Document identification method based on spectral analysis

Abstract

The invention discloses a document identification method based on spectral analysis, which comprises the following steps: acquiring spectral data of a local document, and carrying out data processing and feature extraction on the spectral data to obtain a local spectral feature; training a local teacher model through the local spectral features to obtain spectral feature knowledge representation of the local teacher model, and uploading the spectral feature knowledge representation of the local teacher model to a central server; training a global student model through the spectral feature knowledge representation of the plurality of local teacher models in the central server, and generating an antagonistic spectral sample through the global student model and the spectral feature knowledge representation; and issuing the global student model and the confrontation spectral feature sample to the local for incremental iteration training to obtain a local document identification model, and outputting an identification result for the local document spectral data based on the model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of document authentication, and in particular to a document authentication method based on spectral analysis. BACKGROUND

[0002] With the continuous improvement of requirements for document authenticity and traceability in the fields of justice, finance, and archive management, spectral analysis (especially Raman spectroscopy) has become an important technical means for document authentication due to its advantages of non-destructive, no need for reagents, high resolution, etc. Raman spectroscopy can reveal the molecular structure information of key components such as ink and paper, and is often used to determine the age of documents, detect tampering traces, and identify counterfeit materials. Existing methods usually include sample pretreatment, spectral data acquisition, signal denoising and baseline correction, feature extraction, database comparison, and authenticity determination, and some institutions have introduced deep learning to improve the discrimination ability in complex scenarios.

[0003] However, existing judicial evidence preservation, financial and archive compliance have strict restrictions on data cross-domain flow, and original document spectral data is prohibited or difficult to share between institutions, which prevents the direct aggregation of multi-institutional samples through centralized training. The amount of local samples available to each institution is limited, and there is a significant regional and scenario bias (such as significant differences in ink, paper, and age proportions in different regions). The generalization ability of models trained by a single institution is insufficient, and the cross-domain migration performance is unstable. New ink formulations, antique materials, and chemical treatment processes are constantly emerging, which are beyond the existing experience distribution of a single institution, leading to model updates lagging behind and insufficient recognition of new counterfeit materials. Traditional training lacks systematic adversarial perturbations and physical consistency constraints, making it vulnerable to "unphysical strong adversarial samples" or "physically reasonable but insufficiently attackable samples", making it difficult to balance robustness and credibility. During the local continuous incremental update process, the model is prone to forgetting previously learned knowledge under historical material / ink conditions (catastrophic forgetting), resulting in a decline in old domain performance.

[0004] In view of the above problems existing in the prior art, a document authentication method based on spectral analysis is designed, which is the purpose of the present application. SUMMARY

[0005] Therefore, the purpose of the present application is to provide a document authentication method based on spectral analysis, which can solve the above problems.

[0006] The present application provides a document authentication method based on spectral analysis, comprising: obtaining spectral data of local documents, performing data processing and feature extraction on the spectral data to obtain local spectral features; training a local teacher model through the local spectral features to obtain spectral feature knowledge representation of the local teacher model, and uploading the spectral feature knowledge representation of the local teacher model to a central server; training a global student model in the central server through the spectral feature knowledge representation of a plurality of local teacher models, and generating an adversarial spectral sample through the global student model and the spectral feature knowledge representation; downlinking the global student model and the adversarial spectral feature sample to the local for incremental iterative training to obtain a local document authentication model, and outputting an authentication result based on the model for local document spectral data.

[0007] The present application has the following beneficial effects: First, through PCA dimension reduction and redundancy removal, the problem of large computational overhead and redundant interference caused by high-dimensional spectrum is solved, through wavelet / wavelet packet multi-scale local feature extraction, the problem of PCA insensitivity to local subtle forgery is solved, through autoencoder learning of nonlinear hidden space representation, the problem of complex aging, mixed ink and nonlinear coupling difficult to model is solved, and more stable and higher generalization performance of local spectral features is obtained.

[0008] Second, by uploading only soft labels / features instead of raw spectra, the problem of cross-agency data out-of-domain and privacy compliance restrictions is solved, by filtering soft labels through confidence and information entropy, the problem of noise and uncertain knowledge polluting the central model is solved, by jointly outputting soft labels through "labeled supervision + unlabeled reasoning", the problem of annotation scarcity and insufficient knowledge coverage is solved, bringing wider sample coverage and more comprehensive discrimination boundary.

