Adaptive noise-based financial document anti-reversing verification method, device and medium

Abstract

The application discloses a financial voucher anti-reverse verification method and device based on adaptive noise, and a medium, and belongs to the technical field of financial data security verification. The application fuses structured field features and unstructured visual features of a voucher to splice a joint feature vector. The application discloses a pure sample training model, discards a traditional training end noise mode, combines a device unique physically unclonable feature value and a real-time time stamp only in an inference stage, dynamically generates adaptive noise and superimposes the adaptive noise on a network feature layer to effectively interfere with reverse detection. Then, based on the enhanced feature vector, the authenticity of the voucher is verified, and the verification result, noise parameters and device feature values are stored in an unalterable storage medium. The application realizes two-way consideration of model discrimination accuracy and anti-reverse attack capability, the noise has no fixed rule, the cracking difficulty is extremely high, the protection module and the model training are decoupled, operation and maintenance are convenient, a whole-process storage traceability link is constructed, and the application can efficiently adapt to a high-precision, high-security and traceable verification scene of a financial voucher.

Description

Technical Field

[0001] This invention relates to the field of financial data security verification technology, and in particular to a method, device and medium for reverse verification of financial vouchers based on adaptive noise. Background Technology

[0002] Financial vouchers are crucial primary evidence for corporate financial management and auditing. Verifying the authenticity of vouchers during the entry stage is a vital step in preventing financial fraud and ensuring the security of financial data. Currently, the industry typically verifies either the content of voucher fields or the voucher image separately; these two methods are independent and make comprehensive verification difficult. Furthermore, existing verification models generally use fixed-parameter noise to defend against reverse engineering attacks, resulting in limited and simplistic protection. After verification, the relevant data is often stored using ordinary methods, leading to insufficient data security and reliability.

[0003] Existing technologies have significant drawbacks: verification dimensions are limited, making it impossible to combine field information with visual information for comprehensive judgment, and making it difficult to identify various covert tampering behaviors; model protection measures are rigid, fixed noise is easily cracked, and verification rules are at risk of being reverse-engineered; data storage lacks security guarantees, making it impossible to reliably retain critical information.

[0004] Therefore, there is an urgent need to provide a method that can improve the comprehensiveness of credential verification and the model's resistance to reverse attacks, while ensuring the security and traceability of verification data. Summary of the Invention

[0005] The main objective of this invention is to provide a method, device, and medium for reverse verification of financial vouchers based on adaptive noise, aiming to solve the technical problems in the prior art of low comprehensiveness of financial voucher data verification, poor resistance to reverse attacks of models, and weak traceability of verification data security.

[0006] To achieve the aforementioned objectives, the first aspect of this invention proposes a method for preventing reverse verification of financial documents based on adaptive noise, comprising: Obtain the financial voucher data to be verified, extract the structured field features and unstructured visual features from the financial voucher data, and fuse the two types of features into a joint feature vector; Collect the timestamp of the current verification request and the physical non-cloning feature value of the initiating device. Dynamically generate an amplitude parameter of adaptive noise based on the timestamp and the physical non-cloning feature value. Generate adaptive noise according to the amplitude parameter and add the adaptive noise to the input or output features of at least one network layer of the neural network that processes the joint feature vector to obtain an enhanced feature vector. The authenticity of the financial voucher data is verified based on the enhanced feature vector, and the verification result is output. The verification results, the noise amplitude parameters, and the physically unclonable feature values ​​are stored in an immutable storage medium.

[0007] Furthermore, the structured field features include a numerical string of amounts, a corresponding subject category, and a semantic encoding of the field's summary meaning; The unstructured visual features include the layout information of the voucher and the location area of ​​the seal on the voucher. The fusion into a joint feature vector includes: aligning the structured field features and the unstructured visual features according to a unified dimension and then concatenating them.

[0008] Furthermore, the amplitude parameter for dynamically generating adaptive noise based on the timestamp and the physically non-clonable feature value includes: The timestamp is converted into a binary sequence, the physically unclonable feature value is converted into a binary sequence, a bitwise mixing operation is performed on the two sets of binary sequences, and the operation result is mapped to a noise amplitude parameter.

[0009] Furthermore, the verification of the authenticity of the financial voucher data based on the enhanced feature vector includes: The enhanced feature vectors are classified using a support vector machine, and the local forgery probabilities corresponding to the seal location region and the amount string are output respectively. If the probability of local forgery is greater than the preset forgery threshold, it is determined that local tampering exists, and a verification rejection result is output.

[0010] Furthermore, after classifying the enhanced feature vectors using a support vector machine and outputting the local forgery probabilities corresponding to the seal location region and the amount string, the method further includes: When the local forgery probability is less than or equal to the preset forgery threshold, obtain the historical tampering degree statistics corresponding to the field semantic encoding and page layout information from the historical submission records; The boundary offset of the threshold for the review rule is determined based on the statistical values ​​of the historical degree of tampering. The boundary offset is superimposed on the basic review rules to obtain the judgment conditions of the dynamic review rules; The enhanced feature vector is verified by the discrimination conditions of the dynamic audit rules, and the consistency between the fields covered by the seal position area and the amount number string in the semantics of the voucher is compared to determine the final verification result.

[0011] Furthermore, storing the verification result, the noise amplitude parameter, and the physically unclonable feature value in an immutable storage medium includes: The verification result, noise amplitude parameter, physical non-clonable feature value, discrimination condition and enhanced feature vector are encrypted using a hash algorithm to obtain the device identity record; The device identification records are distributed and written to each evidence storage node in segments according to the time interval corresponding to the timestamp, thereby generating evidence storage data.

[0012] Furthermore, the method also includes a post-audit step, comprising: Upon receiving a post-audit instruction, the device identification record in the stored evidence data is retrieved; Extract the verification value from the record and compare it with the physical non-cloning feature value extracted by the current initiating device to calculate the matching degree between the two; The submitter of the voucher can be traced and identified based on the matching results.

