamically adjust the single / single-day transfer limit. The amount threshold = The user's historical mean + 3 × standard deviation, the amount threshold is updated every day, the user behavior baseline model counts the user's common payee and payment scenario preferences, such as personal transfers and cross-border payments, based on device fingerprints, such as MAC address, SIM card ICCID and Bluetooth device list, to verify the legitimacy of the operation source, rebuild the user behavior baseline model every week to adapt to changes in consumption habits, each new transaction triggers an incremental update of the feature library, calculates the false alarm rate and missed alarm rate indicators through the confusion matrix, dynamically adjusts the parameters of the statistical model, prevents abnormal transactions, sets anomaly detection rules, identifies operations that deviate from the norm, and triggers risk warnings. Time anomalies are non-working hours and large transactions are initiated. Spatial anomalies are operations where the location is more than 50 kilometers away from the usual location. Amount threshold anomalies are single transactions. If the transfer amount exceeds three standard deviations of the historical mean, a multi-dimensional user behavior baseline is established to accurately intercept known anomalies. Incremental learning and rule engine optimization are used to ensure the system's continuous evolution. A dynamic threshold adjustment mechanism is used to set risk level scores. The user behavior baseline model generates a risk score: (historical transaction frequency score × 0.4) + (device credibility score × 0.3) + (geographic location risk score × 0.3). The risk score ranges from 0 to 1, with higher risk scores indicating greater alert sensitivity. Geographic location risk is obtained from the public security filing database. The threshold calibration formula is dynamic threshold = 1 × (1 - risk score). Risk alerts are graded, with a level 1 risk score of cosine similarity ≥ 0.85 and matching known attack patterns in the high-risk transaction pattern library. Level 2 risk refers to cosine similarity ≥ 0.7 and matching dynamic threshold conditions in the high-risk transaction pattern library, such as single transaction amount > historical mean + 3 standard deviations and geographic location offset > 50 kilometers. Level 3 risk refers to regular logical conflicts that only match the high-risk transaction pattern library, such as inconsistent multimodal data and missing key fields. Based on the priority weights of the reinforcement learning model, optimization suggestions are generated, such as adjusting the amount, splitting large transactions, and prompting the recipient to display a list of similar contacts. The name similarity must be ≥ 90%. The analysis results of the two libraries are combined to trigger real-time risk classification warnings.

