An ai large model archive classification method and device

Abstract

This application discloses an AI large-scale model file classification method and apparatus, comprising the following steps: S1: The central server initializes the model parameters of the global AI large-scale model and distributes the initial model parameters to each client; S2: The global AI large-scale model parameters are updated through multiple rounds of iteration; S3: Each client loads the trained global AI large-scale model, performs classification inference on locally collected new file data, and outputs the file classification result. Belonging to the field of artificial intelligence technology, this structure enables distributed clients to improve the model's adaptability to new tasks through meta-learning without leaking local data.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to a method and apparatus for classifying large AI model files. Background Technology

[0002] Traditional archival classification systems rely on supervised learning artificial intelligence models, which require a large amount of labeled data for training.

[0003] However, in practical applications, such as adding new file categories or rare categories, there is often a problem of scarce sample size, which leads to a sharp decline in model performance and makes it difficult to meet the classification accuracy requirements.

[0004] The existing technical solutions have obvious defects: 1. Centralized meta-learning: Algorithms such as Model-Agnostic Meta-Learning (MAML) can enable models to quickly adapt to new tasks, but they require centralized access to all data and cannot be applied to distributed sensitive data environments. They pose serious privacy and security risks in archival classification scenarios.

[0005] 2. Traditional federated learning: It protects data privacy by training the model locally and only uploading model parameter updates. However, in scenarios with few samples, the amount of data from a single client is insufficient to support the effective training of the local model, which can easily lead to model divergence or overfitting.

[0006] 3. Simple technology overlay: Simply combining meta-learning and federated learning cannot solve the problem of global model instability caused by the heterogeneity of task distribution among clients, and lacks an effective privacy protection mechanism, which cannot meet the high privacy protection requirements of archive data.

[0007] Therefore, there is an urgent need to develop an AI large-scale model file classification method and device to solve the problems in the existing technology. Summary of the Invention

[0008] The purpose of this invention is to provide an AI large model file classification method and apparatus that enables distributed clients to improve the model's adaptability to new tasks through meta-learning without leaking local data. The method is simple in structure and easy to use, thereby solving the problems mentioned in the background art.

[0009] To achieve the above objectives, the present invention provides the following technical solution: A method for classifying large AI model archives, implemented in a distributed system comprising a central server and multiple clients, includes the following steps: S1: The central server initializes the model parameters of the global AI model and distributes the initial model parameters to each client; S2: Update the parameters of the global AI large model through multiple rounds of iteration, including the following sub-steps; S21: The central server selects a portion of clients based on a preset strategy and sends the encrypted current global model parameters to the selected clients. S22: The selected client trains the received model parameters locally based on local private archive data and generates model update information; S23: The client performs differential privacy and homomorphic encryption on the model update information in sequence, generates secure model update information, and uploads it to the central server; S24: The central server performs security aggregation calculations on the security model update information received from multiple clients and updates the global model parameters; S3: Each client loads the trained global AI model, performs classification inference on the newly collected archive data locally, and outputs the archive classification results.

[0010] By adopting the above technical solution, the central server and multiple clients work together to train a high-quality global archive classification model while ensuring the data privacy of all participants. This deconstructs the traditional centralized training into a distributed process of local training, privacy protection, and secure aggregation, ensuring that the original archive data is always kept locally and only model update information that has undergone strict privacy processing is uploaded. This allows for the utilization of data value across institutions while meeting data security regulatory requirements.

[0011] As a further aspect of the present invention: the local training and updating in S22 includes sampling from local private archive data to construct multiple few-shot meta-learning tasks, each meta-learning task including a support set for rapid model adaptation and a query set for performance verification; wherein, for each meta-learning task, the following sub-steps are included: S221: Copy the global model parameters to obtain a local model copy; S222: Train the local model copy on the support set in an inner loop and update the model parameters using gradient descent; S223: Calculate the loss value of the adapted model parameters on the query set as the meta-loss, and calculate the model gradient based on the meta-loss; S224: Aggregate the model gradients from multiple meta-learning tasks to obtain the final model update information.

[0012] By adopting the above technical solutions and the dynamic scheduling mechanism based on task difficulty, the model training follows a scientific learning path from easy to difficult, avoiding the training instability problems that may be caused by directly handling complex tasks.

[0013] As a further aspect of the present invention, it also includes a course learning strategy, which evaluates the difficulty level of the meta-learning task based on inter-class distance and sample entropy indicators; in the early stage of training, it prioritizes training with low-difficulty meta-learning tasks, and as the training rounds increase, it gradually increases the difficulty threshold and introduces high-difficulty meta-learning tasks.

[0014] By adopting the above technical solution, a meta-learning task consisting of a support set and a query set is constructed, and the training process of inner loop adaptation and outer loop optimization is performed locally, enabling the model to quickly learn new categories from a small number of samples.