[0013] Further, generating adaptive noise based on the amplitude parameter includes: The joint feature vector is input into the neural network, and the initial confidence score of the authenticity of the certificate is output. The initial confidence score of the authenticity of the certificate is the original discrimination score of the neural network on the joint feature vector under noise disturbance, which is used to dynamically correct the noise amplitude parameter. When the initial confidence level is greater than the first preset threshold, it is determined to be a high-confidence formal certificate, and a negative offset is applied to the noise amplitude parameter to correct it to the first amplitude range; When the initial confidence level is less than or equal to the first preset threshold and greater than or equal to the second preset threshold, it is determined to be a critical confidence level risk certificate. A positive offset is applied to the noise amplitude parameter to correct it to the second amplitude range, wherein the lower limit of the second amplitude range is greater than the upper limit of the first amplitude range. When the initial confidence level is less than the second preset threshold, it is determined to be a low-confidence tampered certificate, and the noise disturbance is terminated or a preset limit amplitude disturbance is applied. Adaptive noise is generated based on the corrected noise amplitude parameters.

[0014] A second aspect of the present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the adaptive noise-based financial document anti-reverse verification method described in any of the preceding claims.

[0015] A third aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the adaptive noise-based financial document anti-reverse verification method described in any of the preceding claims.

[0016] The adaptive noise-based method, device, and medium for preventing reverse verification of financial documents of the present invention have significant substantial improvements and multiple technical advantages compared with the prior art. The specific beneficial effects are as follows: First, this invention integrates the structured field features and unstructured visual features of financial vouchers to construct a joint feature vector, while covering both standardized business field information and visual layout information of the voucher. This overcomes the shortcomings of traditional single-dimensional verification, which is prone to omissions and hidden tampering. It can comprehensively characterize the true attributes of the voucher and effectively identify various forgery behaviors such as field tampering, layout alteration, and abnormal seals, significantly improving the coverage and recognition accuracy of financial voucher verification.

[0017] Second, this invention adopts an innovative architecture that decouples model training and security protection. The model is trained using clean samples to learn accurate and stable authentication logic and retain optimal verification accuracy. Only a small amount of controllable adaptive noise is applied during the inference and verification stage, which only perturbs the model input features and does not change the core authentication features of the voucher. Under the premise of fully ensuring the accuracy of voucher verification, model security protection is achieved, which completely avoids the drawbacks of traditional noise-adding schemes at the training end that sacrifice authentication accuracy and smooth authentication boundaries.

[0018] Third, this invention dynamically generates noise amplitude parameters based on both the device's physically unclonable characteristics and real-time dynamic timestamps. This ensures that the noise parameters for each verification request are unique to the device, dynamic in time, and do not follow a fixed pattern of change. Attackers cannot crack the noise logic and model discrimination rules through reverse engineering methods such as batch sampling, function fitting, and gradient backtracking. This fundamentally defends against reverse engineering and model parameter theft attacks, significantly improving the security level of the credential verification system.

[0019] Fourth, the noise protection logic of this invention is independent of the model training process. Model iteration updates and verification rule optimization do not require synchronous adjustment of the noise strategy. The module has low coupling and strong scalability, which effectively reduces the difficulty of system development and later operation and maintenance, and is suitable for long-term enterprise use.

[0020] Fifth, this invention stores core data such as verification results, noise amplitude parameters, and physical non-clonable characteristics of equipment into an immutable storage medium, completely preserving the equipment information, protection parameters, and verification results of each verification, forming a closed-loop trusted evidence storage link. This provides reliable data support for subsequent financial audits, risk tracing, and liability determination, meeting the application requirements of high compliance and high credibility in financial scenarios. Attached Figure Description

[0021] Figure 1 A flowchart illustrating an embodiment of an adaptive noise-based method for preventing reverse verification of financial documents. Figure 2This is a schematic block diagram of a computer device according to an embodiment of the invention. The realization of the object, functional features, and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0023] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this specification means the presence of features, integers, steps, operations, elements, modules, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, modules, components, and / or groups thereof. It should be understood that when we say an element is “connected” or “coupled” to another element, it can be directly connected or coupled to the other element, or there may be intermediate elements. Furthermore, “connected” or “coupled” as used herein can include wireless connections or wireless coupling. The term “and / or” as used herein includes all or any modules and all combinations of one or more associated listed items.

[0024] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.

[0025] Reference Figure 1 This invention provides a method for preventing reverse verification of financial documents based on adaptive noise, comprising the following steps: S1: Obtain the financial voucher data to be verified, extract the structured field features and unstructured visual features from the financial voucher data, and fuse the two types of features into a joint feature vector.

[0026] Structured field features are standardized text and numeric field data in financial vouchers with fixed formats and quantifiable input; unstructured visual features are visual information in financial vouchers that do not have fixed data formats and rely on image presentation; the joint feature vector is a one-dimensional feature array formed by integrating multiple types of features according to data rules, and is the standard input data for machine learning models.

[0027] This embodiment uses enterprise VAT reimbursement vouchers as the verification object. The system first reads the complete electronic data of the voucher and extracts structured field features such as amount, accounting subject, and summary from the voucher data. At the same time, it extracts unstructured visual features such as voucher layout, seal position, and border style. For example, the extracted structured field features are: amount 5680 yuan, subject "office expenses", and summary "purchase of office supplies"; the extracted unstructured visual features are: standard voucher format and the financial seal located in the lower right corner of the voucher. The above two types of features are concatenated according to data dimension rules to generate a joint feature vector. This step is used to complete the full-dimensional feature representation of the voucher. The joint feature vector fully carries the business attributes and visual attributes of the voucher, breaking the limitations of traditional single-dimensional verification and providing a complete and comprehensive data source for subsequent accurate authenticity judgment and model security protection. This step integrates the two core types of voucher information, text and visual, to achieve multi-dimensional feature collection, effectively covering various forms of forgery such as field tampering, seal displacement, and layout alteration, and improving the comprehensiveness of subsequent verification.