[0009] Third, through entropy reciprocal weighted KL distillation, the problem of global model interference by noise caused by uneven knowledge quality of different agencies is solved, bringing improvement of generalization and stability of the global student model. Through material perception adversarial generation with the global student as the discriminator, the problem of insufficient coverage of scarce / difficult-to-sample boundaries and new forgeries is solved. Through joint optimization of adversarial loss and physical consistency loss, a credible and effective adversarial sample library is brought. Through centralized training and periodic iteration without sharing raw data, the problem of being unable to centrally gather data to build a national model is solved, enabling stronger cross-domain generalization in multi-domain scenarios.

[0010] Fourth, by downlinking global student parameters as initialization and teacher signals, the problem of slow local convergence and inconsistent decision boundaries is solved, bringing fewer training rounds and higher starting performance. Through triple memory replay (samples, soft / true labels, weights), the problem of catastrophic forgetting in local incremental learning is solved, bringing stable performance in old domains and continuous learning ability. Through dynamic proportion mixing training of adversarial samples and local / replay samples, the problem of insufficient robustness to new inks, antique materials and perturbations is solved. Through weighted joint loss (task loss + distillation loss) and sample weighting, the problem of unstable training and insufficient hard example learning caused by heterogeneous samples is solved. BRIEF DESCRIPTION OF DRAWINGS

[0011] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the accompanying drawings needed to be described in the embodiments or the prior art will be briefly introduced. Obviously, the accompanying drawings in the following description only only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0012] Figure 1 is a method flowchart of the present embodiment. DETAILED DESCRIPTION

[0013] For the convenience of those skilled in the art to understand, the structure of the present application will be further described in detail in combination with the drawings. It should be understood that, in the present embodiment, the order of the steps mentioned can be adjusted according to actual needs, even can be executed simultaneously or partially simultaneously, except that the order is particularly specified.

[0014] As Figure 1 shown, the present embodiment provides a document authentication method based on spectral analysis, comprising: S1 obtaining spectral data of local documents, performing data processing and feature extraction on the spectral data to obtain local spectral features; S101 acquiring spectral data of local documents by spectral instrument, performing denoising, baseline correction and normalization processing on the spectral data to obtain normalized spectral data; S102 sequentially performing principal component analysis, wavelet decomposition and autoencoder feature fusion on the normalized spectral data, extracting multi-modal spectral features, and combining into a feature vector as a local spectral feature.

[0015] S1021 performing principal component analysis on the normalized spectral data to obtain the main variation components in the data, reducing the original high-dimensional spectral data to several main component feature vectors to obtain PCA principal component feature vectors; S1022 performing wavelet transform or wavelet packet decomposition on the normalized spectral data to extract features containing different frequency band information to obtain wavelet decomposition coefficient feature vectors; S1023 performing unsupervised training on the normalized spectral data by using a deep autoencoder model, encoding the data into a feature vector in a low-dimensional hidden space to obtain an autoencoder hidden space feature vector; S1024 fusing the PCA principal component feature vectors, the wavelet decomposition coefficient feature vectors and the autoencoder hidden space feature vectors according to a preset dimension to form a multi-modal spectral feature vector as a local spectral feature.

[0016] In this step, the spectral instrument can be selected as a Raman spectrometer. The instrument emits a beam of a specific wavelength (laser / infrared / visible light / X-ray) to the surface of the document sample. The sample material produces a specific absorption, scattering, fluorescence or reflection response to the incident light. After the detector of the instrument receives the response signal, wavelength decomposition is performed to generate a spectral curve (i.e., the light intensity distribution at each wavelength point) representing the sample composition. The instrument generates spectral raw data files in.spc,.csv,.txt formats, which can be used for subsequent preprocessing and analysis. The principal component analysis (PCA) method is used to extract the main components from the high-dimensional spectrum to compress the redundancy, improve the efficiency of subsequent model training and automatic recognition accuracy, and avoid the processing bottleneck and misjudgment risk caused by directly operating high-dimensional raw data. The feature extraction algorithm is prior art, and the present application will not be described in detail.

[0017] The specific type of document forgery may only occur in some local frequency bands, such as the difference between the reflectivity of the imitation ink in some bands and the original. The PCA feature is not sensitive, and the wavelet energy can accurately find it. Some deep imitation shows nonlinear mixed signal changes, and the hidden variable space of the autoencoder can effectively model them. The fusion of the three types of features effectively reduces the misjudgment caused by single-point failure.