[0021] The dynamic decision-making module is based on a deep reinforcement learning framework and optimizes risk response strategies in payment scenarios in real time through a three-dimensional reward function. Its core lies in establishing an adaptive strategy priority adjustment mechanism to comprehensively balance user operation experience, risk interception efficiency and response speed. The module receives the standardized payment intention feature vector output by the input parsing module, integrates the 128-dimensional structured features of multimodal data such as voice, text, and images, and the risk assessment results of the intention simulation module, including the cosine similarity matching value between user operations and adversarial samples, and the anomaly detection results of the high-risk transaction pattern library. Combined with real-time transaction data and user behavior baseline models, it dynamically generates risk level scores and assigns error correction strategy priorities. Through a closed-loop feedback mechanism, based on user experience, the module The model parameters are continuously calibrated based on the user's final operation results and interception effects to ensure the real-time and accuracy of the strategy. The deep reinforcement learning model generates dynamic risk scores based on the risk assessment results of the intent simulation module and the user behavior baseline model, and assigns error correction strategy priorities to payment operations with different risk levels. Level 1 risk triggers mandatory level 3 biometric authentication, voiceprint, iris and touch handwriting pressure detection, level 2 risk pop-up window for secondary confirmation and associated transaction limit prompts, verifies the legitimacy of the device, level 3 risk sends SMS verification codes and intelligently prompts error correction suggestions. Through the three-dimensional reward function mechanism, through the dynamic game of user convenience, transaction security and operation speed in three dimensions, Pareto optimality is achieved, forming a risk decision-making system with self-evolution capabilities. The formula is , It is the weight of user convenience, transaction security, and operation speed. According to the actual situation, the strategy is dynamically adjusted to prevent users from feeling that verification is too troublesome, and to prevent risky transactions while ensuring that the entire process is not stuck. The weight of user convenience α=0.6. This dimension focuses on the active maintenance of user experience. The reinforcement learning model gives priority to verification strategies with high user acceptance, and gives priority to verification methods that users are most accustomed to and have the least operational resistance. It analyzes user historical operations, such as frequently used SMS verification or fingerprint verification, and automatically matches preference strategies. If a certain verification method causes users to cancel frequently, such as a high failure rate in face recognition, the system will lower its priority. In low-risk scenarios, the verification steps are simplified by default. If a certain verification method If it is often canceled by the user midway, the system will automatically reduce its frequency of use. For example, in low-risk scenarios, SMS verification codes are preferred rather than mandatory biometric authentication to avoid interference to users caused by excessive verification. When the user's historical operations show high-frequency use of voice confirmation, the priority of voice secondary verification is automatically increased. By monitoring the user interruption rate in real time, the triggering frequency of highly complex verification strategies is dynamically reduced. The transaction security weight β=0.3. This dimension ensures the achievement of core security goals and matches known fraud patterns. For 12 types of known fraud scenarios, a minimum interception success rate threshold is set, such as high-frequency small-amount transfers. Multiple verifications are mandatory at the same time. For new attack patterns, such as AI forged voiceprints, temporary increases are made. The weight ratio of this dimension ensures the reliability of interception by sacrificing some user experience, even if it increases the verification time. Combined with the fraud feature library shared by the alliance chain, the list of fraud scenarios that need to be protected is dynamically expanded. For example, when the cosine similarity matching value exceeds 0.85, even if it may increase the user operation time, it is still necessary to enforce the high-risk verification strategy. The weight of the operation speed γ=0.1. This dimension controls the complexity of the verification steps to prevent users from waiting too long. For multi-step verification processes, such as biometric authentication plus SMS secondary confirmation, an exponential time penalty is imposed, forcing the model to give priority to single-step verification solutions and set a hard time threshold. For example, a single verification must not exceed 8 seconds, and the timeout strategy will automatically downgrade to a lower security level. A high-level alternative solution uses device performance awareness to automatically enable lightweight authentication processes for low-end mobile devices, such as replacing passwords with fingerprints. For older phones or with poor network connections, lightweight authentication is automatically enabled, such as simply entering a text message code instead of scanning a face. Based on actual user results, such as authentication pass / rejection rates, process interruption rates, and system interception effectiveness, error signals are backpropagated hourly to adjust the weights of α, β, and γ. For example, when a new type of phishing attack causes the interception rate to drop, the β weight is automatically increased to 0.35. Combined with the payment intent knowledge graph, the weighting system is dynamically restructured for different scenarios. For example, in cross-border payment scenarios, β is temporarily increased to 0.4, while in personal small-value transfer scenarios, α is increased to 0.65. When a metric in a particular dimension exceeds a critical threshold, such as a user churn rate exceeding 5%, an emergency rebalancing algorithm is triggered, temporarily freezing weight updates and reactivating a pre-set baseline policy to prevent the system from falling into a local optimal trap. User feedback on final operations, such as verification pass rate and interception false positive rate, and system performance indicators such as average response time, are collected as training samples for reinforcement learning. In collaboration with the audit evidence module, the algorithm utilizes immutable evidence recorded on the blockchain to verify policy execution consistency, such as the timestamp of user confirmation operations and the results of smart contract execution, to calibrate model biases. A real-time hot update mechanism is supported, allowing model parameter iteration without downtime, ensuring that defense strategies continuously adapt to new attack patterns. Integrating the payment intent knowledge graph, such as requiring the linkage of foreign exchange accounts for cross-border payments and verification of recipient qualifications for charitable donations, allows for dynamic adjustment of policy priorities. Fraud signature libraries are shared through the consortium chain, allowing for immediate updates to the policy library to cover unknown threats. Through this mechanism, the dynamic decision-making module ensures a high interception rate while minimizing disruption to the user experience, achieving both improved payment security and efficiency.