[0028] S2: Collect the timestamp of the current verification request and the physical non-cloning feature value of the initiating device, dynamically generate the amplitude parameter of adaptive noise based on the timestamp and the physical non-cloning feature value, generate adaptive noise according to the amplitude parameter, and add the adaptive noise to the input feature or output feature of at least one network layer of the neural network that processes the joint feature vector to obtain the enhanced feature vector.

[0029] The timestamp is a time-series code used to uniquely identify the time of data generation, possessing the characteristic of real-time dynamic change; the Physically Unclonable Function (PUF) is a unique and uncopyable hardware feature code generated by the hardware device based on physical differences in the chip manufacturing process. Each device's feature value is independent and cannot be artificially replicated. The Physically Unclonable Function can be obtained in the following ways: the initiating device has a built-in SRAM (Static Random-Access Memory) type PUF module. During verification, the system sends a challenge value to the device, and the PUF module generates a unique response value based on random doping differences in the chip manufacturing process; or, the device uses an arbiter PUF, generating an unclonable bit string by measuring the delay differences of multiple paths. This embodiment uses an SRAM PUF, where the system reads the random initial state of the SRAM cells when the device starts up as the Physically Unclonable Function. Adaptive noise is perturbation data where parameters dynamically change with external conditions; the neural network is a deep learning network model used for feature depth mining and computational processing. In this invention, the neural network is trained using clean samples, learning pure credential authenticity discrimination logic. In this embodiment, the neural network employs a three-layer fully connected network with an input layer dimension of 32, hidden layers dimensions of 64 and 32, and an output layer dimension of 2 (real / forged). The network uses the ReLU (Rectified Linear Unit) activation function and the cross-entropy loss function, and is trained using the Adam (Adaptive Moment Estimation) optimizer. The training data consists of 100,000 labeled images of real and forged financial documents.

[0030] The system captures the timestamp corresponding to this credential verification request and reads the hardware physical non-cloning feature value of the submitting device. Based on these two sets of data, it calculates a noise amplitude parameter, generates corresponding noise according to the noise amplitude parameter, and superimposes the noise onto the input or output features of the neural network convolutional layer to complete feature enhancement. For example, if the timestamp of this request is 1747952136 and the device's physical non-cloning feature value is 0x2F7A91C3, the noise amplitude parameter 0.18 is obtained through calculation using these two sets of data. Corresponding Gaussian adaptive noise is then generated and superimposed onto the first layer input features of the neural network to finally obtain the enhanced feature vector.

[0031] This step employs a design approach that decouples training from noise. The neural network is trained entirely using clean, noise-free samples, fully learning the correspondence between valid features of the credential and its authenticity, thus forming a precise and stable discrimination boundary. Noise is applied only during the model inference and verification phase, using small-range adaptive noise with controllable amplitude. This perturbation is strictly limited to a reasonable range, causing only a minor disturbance to the joint feature vector. It does not change the core feature distribution, ranking, or differences, and only slightly shifts the discrimination score output by the model, insufficient to overturn the authenticity determination result. The noise only serves to interfere with external reverse detection and does not destroy the valid feature information required for credential authenticity determination, fundamentally avoiding the shortcomings of existing technologies where noise added during the training phase leads to a decrease in recognition accuracy.

[0032] Meanwhile, this step abandons the traditional fixed-amplitude noise scheme. The noise amplitude parameter is generated by combining the device's unique, physically unclonable feature value with a real-time updated timestamp. On the one hand, the physically unclonable feature value is bound to each hardware device, and the noise parameters generated by different devices are different. On the other hand, the same device initiates verification at different times, and the timestamp changes continuously, so the corresponding noise parameters also change. This ensures that the noise parameters for a single request and a single device are completely unique and have no fixed change pattern. Attackers cannot fit the noise function by collecting a large number of input and output samples, nor can they remove noise interference and reverse-derive neural network weights and verification rules through gradient backtracking, regression analysis, etc. Compared with existing fixed noise and training-end noise-adding schemes, the resistance to reverse attacks is significantly improved. In addition, the noise logic is independent of the model training process, and there is no need to adjust the noise parameters when the model is iterated or the rules are updated, which greatly reduces the difficulty of system development and maintenance.

[0033] This step, on the one hand, relies on the unique characteristics of the hardware and time-series data to generate dynamic noise, thereby blocking the risk of the model being reverse-engineered from the root and ensuring the security of the verification rules and model parameters; on the other hand, it adopts an architecture of independent noise addition at the inference end and pure sample learning at the training end, which balances high recognition accuracy and strong security protection capabilities, while simplifying system operation and maintenance.

[0034] S3: Verify the authenticity of the financial voucher data based on the enhanced feature vector and output the verification result.

[0035] The enhanced feature vector obtained in the previous step is input into the verification model. The model performs the verification of the voucher's authenticity according to the preset discrimination logic and outputs a clear result. For example, the model analyzes the enhanced feature vector, relies on the stable discrimination logic learned during the training phase, filters out score shifts caused by small noises, compares the matching degree between the features and standard voucher samples, determines that the VAT reimbursement voucher has not been tampered with, and outputs a "verification passed" result. The neural network forms a mature discrimination capability based on clean samples, and minor noises in the inference phase will not affect the core judgment logic. This step relies on the security-hardened features to complete the verification, ensuring the accuracy and reliability of the voucher authenticity determination result while resisting external attacks.

[0036] S4: Store the verification result, the noise amplitude parameter, and the physically unclonable feature value in an immutable storage medium.

[0037] Immutable storage media are storage carriers built using technologies such as blockchain and read-only distributed ledgers, where data cannot be deleted or tampered with once written.