[0015] As a further aspect of the present invention: the privacy protection processing step in S23 includes the following steps: S231: Perform gradient clipping on the model update information, restrict the gradient norm to a preset range, and then add random noise that conforms to a Gaussian distribution; S232: Use a homomorphic encryption algorithm to encrypt the model update information that has undergone differential privacy processing, and generate secure model update information that can be aggregated and calculated in ciphertext.

[0016] By adopting the above technical solutions, differential privacy processing ensures that individual sample information cannot be reverse-inferred by adding noise to the model update; while homomorphic encryption ensures the confidentiality of the model update during transmission and aggregation. The two work together to provide strict mathematical privacy guarantees while ensuring the availability of the model.

[0017] As a further aspect of the present invention, it also includes online few-sample adaptation, comprising the following steps: ST1: When the client receives fewer than the preset number of labeled samples provided by the user for a new category of archives; ST2: Using the labeled samples as the support set, perform local inner loop training based on the current global model to quickly adapt and obtain a personalized model for the new category; ST3: Uses a personalized model to perform classification reasoning on new category files, with the entire process completed locally without uploading the original data.

[0018] By adopting the above technical solution and using a very small number of labeled samples provided by users for local fine-tuning, the system can improve its classification ability for new categories of files in real time without disconnecting the connection or leaking data.

[0019] As a further aspect of the present invention, the distributed training step further includes: the central server uses the meta-validation dataset to evaluate the performance of the updated global model, and dynamically adjusts the client selection strategy and learning rate parameters for the next round of training based on the evaluation results.

[0020] By adopting the above technical solutions, performance monitoring based on the meta-validation set can promptly detect training anomalies or performance bottlenecks, while the dynamic adjustment strategy can adaptively optimize key parameters according to the training status, thereby improving the resource utilization efficiency of the entire training process while ensuring model quality.

[0021] This application also discloses an AI large model archive classification device, including a central server device and multiple client devices connected via a network; The central server device includes: A model management module used for storing, maintaining, and updating global AI model parameters; A secure aggregation module for securely aggregating and calculating encrypted model update information uploaded by multiple clients; A scheduling management module for managing the distributed training process, including client selection strategies and training round control; An encrypted communication module for enabling secure data transmission with client devices; The client device includes: A local training module used to perform model training based on local private archive data and generate model update information; A privacy protection module for performing privacy protection processing on model update information, including differential privacy units and homomorphic encryption units; A classification reasoning module used to classify archival data using a trained model; A local storage module used for encrypted storage of local private archive data and model parameters.

[0022] By adopting the above technical solutions, the modular design of the central server device and the client device enables the functions of distributed training, privacy protection and file classification to be realized at the hardware level.

[0023] As a further aspect of the present invention: the local training module of the client further includes a meta-learning engine for model training based on a few-shot meta-learning task. The meta-learning engine includes: a task building unit for sampling and constructing a meta-learning task from local data; a model adaptation unit for performing inner loop training on the support set; and a meta-optimization unit for calculating meta-loss and generating gradients on the query set.

[0024] By adopting the above technical solution, the client device can not only perform conventional model training, but also has the ability to handle meta-learning specific tasks, thereby ensuring the execution efficiency and effectiveness of the few-shot learning algorithm at the hardware level.

[0025] As a further aspect of the present invention, it also includes a course learning scheduling unit, used to schedule the training order of meta-learning tasks according to a progressive difficulty strategy, the course learning scheduling unit comprising: A difficulty assessment subunit based on inter-class distance and sample entropy to calculate task difficulty; A scheduling control subunit that dynamically adjusts the difficulty threshold based on the training progress.

[0026] By adopting the above technical solution, intelligent training scheduling capability is provided for the meta-learning process. The collaborative work of the two sub-units of difficulty assessment and scheduling control enables the device to automatically identify the difficulty of the task and optimize the training order, thereby ensuring the stability and efficiency of the training process at the hardware level and reducing the dependence on manual parameter tuning.

[0027] As a further aspect of the present invention: the client also includes a multimodal archive acquisition interface for acquiring different types of archive data, including: Document scanning interface for acquiring digitized images of paper archives; A data import interface for receiving electronic archive files; A voice capture interface used to record voice files and convert them into text data.

[0028] By adopting the above technical solutions, the device supports the collection and preprocessing of various forms of archives, including paper documents, electronic files, and voice data. It can adapt to archive data from different sources and in different formats, providing a complete data access solution for building a unified intelligent archive classification system.

[0029] This application also discloses a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the above-described AI large model file classification method.

[0030] Compared with existing technologies, the beneficial effects of this invention are: it deeply embeds meta-learning into the local training process of federated learning, constructs a federated meta-learning framework, and solves the problems of few-shot learning and privacy protection in a distributed environment; on the one hand, through the combination mechanism of differential privacy, homomorphic encryption, and HSM hardware, it achieves complete privacy protection from the data end to the transmission end and then to the computing end; at the same time, the established few-shot adaptive mechanism can also achieve rapid model optimization when users provide a small number of samples, greatly reducing the labeling cost.