[0022] The biometric verification module achieves high-security identity authentication by integrating the user's multi-dimensional biometric features, combining voiceprints, touch pen pressure and iris dynamic characteristics. Its core lies in the collaborative analysis and dynamic calibration of multimodal features to comprehensively determine the authenticity of the user's identity and resist forgery attacks. The module generates comprehensive verification confidence through real-time collection, feature extraction and fusion analysis, dynamically adapts to transaction scenarios with different risk levels, and ensures a balance between security and user experience. The user's voice commands are collected through the microphone, covering the voiceprint frequency band of 0.8kHz to 12kHz, ensuring that the core frequency characteristics of the human voice are captured. Use environmental noise compensation algorithm to eliminate background noise and increase the signal-to-noise ratio to more than 15dB to ensure the purity of voiceprint data. Analyze the spectral characteristics of voiceprint, including pitch, resonance peak distribution and pronunciation habits, generate a unique voiceprint feature template, and characterize the personalized pattern of user voice through Mel-frequency cepstral coefficients. Combined with the natural coherence of pronunciation, it can distinguish between real-person voice and forged attacks such as recording playback. Detect touch handwriting pressure, record the user's operation trajectory on the touch device in real time at a sampling rate of 100 times per second, capture the curve of pressure value changing over time, and extract the "dynamic entropy value" of touch handwriting, such as pressure. The fluctuation pattern of the distribution of touch and the change of touch speed form the unique touch behavior characteristics of the user. For example, the distribution of the force of the user pressing the screen and the smoothness of the finger sliding trajectory can be used as the basis for judgment. By analyzing the continuity and naturalness of the touch operation, the real user operation and the attack behavior simulated by the mechanical script are distinguished. The 3D structured light module is used to scan the user's iris, capture the detailed information of the iris texture, and detect the tiny vibration frequency of the iris in the living state, such as pupil contraction and eye movement, to ensure that the verification object is a real human body. The three-dimensional structural characteristics of the iris texture can resist the detection of photos, videos or 3D printing. To prevent forgery attacks on fingerprint models, the verification results of voiceprint, touch handwriting pressure and iris features are integrated and analyzed, and each modality is weighted to generate a comprehensive verification confidence score. Different levels of security measures are triggered according to the confidence score. High confidence directly passes the verification, medium confidence triggers secondary verification, low confidence is forced to intercept and prompts risks, and the weight distribution is dynamically optimized based on historical verification data. For example, when the user's touch handwriting is abnormal multiple times, the priority of iris verification is automatically increased. Through the above mechanism, the biometric verification module achieves the synergistic effect of high precision and low error rate in payment scenarios, significantly improving transaction security.