[0038] The system collects all key data generated during this verification process and stores it on an immutable storage medium. For example, the "verification passed" result, noise amplitude parameter 0.18, and the device's physical non-cloning characteristic value 0x2F7A91C3 are all stored in the blockchain evidence storage node. This step simultaneously retains the verification results, noise parameters, and the device's unique identifier, completely recording the entire verification process. This ensures the permanent and secure retention of key data, constructs a complete evidence chain, and provides reliable data support for subsequent audits, investigations, problem tracing, and liability determination.

[0039] In this embodiment, a complete closed-loop process is constructed, consisting of "multi-dimensional feature fusion - dynamic noise protection at the inference end - high-precision authenticity verification - tamper-proof data storage". First, this invention breaks through conventional technical approaches in the field by decoupling model training from reverse engineering noise protection. Existing technologies generally use noise as a data augmentation method during the training phase, inevitably resulting in smoothed model discrimination boundaries and compromised recognition accuracy. This invention creatively applies noise only during the inference verification phase, using clean samples during the training phase to allow the model to learn the optimal discrimination logic. This preserves the high-precision recognition advantage of deep learning models while utilizing noise to achieve the security goal of resisting reverse engineering, resolving the core technical contradiction of "mutual constraint between protection capability and recognition accuracy" in existing technologies. Furthermore, this invention abandons the industry-standard fixed-amplitude noise and global random noise patterns, combining the device's physically unclonable feature values ​​and real-time timestamps as dual dynamic factors to generate adaptive noise. The uniqueness of device hardware characteristics and the real-time dynamic nature of timestamps ensure that the noise parameters of each verification request are unique and unpredictable, completely eliminating the path for attackers to steal model parameters and verification rules through sample fitting and reverse engineering. This anti-attack effect is far superior to traditional solutions. Simultaneously, the noise module is deployed independently, decoupled from the model training process, significantly reducing the complexity of system iteration and maintenance. Finally, this invention integrates structured field features of vouchers with unstructured visual features for joint verification, overcoming the limitations of a single verification dimension and accurately identifying various covert fraudulent behaviors. The backend uses tamper-proof media to store key data throughout the entire process, meeting the compliance requirements of auditing, supervision, and accountability in the financial industry.

[0040] In summary, the adaptive noise-based anti-reverse verification method for financial vouchers in this embodiment integrates text fields and visual features of the voucher for joint judgment, effectively identifying various forgery behaviors such as field tampering, seal displacement, and layout alteration. The model is trained on clean samples, and minor noise at the inference end does not affect the judgment results. Compared to noise-adding schemes at the training end, this significantly reduces the probability of false positives and false negatives, meeting the high-accuracy verification requirements of financial vouchers. The decoupled noise-adding architecture of training and inference, combined with dynamic noise jointly generated by hardware features and time-series data, produces noise with no fixed pattern, making it impossible for attackers to reverse engineer the model parameters and verification rules. This fundamentally avoids the risk of model breaches, significantly improving security. The noise logic is independent of the model training process; model updates and rule iterations do not require synchronous adjustment of noise parameters, reducing development and maintenance costs and facilitating long-term system deployment and functional expansion. Verification results, noise parameters, and unique device characteristics are uniformly stored in an immutable storage medium, preventing data deletion and tampering, forming a complete and reliable traceability chain, fully meeting the compliance requirements of corporate financial audits and external supervision.

[0041] In one embodiment, the structured field features described above include a string of numerical amounts, a corresponding subject category, and a semantic encoding of the field's meaning.

[0042] Semantic field encoding involves converting textual categories and summary content into machine-recognizable numerical codes according to semantic rules. For example, in the aforementioned VAT reimbursement voucher, the core text fields are split and encoded: the amount string is 5680; the category "office expenses" is semantically encoded as 12; and the summary "purchase of office supplies" is semantically encoded as 27. Converting natural language text into standardized codes adapts to vector operation rules, improving feature processing efficiency.

[0043] The aforementioned unstructured visual features include the layout information of the voucher and the location of the seal on the voucher.

[0044] Extract fixed layout data and seal coordinate information from the voucher image as core visual features. For example, the layout is a standard, general reimbursement voucher format; the seal location area coordinates are (720, 510) to (860, 630). Accurately locate the core visual identifiers of the voucher to achieve targeted identification of forgery behaviors such as seal misplacement and layout alteration.

[0045] The above-mentioned fusion into a joint feature vector includes: aligning the structured field features and the unstructured visual features according to a unified dimension and then concatenating them.

[0046] Scale alignment normalizes two types of feature data with different dimensions and scales, ensuring consistency in data value range and dimensional length. Normalized scale alignment is performed on the encoded field features and coordinate-based visual features, then they are concatenated sequentially into a complete vector. For example, unifying the field feature dimension to 16 dimensions and the visual feature dimension to 16 dimensions results in a 32-dimensional joint feature vector after alignment. This eliminates data differences between different feature types, ensuring effective feature fusion and stable model operation.

[0047] This embodiment discloses refined feature extraction and fusion rules, making the feature composition more closely aligned with the business attributes of financial vouchers, standardizing the feature fusion method, and improving the stability and verification accuracy of subsequent model operations.

[0048] In one embodiment, the amplitude parameter for dynamically generating adaptive noise based on the timestamp and the physically non-clonable eigenvalue includes: S201: Convert the timestamp into a binary sequence and the physically unclonable feature value into a binary sequence; Continuing with the previous example of expense reimbursement vouchers, the decimal timestamps and hexadecimal physically unclonable characteristic values ​​are uniformly converted into equal-length binary sequences. For example, the timestamp 1747952136 is converted into a 32-bit binary sequence T = 1101000110111001010000011000000; the physically unclonable characteristic value 0x2F7A91C3 is converted into a 32-bit binary sequence P = 00101111011110101001000111000011. This unified data format provides a foundation for subsequent bitwise operations.