[0018] The spectral absorption of aged paper increases in some near-infrared bands. Assuming that the original spectrum has 1000 wavelength channels, only after PCA processing, the principal component feature reflects the overall absorption change, but it is not clear which small part of the change may be diluted. The wavelet decomposition can accurately capture the local high-frequency mutation position of the spectral signal, such as the sharp change in [820-830nm], which is beneficial to distinguish between new and old paper. The autoencoder further models the nonlinear features of the overall spectrum and finds that some local aging shows a group of cross-section combination features (such as [820-830nm]&[950-970nm] jointly abnormal). Finally, the three types of features are spliced, which greatly improves the accuracy and robustness of aging identification.

[0019] The imitation ink only has a slight difference in reflectivity with the genuine product in some sub-bands (such as 640-660nm). PCA cannot capture it, and the component coefficients differ significantly after wavelet decomposition (such as high-frequency coefficient c4,35: 0.02 for the original product and 0.15 for the imitation product). Some imitation products show global nonlinear shift through mixed ink making techniques. The autoencoder can distinguish this overall distribution difference through end-to-end learning. After the fusion of the three features, the imitation detection precision can be improved.

[0020] S2 trains a local teacher model through local spectral features, obtains a spectral feature knowledge representation of the local teacher model, and uploads the spectral feature knowledge representation of the local teacher model to a center server; S201 constructing a local teacher model based on a deep neural network, using a labeled local spectral feature to train the local teacher model with cross-entropy loss, obtaining a first set of local soft labels; S202 inputting an unlabeled local spectral feature into the local teacher model after cross-entropy loss training, obtaining a second set of local soft labels; S203 calculating the confidence and information entropy of the soft label, filtering out soft labels with low confidence, and uploading the local spectral feature and its corresponding first soft label and second local soft label to the center server as spectral feature knowledge representation.

[0021] S2031 calculating the maximum class probability as the confidence of all soft labels, and calculating the information entropy, the calculation formula is as follows: , , wherein, represents the spectral feature of a sample, represents the confidence, represents the probability of the soft label of the sample corresponding to the i-th class, i represents the class number, represents the information entropy; S2032 retaining soft labels with confidence higher than a confidence threshold and entropy value lower than an entropy threshold, and removing other soft labels, and uploading to the center server.

[0022] In this step, the confidence is used to evaluate the "self-confidence" degree of the model in classifying a certain sample. High confidence represents that the model is clear in judgment and the output distribution is sharp; low confidence means that the model is more "hesitant" in classifying the sample, which may be a boundary sample difficult to distinguish between multiple categories.

[0023] Information entropy is used to reflect the "determinacy" or "uncertainty" of the prediction distribution of the model for a certain sample. Low entropy means that the probability is concentrated, and the model is "firm"; high entropy means that the probability distribution of each category is relatively uniform, indicating that the model is "swaying" or "confused", and such soft labels have large signal noise and are easy to mislead downstream training.

[0024] Only uploading clear soft labels with high confidence and low entropy can prevent a large number of low-quality samples from being mixed in the training process, reduce invalid or misleading gradient signals, and improve the stability of student model learning. Through filtering, only reliable knowledge representation is uploaded, which can achieve efficient modeling and reduce the risk of leakage caused by a large amount of invalid data.

[0025] The spectral data itself is highly sensitive, and the original spectral transmission has the risk of material information leakage. After the original data is trained locally to generate knowledge representation (soft label, feature embedding, etc.) and uploaded, multi-center collaborative modeling can be realized without leaking specific original document details. Further, the labeled sample collection of document spectral data is expensive, but the unlabeled data is easy to obtain. The design allows the local teacher network to be directly trained with labeled data, while the unlabeled data is inferred by the trained model, which expands the knowledge coverage and effectively utilizes all available data. The soft label (probability output) obtained using the DNN model not only reflects the model confidence, but also contains the "fuzzy boundary" between classes, which is more helpful for the center student model to inherit the discrimination details of the local model and promote better generalization ability.

[0026] The local model first trains knowledge (soft label) through labeled and unlabeled data, and these abstract knowledge is uploaded to the center. The center distills the student model according to the uploaded knowledge from each place. Decentralization and cross-domain heterogeneous knowledge fusion aggregate knowledge / representation / behavior rather than simply model parameters, which is suitable for distributed heterogeneous and data privacy sensitive document identification scenarios.