[0023] The audit evidence module builds a distributed ledger network based on blockchain technology, and realizes the tamper-proof evidence and full life cycle tracking of key operation data in payment scenarios through multi-node consensus mechanism and cryptographic algorithm. Its core principle is to use the distributed characteristics of the alliance chain to encapsulate the data of key nodes such as transaction decision-making, user confirmation and system correction into blocks with timestamps, combine with smart contracts to automatically execute audit rules, and use zero-knowledge proof technology to achieve privacy protection of sensitive information, ensure the integrity, traceability and compliance of evidence data, receive the original data of payment instructions from the input parsing module, including semantic features of modalities such as voice, text, and image, the error correction strategy version identifier generated by the dynamic decision module, and the comprehensive verification device of the biometric verification module. The trust score, as well as the timestamp of the user's final confirmation operation, device fingerprint, geographic location information and other key fields, are structured and encapsulated by a unified semantic tag system for multi-source data to generate a standardized evidence data package. The standardized evidence data package is preprocessed, the core fields are extracted and their semantic fingerprints are calculated, and the SHA-256 hash algorithm is used to generate a unique data fingerprint to ensure that any data tampering can be quickly detected. At the same time, sensitive fields are encrypted through asymmetric encryption technology, and zero-knowledge proof technology is used to realize desensitized evidence storage of key information. For example, only the hash value is disclosed and the original data content is hidden. The encrypted evidence data package and the corresponding hash value are packaged into a block, and multi-node verification and distribution are completed through the practical Byzantine fault-tolerant consensus mechanism of the alliance chain. Distributed storage, each block contains the hash value of the previous block, forming an irreversible chain structure. The evidence data covers the semantic fingerprint of the payment instruction: the hash value and standardized feature vector of the original instruction, dynamic decision correction record: error correction strategy version identification and triggering conditions, user confirmation credentials: the timestamp of the final operation, device fingerprint and geographic location, smart contract execution summary: the call results and status log of the associated contract, deployment of on-chain smart contracts, preset audit rules, such as cross-border payments require verification of foreign exchange account qualifications, large transactions require the digital signature of the legal person, when the payment process triggers specific events such as successful risk interception and user confirmation timeout, the smart contract automatically performs compliance verification and writes the execution results into the blockchain, such as pass and rejection and the reason code. , contract status changes are synchronized to other modules in real time through an event-driven mechanism, such as the strategy optimizer of the dynamic decision-making module, supporting cross-chain interaction with external blockchain networks, and realizing cross-chain hash lock verification through relay chain technology to ensure the consistency of evidence data in a multi-chain environment. It adopts zero-knowledge concise non-interactive knowledge argumentation technology to provide transaction compliance proof to regulators without leaking the original data. For example, it verifies the authenticity of "a certain charitable donation has passed the qualification review" without disclosing the specific information of the payee. It provides a standardized API interface for authorized parties to query evidence records based on time range, transaction type, risk level and other dimensions, and supports reverse verification of data integrity through hash value, that is, input evidence hash value and match it with the on-chain block hash.If they are consistent, it proves that the data has not been tampered with. At the same time, based on the Merkle tree structure, the associated evidence nodes of a specific transaction can be quickly located to improve audit efficiency. The consortium chain is jointly maintained by pre-selected authoritative institutions such as banks, payment platforms, and regulatory agencies as consensus nodes. Node access adopts a digital certificate authentication mechanism. Each participant must submit an institutional qualification certificate and complete identity binding through the issuance of an X.509 certificate by a certificate authority. Nodes are divided into three roles. The core accounting node is responsible for block generation and verification, such as commercial banks. The regulatory audit node only has data query permissions, such as central bank branches. Lightweight access nodes support data on the chain but have no consensus rights, such as third-party payment platforms. A layered network architecture is adopted, and the core accounting nodes are interconnected through dedicated lines to form a main chain network. , the regulatory audit nodes are connected to the main chain through an encrypted tunnel, and the consensus mechanism uses practical Byzantine fault tolerance, requiring more than 2 / 3 of the nodes to reach a consensus to confirm the new block, ensuring that the system can still run reliably when the number of malicious nodes does not exceed 1 / 3, the block generation cycle is fixed at 2 seconds, and the capacity of a single block is limited to 2MB, balancing throughput and latency. The transaction evidence data contains four core fields. The semantic fingerprint is the SHA-256 hash value of the payment instruction. The correction record is the policy version ID and trigger conditions generated by the dynamic decision module. The user credential is the final confirmation timestamp, device fingerprint, and geographic location. The contract summary is the execution result log of the associated smart contract. Each block contains the hash value of the previous block, forming an irreversible chain structure. Tampering with any block will result in subsequent hash chains. It uses relay chain technology to interact with external blockchains, adopts hash time lock protocol, generates encrypted hash locks in cross-chain transactions, and requires the recipient to provide original data proof within the specified time, otherwise the transaction will be automatically rolled back. For example, cross-border payment instructions need to generate matching hash locks on both the local chain and the target chain to ensure cross-chain data consistency. Sensitive fields are stored through asymmetric encryption and can only be decrypted by authorized nodes. When providing data to regulators, zero-knowledge proof technology is used to generate compliance proof, such as verifying the authenticity of "a certain transaction has completed foreign exchange qualification review" without disclosing user account details. The audit interface supports retrieval of evidence by time range, transaction type and other dimensions, and quickly verifies data integrity through Merkle tree. Three types of deployments On-chain smart contracts: Compliance review contracts: Automatically verify the qualifications of foreign exchange accounts for cross-border payments, the registration status of charitable donation recipients, and strategy execution contracts: According to the priority weight of the dynamic decision-making module, trigger the specified verification process, such as mandatory three-level biometric authentication, cross-chain verification contracts: Manage hash lock generation and release to ensure the atomicity of cross-chain transactions. The contract code is solidified on the chain after multi-party joint audits. The execution results are synchronized to each business module in real time. The device fingerprint and geographic location data standards of the audit evidence module are used, and the smart contract execution results are shared with the dynamic decision-making module to form a closed-loop strategy execution. The semantic fingerprint generation algorithm of the intent simulation module is used to ensure the consistency of cross-module data association. The architecture uses strict node access, layered consensus and privacy enhancement technology.Build a trusted alliance chain that meets financial regulatory requirements and supports the full-link audit and evidence storage needs of payment scenarios.