[0049] S202: Perform a bitwise mixing operation on two sets of binary sequences and map the result to a noise amplitude parameter.

[0050] Bitwise mixed operations involve performing logical operations on each bit of a binary sequence of the same length. This embodiment uses the following operations and mapping method to obtain the noise amplitude parameter, assuming T... i Let P be the value of the i-th bit of the timestamp binary sequence. i Let be the value of the i-th bit of the binary sequence of physically non-clonable eigenvalues, n be the total number of bits in the binary sequence, and A be the noise amplitude parameter. The calculation formula is:

[0051] In the formula, This represents a bitwise XOR operation, where α is a normalization coefficient within the interval [0.05, 0.3], used to map the operation result to a reasonable noise amplitude range. For example, with a sequence bit length n=32 and a normalization coefficient α=0.2, the noise amplitude parameter A=0.18 is calculated by summing the bits of T and P after bitwise XOR and substituting the sum into the formula. By combining hardware features with timing data for joint operations, each verification request generates a unique noise amplitude parameter, which has no fixed pattern. This embodiment uses a bitwise XOR operation, but in other implementations, bitwise AND, bitwise OR, or bitwise NAND operations can also be used, as long as the two binary sequences can be mixed into a single value.

[0052] In this embodiment, the noise amplitude parameter is generated entirely based on dynamic timing and the unique hardware characteristics of the device. The noise amplitude parameter changes in real time with the device and time, and there is no fixed pattern. Attackers cannot crack the noise rules through sample fitting, which greatly enhances the model's resistance to reverse engineering attacks. The formula calculation logic is simple and the calculation efficiency is high.

[0053] In one embodiment, the above-mentioned verification of the authenticity of the financial voucher data based on the enhanced feature vector includes: S301: Use a support vector machine to classify the enhanced feature vectors and output the local forgery probability corresponding to the seal location region and the amount string, respectively.

[0054] Support Vector Machines (SVMs) are machine learning classification models suitable for small sample sizes and high-dimensional feature classification. Local forgery probability represents the probability of tampering in a local area of ​​a document, with a value range of [0,1]. The enhanced feature vector, formed by adding noise to the joint feature vector obtained in the above embodiments, is input into the SVM to calculate the forgery probability by region. For example, with a 32-dimensional enhanced feature vector as input, the model outputs: a forgery probability of 0.06 for the amount string and a forgery probability of 0.04 for the seal location area. This precise regional forgery risk assessment, locating suspected tampering locations, differs from a general overall assessment.

[0055] S302: If the probability of local forgery is greater than the preset forgery threshold, it is determined that there is local tampering, and a verification rejection result is output.

[0056] A fixed forgery threshold is preset. The calculated local forgery probability is compared with the threshold, and the corresponding judgment logic is executed. For example, if the preset forgery threshold is 0.3, and the probabilities of forgery in both instances are less than the threshold, tampering is not judged at this time; however, if the probability of forgery of a certain voucher amount is 0.42, which is greater than the threshold of 0.3, then local tampering is judged, and "Verification Rejected" is output. The rules are clear and the judgment speed is fast, which can quickly intercept obviously tampered vouchers. Furthermore, the preset forgery threshold can be set according to the business scenario. For example, for high-risk vouchers (such as large expense reimbursements), the threshold can be set to 0.2; for low-risk vouchers, it can be set to 0.4. This embodiment uses 0.3 as a general threshold.

[0057] In this embodiment, a support vector machine is used to perform probabilistic judgment by region, and a preliminary authenticity screening is completed in combination with preset thresholds, focusing on detecting the two high-frequency tampering areas: amount and seal. Targeted detection is carried out by focusing on the core easily tampered areas of financial documents. The screening logic is simple and efficient, which can quickly intercept explicit tampering of documents, reduce the pressure of subsequent verification, and improve the overall verification efficiency.

[0058] In one embodiment, after classifying the enhanced feature vectors using a support vector machine and outputting the local forgery probabilities corresponding to the seal location region and the amount string, the method further includes: S303: When the local forgery probability is less than or equal to the preset forgery threshold, obtain the historical tampering degree statistics corresponding to the field semantic encoding and layout information from the historical submission records.

[0059] The historical tampering severity statistics are a comprehensive value derived from historical voucher data, representing the frequency and severity of historical forgery for a specific field and format. When the probability does not exceed a threshold, historical tampering statistics for that type of voucher are retrieved. For example, if the voucher item is "office expenses" and the format is a general reimbursement format, retrieving historical records yields a tampering severity statistics value of 0.12. The calculation method for the historical tampering severity statistics is as follows: Count the number of vouchers in the past N submissions (e.g., N=1000) where the field semantic code is a specific value (e.g., 'office expenses') and was ultimately determined to be forged; divide this number by the total number of submissions to obtain the forgery rate; simultaneously, combine this with a weighted average of tampering severity scores (e.g., percentage deviation of amount) to obtain the comprehensive tampering severity statistics value. Specifically, let the historical forgery rate for a certain field type be r, and the average tampering severity be s, then the tampering severity statistics value V = λr + βs, where λ and β are preset weights. Combining historical forgery patterns provides data support for dynamically adjusting audit standards.

[0060] S304: Determine the boundary offset of the audit rule discrimination threshold based on the historical tampering statistics.

[0061] Threshold offsets are generated based on the statistical values; the higher the degree of tampering, the larger the offset. For example, a threshold offset of 0.05 is calculated from a statistical value of 0.12. This allows the review standards to adaptively adjust based on historical fraud cases.

[0062] S305: The boundary offset is superimposed on the basic audit rules to obtain the judgment conditions of the dynamic audit rules.

[0063] The offset is added to the basic semantic verification rules to update the verification judgment conditions. For example, an offset of 0.05 is added to the basic semantic rules to form dynamic judgment conditions. This breaks the traditional fixed review rules and improves the ability to identify new and covert tampering.