[0027] S3 trains a global student model in the center server through spectral feature knowledge representations of a plurality of local teacher models, and generates adversarial spectral samples through the global student model and the spectral feature knowledge representations; S301 trains a global student model in the center server by adaptively weighting knowledge distillation according to the entropy weight of the uploaded spectral feature knowledge representation; S3011 obtains a corresponding soft label of each uploaded spectral feature knowledge representation, calculates the information entropy according to the corresponding soft label, normalizes the inverse of the information entropy of each spectral feature knowledge representation as its weight, and forms a sample triple set; S3012 uses each sample weight to weight the KL divergence loss based on the sample triple set, takes the weighted loss as the objective function for optimizing the global student model, and iteratively trains until convergence.

[0028] In this step, in a real application scenario, the knowledge uploaded by different local models has different confidence / reliability / representativeness. Some soft labels are very certain (high confidence, low entropy), and others are very uncertain (low confidence, high entropy). If they are not distinguished, weak signals will be mixed into mainstream knowledge, polluting the global student model. The inverse entropy weight is used in the sample triple (spectral feature, soft label, weight) to realize differentiated use of samples, so that local samples with high reliable knowledge can contribute more to the global model, and the influence of noise / fuzzy samples is reduced.

[0029] S302 generates an adversarial spectral feature sample based on the trained global student model, in combination with the spectral feature knowledge representation, through the material perception adversarial generation network and the physical property constraint.

[0030] S3021 uses the trained global student model as the discriminator of the material perception adversarial generation network, and inputs the spectral feature knowledge representation and its corresponding document material, ink category, and physical property as an input sample combination; S3022 inputs the input sample combination combined with random noise into the generator of the material perception adversarial generation network to generate a fake spectral feature sample; S3023 evaluates the adversarial loss of the fake spectral feature sample through the discriminator of the adversarial generation network, and evaluates the physical loss through the physical consistency constraint; S3024 combines the adversarial loss and the physical loss by weighting, and optimizes the generator parameters to enable the generation of adversarial spectral feature samples under the constraint of physical properties.

[0031] In this step, the trained global student model is used as the discriminator of the material perception adversarial generation network to ensure that the discrimination target is consistent with the real deployment model, and the generated sample can be finely controlled according to the specified material / ink, covering more real working conditions. Optimizing only the adversarial loss may produce fake samples that violate physical laws, and optimizing only the physical consistency is not aggressive enough to expose the real weaknesses of the model. By combining the adversarial loss and the physical loss by weighting, the model training defects caused by a single loss are avoided.

[0032] S4 distributes the global student model and the adversarial spectral feature sample to the local node for incremental iterative training to obtain a local document identification model, and outputs an identification result based on the model.

[0033] S401 The central server distributes the global student model parameters and the adversarial spectral feature sample to the local node; S402 The local node performs incremental training using a three-tuple memory replay mechanism to obtain an updated local document identification model.

[0034] S4021 constructs historical local spectral features into a local memory library, and continuously supplements and maintains them with new samples, records corresponding soft labels / true labels and sample weights; S4022 samples local spectral features from the local memory library according to a dynamic proportion, and mixes them with adversarial spectral samples to form a memory replay three-tuple (local spectral feature, soft label / true label, sample weight); S4023 constructs a weighted training batch based on the memory replay three-tuple and the distillation mechanism, uses a joint loss containing task loss and distillation loss for forward and loss aggregation, and iteratively updates to obtain a local document identification model.

[0035] In this step, through the global student model issued by the center, the local training round can be reduced, the consistency of the global and local decision boundaries is ensured, and the subsequent distillation is beneficial. The historical knowledge is continuously consolidated through the memory library playback, and the performance under the old material / ink condition is maintained stable; the sample weight allows the difficult examples and key conditions to be trained more fully. The historical decision boundary is represented by a soft label, and the important sample contribution is controlled by a weight. The memory library represents the representativeness and timeliness. The task loss guarantees the current local target performance, the distillation loss aligns the local model with the global student / historical soft label, slows down the forgetting, and combines the two losses to avoid unstable training caused by sample heterogeneity.

[0036] Those skilled in the art will appreciate that embodiments of the application can be provided as methods, systems, or computer program products. Accordingly, the application can be embodied in the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the application can be embodied in the form of a computer program product on one or more computer readable storage media (including, but not limited to, disk memory, CD-ROMs, optical storage media, etc.) having computer usable program code embodied thereon.