[0024] The cross-scenario knowledge transfer module uses graph neural network and knowledge graph technology to achieve the generalization transfer and model sharing of risk features between different payment scenarios. At the same time, relying on the technical framework of existing modules, it ensures the overall technical consistency and functional complementarity of the system. Its core goal is to solve the problem of model performance degradation caused by sparse data in low-frequency scenarios. By modeling the relationship between scenarios, the risk identification capabilities trained in high-frequency scenarios such as personal transfers are transferred to low-frequency scenarios such as charitable donations, while avoiding functional overlap with modules such as intent simulation and dynamic decision-making. The standardized output of the 128-dimensional structured feature vector is obtained from the input parsing module, including multimodal fusion features such as transaction type, amount, and payee attributes, and combined with the scenario risk label generated by the intent simulation module. Signatures, such as "cross-border currency tampering attack", construct scenario feature vectors for each payment scenario, use knowledge graph embedding technology to map scenario semantic labels into low-dimensional vectors to form scenario semantic representations, and construct payment scenario heterogeneous graphs based on scenario feature vectors and semantic labels. The nodes in the graph are divided into two categories: scenario nodes represent specific payment scenarios, such as "corporate payment" and "cross-border transfer", and node attributes include scenario feature vectors and risk labels. Risk pattern nodes represent the 12 types of attack patterns defined by the intent simulation module, and node attributes include the semantic fingerprint of the adversarial sample and the cosine similarity threshold. The edge weight is calculated by cosine similarity to measure the correlation between risk features between scenarios, such as the similarity of transfer amount distribution between "corporate payment" and "commercial payment", as well as the scenario The strength of association with attack patterns, such as the matching probability of "cross-border payment" and "currency unit tampering attack", the graph neural network uses a layered graph attention network to perform message passing and feature aggregation on the scene graph, propagates between scene nodes, and calculates the contribution weight of high-frequency scene nodes to low-frequency scene nodes through the attention mechanism. For example, the amount anomaly detection capability of the "personal transfer" scenario is transferred to the "charitable donation" scenario, and the scenario-risk pattern association learning is carried out to jointly optimize the embedding representation of the scene node and the risk pattern node to enhance the generalization recognition capability of the low-frequency scenario for unknown attack patterns. For example, the "recipient similar name phishing attack" detection logic trained in the high-frequency scenario is used to assist in the interception of similar risks in the low-frequency scenario, and the training objective function is integrated into the scenario. Classification loss and risk pattern matching loss ensure that the migrated model meets both high accuracy and low false alarm rate in the target scenario. The migrated risk model parameters are synchronized in real time to the high-risk transaction pattern library of the intention simulation module and the policy optimizer of the dynamic decision-making module through the self-healing strategy update interface, such as updating the cosine similarity threshold of the intention simulation module, adapting to the risk distribution characteristics of the target scenario, adjusting the three-dimensional reward function weight of the dynamic decision-making module, optimizing the verification strategy priority of low-frequency scenarios, and using differential privacy technology in the update process to ensure the desensitization protection of user data during cross-scenario migration. Collaborating with the alliance chain network of the audit evidence module, the topological structure of the scene graph, GNN model parameters and migration logs are encapsulated as evidence data readable by smart contracts.Cross-chain verification is achieved through hash locks and zero-knowledge proof technology. For example, when a newly connected payment platform lacks data on "cross-border donation" scenarios, the migrated risk model can be quickly loaded based on the scenario-related weights stored on the chain without the need for retraining.