[0064] S306: The enhanced feature vector is verified by the discrimination conditions of the dynamic audit rules, and the consistency between the field covered by the seal position area and the amount number string in the semantics of the voucher is compared to determine the final verification result.

[0065] Based on dynamic rules, the system validates the semantic logic of text and images to determine whether the seal, amount, and account conform to financial business logic. For example, if the validation finds that the seal-covered field matches the semantics of the amount and account, the system determines that the voucher is completely authentic and outputs the final "verification passed" result. This achieves linked semantic validation of text and images, identifying hidden behaviors such as logical forgery.

[0066] In this embodiment, historically tampered data is used to dynamically adjust the audit rules, and a second deep verification of semantic consistency is conducted. The audit standards are dynamically updated based on historical forgery patterns, and hidden tampering that cannot be detected by single-probability detection is identified through image-text semantic linkage verification, significantly improving the accuracy of verification in complex forgery scenarios.

[0067] In one embodiment, storing the verification result, the noise amplitude parameter, and the physically non-clonable characteristic value in an immutable storage medium includes: S401: The verification result, noise amplitude parameter, physical non-cloning feature value, discrimination condition and enhanced feature vector are encrypted using a hash algorithm to obtain the device identity record.

[0068] Hash algorithms are one-way encryption algorithms; even minor changes to the original data can lead to drastic changes in the hash result, making them useful for data integrity verification. Core data across the entire process is collected and uniformly hashed to generate unique identifier records. For example, hash operations are performed on "verification passed," amplitude parameter 0.18, the device's physical non-cloning characteristic value, dynamic discrimination conditions, and enhanced feature vectors to generate unique device identity records. This encryption solidifies all process data, preventing data tampering and enabling rapid identification of data tampering attempts.

[0069] S402: Distribute the device identity record and write it to each evidence storage node in segments according to the time interval corresponding to the timestamp to generate evidence storage data.

[0070] The encrypted records are split and distributed to different distributed storage nodes according to time intervals. For example, based on the time period to which the timestamp belongs, the record is written to the three storage nodes of the same day's shard, generating complete storage data. Specifically: taking each hour as a time interval, all storage data within the hour of the timestamp is packaged into a data block, and the Raft consensus algorithm (a distributed consensus algorithm) is used to write this data block to the three storage nodes simultaneously; for verification requests that span time intervals, they are written according to the interval to which the starting timestamp belongs. Distributed sharded storage avoids data loss due to single points of failure, further improving storage reliability.

[0071] In this embodiment, the evidence storage data covers all dimensions of verification results, model parameters, device characteristics, verification rules, and feature data. Combined with hash encryption and distributed storage, it thoroughly ensures that the evidence storage data is tamper-proof and not easily lost, providing a complete chain of evidence for audit tracing.

[0072] In one embodiment, the above-described adaptive noise-based method for preventing reverse verification of financial documents further includes a post-audit step, including: S5: Upon receiving a post-audit instruction, retrieve the device identification record from the stored evidence data.

[0073] When auditors issue an audit instruction, the corresponding records are retrieved from the distributed evidence storage nodes. For example, upon receiving an audit instruction, the previously generated device identification record is retrieved based on the voucher number. This allows for rapid retrieval of historical evidence storage data to support the audit process.

[0074] S6: Extract the check value from the record and compare it with the physical non-cloning feature value extracted by the current initiating device to calculate the matching degree between the two.

[0075] The hardware verification value is extracted from the evidence storage record and compared with the physical non-clonable feature value read in real time by the current device. For example, the verification value extracted from the evidence storage is compared with the physical non-clonable feature value of the current device, and the matching degree is calculated to be 100%. Accurate identity comparison is achieved based on the unique hardware feature of the device.

[0076] S7: Based on the matching results, trace back to determine the submitter of the voucher.

[0077] The consistency of equipment is determined by the degree of matching, thereby identifying the entity that submitted the voucher. If the matching degree is perfect, it is determined that the voucher was submitted by the operator of the corresponding equipment. This enables precise identification of responsibility and facilitates audit accountability.

[0078] This embodiment relies on the unique characteristics of hardware and the tamper-proof evidence data to achieve traceability of the audit process and identification of responsibility, fully meeting the compliance requirements of corporate financial supervision and external audit, and making up for the shortcomings of insufficient traceability capabilities of traditional solutions.

[0079] In one embodiment, generating adaptive noise based on the amplitude parameter includes: S211: Input the joint feature vector into the neural network and output the initial confidence score of the authenticity of the certificate. The initial confidence score of the authenticity of the certificate is the original discrimination score of the neural network on the joint feature vector under noise disturbance, which is used to dynamically correct the noise amplitude parameter.

[0080] In this step, the neural network is a discrimination model trained entirely on clean, noise-free samples without any data augmentation. The model learns pure financial document authenticity discrimination logic and the optimal discrimination boundary, avoiding boundary smoothing issues caused by training noise. The initial confidence score for document authenticity is the original discrimination score output by the neural network under the input of the original joint feature vector without any noise perturbation, with a value range of [0,1]. This initial confidence score is only used to dynamically adjust the noise amplitude parameter and is not used as the final document verification result output.

[0081] Using the example of the aforementioned enterprise VAT reimbursement voucher, after generating the basic noise amplitude parameter (0.18) in step S2 and before adding adaptive noise in step S2, the system inputs the 32-dimensional original joint feature vector (without noise perturbation) obtained in step S1 into the same neural network, and the network outputs the initial confidence level of the voucher. For example, for a travel expense reimbursement voucher, after inputting the noiseless joint feature vector, the neural network outputs an initial confidence level of 0.94, indicating that the voucher features are regular and there are no significant traces of tampering, belonging to a high-confidence sample; for a reimbursement voucher with a slight deviation in layout, the output initial confidence level is 0.78, indicating that its features are near the boundary of authenticity and there is a risk of ambiguity; for a voucher with an obviously abnormal amount string, the output initial confidence level is 0.62, indicating that it is a high-risk tampering sample.