[0037] The present application is described in reference to the flowcharts and / or block diagrams of the methods, apparatus (systems) and computer program products according to the embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, as well as combinations of flows and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing apparatus to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing apparatus produce an apparatus that implements the functions specified in the flowcharts and / or block diagrams. Figure 1 The functions specified in a flow or multiple flows and / or blocks Figure 1 The functions specified in a flow or multiple flows and / or blocks

[0038] These computer program instructions can also be stored in a computer readable memory that can direct the computer or other programmable data processing apparatus to work in a specific manner, so that the instructions stored in the computer readable memory produce a manufactured product including instruction apparatus, which implements the functions specified in the flowcharts and / or block diagrams. Figure 1 The functions specified in a flow or multiple flows and / or blocks Figure 1 The functions specified in a flow or multiple flows and / or blocks

[0039] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus, so that a series of operation steps are performed on the computer or other programmable data processing apparatus to produce a computer implemented process, so that the instructions executed on the computer or other programmable data processing apparatus provide a process for implementing the functions specified in the flowcharts and / or block diagrams.Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0040] It should be noted that in the claims, any reference signs placed between parentheses shall not be construed as limiting the claims. The word "comprising" does not exclude the presence of components or steps not listed in the claim. The word "a" or "an" preceding a component does not exclude the presence of a plurality of such components. The invention can be implemented by means of hardware comprising several distinct components and by means of a suitably programmed computer. In a unit claim enumerating several means, several of these means may be embodied by one and the same item of hardware. The use of the words first, second, third etc. does not indicate any order. These words may be interpreted as names.

[0041] Although the preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0042] Obviously, those skilled in the art may make various modifications and variations to the present invention without departing from the spirit and scope of the present invention. Thus, if such modifications and variations fall within the scope of the claims and their equivalents, the present invention is intended to include such modifications and variations.

[0043] In the present invention, unless otherwise expressly specified or limited, the terms "mounted," "connected," "connect," "fixed," etc. should be understood broadly. For example, they may refer to fixed connection, detachable connection, or integration; mechanical connection or electrical connection; direct connection or indirect connection through an intermediate medium; internal communication between two components or interaction between two components. Those skilled in the art will understand the specific meanings of the above terms in the present invention based on specific circumstances.

[0044] In the description of the specification, the description of the terms "one embodiment", "some embodiments", "an example", "a specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. In the specification, the illustrative description of the above terms should not be understood as necessarily referring to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any suitable manner in one or more embodiments or examples. In addition, the person skilled in the art can combine and combine the different embodiments or examples described in the specification and the features of the different embodiments or examples, without contradiction.

Claims

1. A document identification method based on spectral analysis, characterized in that: include: Acquire spectral data of local documents, perform data processing and feature extraction on the spectral data, and obtain local spectral features; The local teacher model is trained using local spectral features to obtain the spectral feature knowledge representation of the local teacher model, and the spectral feature knowledge representation of the local teacher model is uploaded to the central server; In the central server, a global student model is trained through the spectral feature knowledge representation of several local teacher models, and adversarial spectral samples are generated through the global student model and the spectral feature knowledge representation; The global student model and adversarial spectral feature samples are sent to the local area for incremental iterative training to obtain a local document identification model, and the identification results of the local document spectral data are output based on the model.

2. The document identification method based on spectral analysis according to claim 1, characterized in that: The obtaining of spectral data of a local document, performing data processing and feature extraction on the spectral data, and obtaining local spectral features includes: The spectral data of local documents are collected by a spectral instrument, and the spectral data are subjected to denoising, baseline correction, and normalization to obtain normalized spectral data; The normalized spectral data are subjected to principal component analysis, wavelet decomposition and autoencoder feature fusion in sequence to extract multimodal spectral features and combine them into feature vectors as local spectral features.

3. The document identification method based on spectral analysis according to claim 2, characterized in that: The normalized spectral data is subjected to principal component analysis, wavelet decomposition, and autoencoder feature fusion in sequence to extract multimodal spectral features and combine them into feature vectors as local spectral features, including: Perform principal component analysis on the normalized spectral data to obtain the main variation components in the data, reduce the original high-dimensional spectral data into several main component eigenvectors, and obtain the PCA main component eigenvectors; Perform wavelet transform or wavelet packet decomposition on the normalized spectral data to extract features containing information of different frequency bands and obtain the wavelet decomposition coefficient feature vector; The deep autoencoder model is used to perform unsupervised training on the normalized spectral data, encoding the data into a feature vector of a low-dimensional latent space, and obtaining the feature vector of the autoencoder latent space; The PCA principal component eigenvector, wavelet decomposition coefficient eigenvector, and autoencoder latent space eigenvector are fused according to the preset dimension to form a multimodal spectral feature vector as the local spectral feature.