[0025] Finally, it should be noted that the above is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or make equivalent substitutions for some of the technical features therein. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. An interactive intention recognition and error correction system for payment scenarios, featuring: It includes an input parsing module, an intention simulation module, a dynamic decision-making module, a biometric verification module, an audit evidence module, and a cross-scenario knowledge transfer module. The input parsing module performs multi-source data collection, modal feature extraction, heterogeneous data alignment and structured feature fusion, and outputs a payment intention feature vector. The intention simulation module is used to generate virtual attack samples through a generative adversarial network, and combines the statistical model of user historical behavior analysis to calculate the cosine similarity between the user's current operation characteristics and the virtual attack samples. The dynamic decision-making module is based on a deep reinforcement learning framework, combined with the payment intention feature vector and risk assessment results, and dynamically allocates the error correction strategy priority through a three-dimensional reward function. The biometric verification module generates a comprehensive verification confidence by fusing voiceprints, touch handwriting pressure and iris dynamic characteristics, and adapts to transaction scenarios with different risk levels. The audit evidence module builds a consortium chain network based on blockchain technology, generates semantic fingerprints of payment instructions, and performs compliance verification through smart contracts. The cross-scenario knowledge transfer module models a heterogeneous graph of payment scenarios through graph neural networks and knowledge graph technologies to achieve generalized transmission of risk features and sharing of model parameters.

2. The interactive intention recognition and error correction system for payment scenarios according to claim 1 is characterized by: The input parsing module collects voice commands through the microphone array sampling rate, combines the environmental noise suppression algorithm to improve the signal-to-noise ratio, converts the voice into text commands through the voice recognition engine, captures user operation behavior through the camera, uses the YOLOv7 algorithm to identify payment credentials, and records the pressure and time curve of the touch trajectory. It achieves multi-device time synchronization based on the NTP protocol, attaches millisecond timestamps and device geographic location information to the input data, and outputs the payment intention feature vector through cascade fusion and cross-modal attention weighted dimensionality reduction of voice, text and image features.

3. The interactive intention recognition and error correction system for payment scenarios according to claim 1 is characterized by: The intent simulation module builds an adversarial sample library covering 12 types of payment fraud scenarios. It uses a statistical model built based on the user's historical transaction frequency, device credibility, and geographic location risk to predict periodic transaction behavior and match user operation characteristics with adversarial samples through cosine similarity.

4. The interactive intention recognition and error correction system for payment scenarios according to claim 1 is characterized by: The dynamic decision module triggers the following verification strategies based on the risk score: for level one risk, voiceprint, iris and touch handwriting pressure authentication is mandatory; for level two risk, a pop-up window is displayed for secondary confirmation and an associated transaction limit reminder is provided; for level three risk, a text message verification code is sent and a correction suggestion is provided if the payee name similarity is ≥90%.

5. The interactive intention recognition and error correction system for payment scenarios according to claim 1 is characterized by: The biometric verification module extracts the Mel-frequency cepstral coefficients of the voiceprint, detects the coherence of natural pronunciation to distinguish forgery attacks, analyzes the dynamic entropy value of touch handwriting pressure to identify mechanical script attacks, and uses a 3D structured light module to capture the three-dimensional characteristics of iris texture and living body vibration frequency.

6. The interactive intention recognition and error correction system for payment scenarios according to claim 1 is characterized by: The audit evidence module builds a consortium chain network through a practical Byzantine fault-tolerant consensus mechanism, deploys compliance review contracts, policy execution contracts and cross-chain verification contracts, and performs transaction compliance verification.

7. The interactive intention recognition and error correction system for payment scenarios according to claim 1, characterized in that: The cross-scenario knowledge transfer module constructs a payment scenario heterogeneous graph based on scenario feature vectors and risk pattern nodes, aggregates high-frequency scenario features to low-frequency scenario nodes through a layered graph attention network, and uses differential privacy technology to synchronize the migrated model parameters to the intention simulation module and the dynamic decision-making module.

Images