[0082] This step relies on a clean model to output a true and interference-free confidence level for voucher risk, providing accurate data for subsequent noise grading adjustments and achieving a precise match between noise protection intensity and voucher risk level.

[0083] S212: When the initial confidence level is greater than the first preset threshold, it is determined to be a high-confidence legitimate certificate, and a negative offset is applied to the noise amplitude parameter to correct it to the first amplitude range; when the initial confidence level is less than or equal to the first preset threshold and greater than or equal to the second preset threshold, it is determined to be a critical confidence risk certificate, and a positive offset is applied to the noise amplitude parameter to correct it to the second amplitude range, wherein the lower limit of the second amplitude range is greater than the upper limit of the first amplitude range; when the initial confidence level is less than the second preset threshold, it is determined to be a low-confidence tampered certificate, and the noise disturbance is terminated or a preset limit amplitude disturbance is applied.

[0084] In this step, two confidence thresholds are preset to divide the initial confidence level into three risk intervals, adapting to differentiated noise adjustment strategies. In this embodiment, the first threshold (high confidence threshold) is preset to 0.9 and the second threshold (low confidence threshold) is preset to 0.7, which can be fine-tuned according to actual business scenarios.

[0085] The specific division rules are as follows: High-confidence legitimate samples: When the initial confidence level is >0.9, it is judged to be a compliant certificate with no obvious traces of tampering and regular features, and the overall risk of forgery is extremely low; Critical fuzzy risk samples: When 0.7 ≤ initial confidence level ≤ 0.9, they are judged as fuzzy credentials whose features are on the boundary between true and false and have slight format or field deviations. Such samples are the key attack targets for attackers to reverse probe and fit the model's discrimination rules by submitting a large number of critical samples. Low-confidence tampered samples: When the initial confidence level is <0.7, it is judged as a high-risk credential with abnormal features and obvious tampering traces, and no refined anti-reverse protection is required.

[0086] For example, travel expense vouchers with a confidence level of 0.94 are classified as high-confidence regular samples; reimbursement vouchers with a confidence level of 0.78 are classified as borderline fuzzy risk samples; and abnormal vouchers with a confidence level of 0.62 are classified as low-confidence tampered samples.

[0087] The basic noise amplitude parameter (0.18 in this example) generated by the timestamp and the physically non-clonable feature value is combined with the confidence offset coefficient corresponding to the sample risk level to dynamically correct the noise amplitude used for feature perturbation, so as to realize the adaptive adjustment logic of "the lower the risk, the weaker the perturbation, and the higher the risk, the stronger the perturbation".

[0088] The specific correction method is as follows: High-confidence regular samples: A negative offset coefficient is superimposed to constrain the final noise amplitude to a low amplitude range of 0.05~0.1 (the first amplitude range). In this embodiment, the basic amplitude of 0.18 is corrected to 0.09, and a very slight feature perturbation is applied to achieve only basic anti-reverse protection, preserving the original valid features of the voucher to the maximum extent and ensuring that the verification accuracy is not affected.

[0089] Critical fuzzy risk samples: A large positive offset coefficient is superimposed to increase the final noise amplitude to a high amplitude range of 0.2~0.3 (the second amplitude range). In this embodiment, the base amplitude of 0.18 is corrected to 0.27, strong perturbation noise is applied, and the feature output distribution of the critical samples is completely disrupted, thus destroying the attacker's sample fitting and gradient backtracking conditions.

[0090] Low-confidence tampering samples: The noise perturbation process is directly terminated (i.e., no noise is added), or a preset maximum amplitude limit perturbation (e.g., high-intensity random noise with an amplitude of 1.0) is applied to block the attack path of attackers using tampered samples to collect model data. This embodiment adopts the noise perturbation termination method to save unnecessary computing power.

[0091] This step completely solves the technical drawbacks of traditional fixed amplitude noise addition, and realizes intelligent adaptive adjustment of noise protection intensity, taking into account both the verification accuracy of regular samples and the strong protection requirements of critical high-risk samples.

[0092] S213: Generate adaptive noise based on the corrected noise amplitude parameters.

[0093] Dynamically corrected adaptive noise is added to the input or output features of at least one layer of the neural network to generate an enhanced feature vector. This enhanced feature vector is then input into a verification model, which uses stable discrimination logic learned through pure training to perform the final verification.

[0094] Differential perturbations of samples at different levels only interfere with external reverse engineering data and do not damage the core valid features of the credentials, thus ensuring stable verification accuracy. For example, for high-confidence credentials corrected to weak noise of 0.09, the final discrimination score output by the model is basically consistent with the initial confidence level (0.94), and it is judged as "verification passed"; for critical samples corrected to strong noise of 0.27, the model output score deviates from the initial confidence level (0.78) by about 0.05~0.1, but still falls stably on the true side of the true / false discrimination boundary (e.g., the final score is 0.72), and it is still judged as "verification passed"; for low-confidence samples, due to the termination of noise addition, the model directly judges as "verification rejected" based on the original features. This step maximizes the model's resistance to reverse engineering attacks in high-risk scenarios without sacrificing the accuracy of business verification, achieving the optimal balance between accuracy, computing power, and security protection.

[0095] In this embodiment, the limitations of fixed-amplitude noise addition are overcome, enabling dynamic adaptation between the risk level of the voucher and the noise protection strength. This solves the inherent defect of traditional solutions where "protection accuracy and security strength cannot be dynamically balanced." Noise perturbation is strengthened for the most vulnerable critical discrimination boundary samples, preventing attackers from submitting a large number of critical samples to fit the model's discrimination boundary and steal model parameters and verification rules, resulting in an exponential improvement in protection effectiveness. Protection is weakened for highly reliable samples and terminated for invalid risk samples, reducing unnecessary computational power consumption, improving the overall system operating efficiency, and adapting to large-scale batch financial voucher verification scenarios.