4. The document identification method based on spectral analysis according to claim 1, characterized in that: The training of the local teacher model by using the local spectral features to obtain the spectral feature knowledge representation of the local teacher model and uploading the spectral feature knowledge representation of the local teacher model to the central server includes: A local teacher model is constructed based on a deep neural network. The local spectral features with labels are used to train the local teacher model with cross entropy loss to obtain the first set of local soft labels. The unlabeled local spectral features are input into the local teacher model trained with cross-entropy loss to obtain a second set of local soft labels. The confidence and information entropy of the soft labels are calculated, and the soft labels with low confidence are filtered out. The local spectral features and their corresponding first and second local soft labels are used as spectral feature knowledge representations and uploaded to the central server.

5. The document identification method based on spectral analysis according to claim 4, characterized in that: The calculating of the confidence and information entropy value of the soft label, filtering out the soft labels with low confidence, and uploading the local spectral feature and its corresponding first soft label and second local soft label as spectral feature knowledge representation to the central server includes: Calculate the maximum category probability as the confidence for all soft labels and calculate its information entropy. The calculation formula is as follows: ， ， in, Represents the spectral characteristics of a sample, Indicates confidence, Indicates the probability that the sample soft label corresponds to the i-th category, i represents the category number, represents information entropy; The soft labels with confidence higher than the confidence threshold and entropy lower than the entropy threshold are retained, and the other soft labels are removed and uploaded to the central server.

6. The document identification method based on spectral analysis according to claim 1, characterized in that: The method of training a global student model in the central server by using the spectral feature knowledge representations of several local teacher models and generating adversarial spectral samples by using the global student model and the spectral feature knowledge representations includes: In the central server, adaptive weighted knowledge distillation is performed according to the entropy weights of the uploaded spectral feature knowledge representation to train a global student model. Based on the trained global student model, combined with the spectral feature knowledge representation, adversarial spectral feature samples are generated through material-aware adversarial generative network and physical property constraints.

7. The document identification method based on spectral analysis according to claim 6, characterized in that: In the central server, adaptive weighted knowledge distillation is performed according to the entropy weight of the uploaded spectral feature knowledge representation, and the global student model obtained by training includes: Obtain the corresponding soft label of each uploaded spectral feature knowledge representation, calculate the information entropy based on the corresponding soft label, and normalize the inverse of the information entropy of each spectral feature knowledge representation as its weight to form a sample triple set; Based on the set of sample triplets, the KL divergence loss is weighted by using the weight of each sample, and the weighted loss is used as the objective function to optimize the global student model. The training is iterated until convergence.

8. The document identification method based on spectral analysis according to claim 6, characterized in that: The method of generating adversarial spectral feature samples based on the trained global student model, combined with spectral feature knowledge representation, material-aware adversarial generative network and physical property constraints includes: The trained global student model is used as the discriminator of the material-aware adversarial network, and the spectral feature knowledge representation and its corresponding document material, ink type, and physical properties are used as input sample combinations. The input sample combination is combined with random noise and fed into the generator of the material-aware adversarial network to generate fake spectral feature samples; The adversarial loss of the forged spectral feature samples is evaluated by the discriminator of the adversarial network, and the physical loss is evaluated by the physical consistency constraint; The adversarial loss and the physical loss are weighted together to optimize the generator parameters so that it can generate adversarial spectral feature samples under the constraints of physical properties.

9. The document identification method based on spectral analysis according to claim 1, characterized in that: The global student model and the adversarial spectral feature samples are sent to the local for incremental iterative training to obtain a local document identification model, and the identification results of the local document spectral data are output based on the model, including: The central server sends the global student model parameters and adversarial spectral feature samples to the local node; The local node performs incremental training using the triplet memory playback mechanism to obtain an updated local document authentication model.

10. The document identification method based on spectral analysis according to claim 9, characterized in that: The local node performs incremental training using a triplet memory playback mechanism to obtain an updated local document authentication model including: The historical local spectral features are constructed into a local memory library, which is continuously supplemented and maintained with newly collected samples, and the corresponding soft labels / true labels and sample weights are recorded; Local spectral features are sampled from the local memory bank at a dynamic ratio and mixed with adversarial spectral samples to form a memory replay triplet (local spectral feature, soft label / true label, sample weight); Based on the memory replay triples and distillation mechanism, weighted training batches are constructed. A joint loss including task loss and distillation loss is used for forward transmission and loss aggregation, and the local document authentication model is obtained by iterative updating.

Images