[0096] Reference Figure 2 The present invention also provides a computer device, the internal structure of which can be as follows: Figure 2 As shown. The computer device includes a processor, memory, network interface, and database connected via a system bus. The processor is designed to provide computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores operating devices, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database stores financial document data, etc. The network interface is used to communicate with external terminals via a network connection. Furthermore, the computer device may also include input devices and a display screen. When the computer program is executed by the processor, it implements the adaptive noise-based reverse verification method for financial documents described in any of the above embodiments. Those skilled in the art will understand that... Figure 2 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer equipment on which the present application is applied.

[0097] One embodiment of this application also provides a computer-readable storage medium storing a computer program thereon. When the computer program is executed by a processor, it implements the adaptive noise-based reverse verification method for financial documents described in any of the above embodiments. It is understood that the computer-readable storage medium in this embodiment can be a volatile readable storage medium or a non-volatile readable storage medium.

[0098] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in this application and in the embodiments can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual-speed SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0099] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, apparatus, article, or method. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.

[0100] The above description is merely a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for preventing reverse verification of financial documents based on adaptive noise, characterized in that, include: Obtain the financial voucher data to be verified, extract the structured field features and unstructured visual features from the financial voucher data, and fuse the two types of features into a joint feature vector; Collect the timestamp of the current verification request and the physical non-cloning feature value of the initiating device. Dynamically generate an amplitude parameter of adaptive noise based on the timestamp and the physical non-cloning feature value. Generate adaptive noise according to the amplitude parameter and add the adaptive noise to the input or output features of at least one network layer of the neural network that processes the joint feature vector to obtain an enhanced feature vector. The authenticity of the financial voucher data is verified based on the enhanced feature vector, and the verification result is output. The verification results, the noise amplitude parameters, and the physically unclonable characteristic values ​​are stored in an immutable storage medium. The structured field features include a numerical string of amounts, corresponding subject categories, and semantic encoding of the field meaning; The unstructured visual features include the layout information of the voucher and the location area of ​​the seal on the voucher. The fusion into a joint feature vector includes: aligning the structured field features and the unstructured visual features according to a unified dimension and then concatenating them; The verification of the authenticity of the financial voucher data based on the enhanced feature vector includes: The enhanced feature vectors are classified using a support vector machine, and the local forgery probabilities corresponding to the seal location region and the amount string are output respectively. If the probability of local forgery is greater than the preset forgery threshold, it is determined that local tampering exists, and a verification rejection result is output. The step of generating adaptive noise based on the amplitude parameter includes: The joint feature vector is input into the neural network, and the initial confidence score of the authenticity of the certificate is output. The initial confidence score of the authenticity of the certificate is the original discrimination score of the neural network on the joint feature vector under noise disturbance, which is used to dynamically correct the noise amplitude parameter. When the initial confidence level is greater than the first preset threshold, it is determined to be a high-confidence formal certificate, and a negative offset is applied to the noise amplitude parameter to correct it to the first amplitude range; When the initial confidence level is less than or equal to the first preset threshold and greater than or equal to the second preset threshold, it is determined to be a critical confidence level risk certificate. A positive offset is applied to the noise amplitude parameter to correct it to the second amplitude range, wherein the lower limit of the second amplitude range is greater than the upper limit of the first amplitude range. When the initial confidence level is less than the second preset threshold, it is determined to be a low-confidence tampered certificate, and the noise disturbance is terminated or a preset limit amplitude disturbance is applied. Adaptive noise is generated based on the corrected noise amplitude parameters.

2. The method according to claim 1, characterized in that, The amplitude parameters for dynamically generating adaptive noise based on the timestamp and the physically non-clonable eigenvalue include: The timestamp is converted into a binary sequence, the physically unclonable feature value is converted into a binary sequence, a bitwise mixing operation is performed on the two sets of binary sequences, and the operation result is mapped to a noise amplitude parameter.

3. The method according to claim 1, characterized in that, After classifying the enhanced feature vectors using a support vector machine and outputting the local forgery probabilities corresponding to the seal location region and the amount string, the method further includes: When the local forgery probability is less than or equal to the preset forgery threshold, obtain the historical tampering degree statistics corresponding to the field semantic encoding and page layout information from the historical submission records; The boundary offset of the threshold for the review rule is determined based on the statistical values ​​of the historical degree of tampering. The boundary offset is superimposed on the basic review rules to obtain the judgment conditions of the dynamic review rules; The enhanced feature vector is verified by the discrimination conditions of the dynamic audit rules, and the consistency between the fields covered by the seal position area and the amount number string in the semantics of the voucher is compared to determine the final verification result.

4. The method according to claim 3, characterized in that, The step of storing the verification result, the noise amplitude parameter, and the physically non-clonable feature value into an immutable storage medium includes: The verification result, noise amplitude parameter, physical non-clonable feature value, discrimination condition and enhanced feature vector are encrypted using a hash algorithm to obtain the device identity record; The device identification records are distributed and written to each evidence storage node in segments according to the time interval corresponding to the timestamp, thereby generating evidence storage data.

5. The method according to claim 4, characterized in that, The method also includes a post-audit step, including: Upon receiving a post-audit instruction, the device identification record in the stored evidence data is retrieved; Extract the verification value from the record and compare it with the physical non-cloning feature value extracted by the current initiating device to calculate the matching degree between the two; The submitter of the voucher can be traced back and identified based on the matching results.

6. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the anti-reverse verification method for financial documents based on adaptive noise as described in any one of claims 1 to 5.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the adaptive noise-based anti-reverse verification method for financial documents as described in any one of claims 1 to 5.

Images