[0031] Other features and advantages of the present invention will be disclosed in detail in the following detailed description and accompanying drawings. Attached Figure Description

[0032] Figure 1 This is a schematic diagram of the central server structure of an AI large model file classification device in an embodiment of the present invention; Figure 2 This is a schematic diagram of the client structure of an AI large model file classification device in an embodiment of the present invention; Figure 3 This is a schematic diagram of a computer-readable storage medium structure in an embodiment of the present invention. Detailed Implementation

[0033] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0034] In this embodiment of the invention, an AI large model file classification method and apparatus are described, see [link to relevant documentation]. Figures 1-2 As shown, it includes the following: 1. System Architecture and Hardware This system adopts a client-server architecture, including a central server device and multiple client devices, which are connected through a secure and encrypted communication channel.

[0035] (1) Central server device The hardware components include: High-performance computing unit: specifically designed for handling the massive model parameter transfer and real-time aggregation calculation in federated learning.

[0036] Secure communication module: Supports TLS / SSL secure communication protocol to ensure the security of the transmission link and prevent parameters from being stolen or tampered with during transmission.

[0037] High-capacity non-volatile storage: Used to store global model parameters, meta-validation datasets, and system logs, supporting fast data read / write and long-term secure storage.

[0038] The software modules include: 1) Global Model Manager Input: Encrypted model gradients uploaded by each client; Processing: A secure aggregation algorithm based on secret sharing is executed. By splitting the gradient across multiple nodes for storage and computation, attacks that maliciously tamper with the gradient are prevented, ensuring the accuracy and security of the aggregation results. The algorithm is tailored to the proportion of local data on each client. ×100%, For the client The amount of data, Calculate a weighted average for the total amount of data from all selected clients; Output: Updated global model parameters.

[0039] 2) Meta-verification and dynamic scheduling engine Input: Candidate global model, meta-validation dataset; Processing: Evaluate model performance on the meta-validation set by calculating classification accuracy and cross-entropy loss; if the improvement in model accuracy is less than 3% for two consecutive rounds, reduce the outer loop learning rate. Reduce by 50%, while also adjusting the client selection ratio. Increase by 0.1; when the loss value increases for two consecutive rounds, stop adjusting the current learning rate and restore the hyperparameter configuration from the previous round. Output: Scheduling parameters such as client-selected strategy and learning rate adjustment instructions.

[0040] (2) Client device The hardware components include: Edge AI Computing Unit: Integrates a dedicated AI chip and a built-in hardware accelerator, supporting efficient model training and inference in local memory.

[0041] Hardware Security Module (HSM): An independent tamper-proof physical chip that provides secure key management, protected cryptographic operations, and true random number generation functions to prevent key leakage and tampering of the calculation process.

[0042] Multimodal data acquisition interface: including high-speed document scanner, high-definition camera, microphone, etc., supporting the digital acquisition of various paper archives, electronic archives, and voice archives.

[0043] Local encrypted storage: Self-encrypting solid-state drives ensure the security of locally stored data; even if the hard drive is physically lost, the data cannot be cracked.

[0044] The software modules include: 1) Local Meta-Learning Engine Input: Local private archive dataset, global model parameters; Processing: Construct meta-tasks, including support sets and query sets, execute the MAML algorithm. The MAML algorithm is a model-independent meta-learning algorithm that does not depend on a specific model structure. It adapts the model to new tasks quickly with only a small number of samples by combining an inner loop to adapt the task with an outer loop to optimize meta-parameters. Calculate the model gradient. Output: Model gradient update amount.

[0045] 2) Privacy Processing Unit Input: Gradient of the original model; Processing: First, differential privacy processing is performed, adding controllable Gaussian noise to the gradient to prevent attackers from retrieving the original data from the output, while ensuring data availability; then, homomorphic encryption is performed. Homomorphic encryption is a special encryption technique that allows computation to be performed directly in ciphertext without decryption, ensuring privacy and security during data transmission and computation. Output: Safety gradient.

[0046] 2. Methodology and Artificial Intelligence Model Training Process (1) System initialization phase The central server loads the initial parameters of the pre-trained large AI model. The pre-trained model uses the Transformer-based BERT-Large model. The pre-training data is a mixture of publicly available general text datasets and de-identified archive datasets. Federated learning hyperparameters and meta-learning hyperparameters are configured, and the output is the initialized global model. .

[0047] In one feasible embodiment, the client selects a proportion when configuring federated learning hyperparameters. : 0.1-0.5; Total number of communication rounds T: 50-200; In the meta-learning hyperparameters, the inner loop learning rate 0.001-0.01, outer loop learning rate : 0.0001-0.001.

[0048] (2) Federal Meta-learning Training Cycle 1. Client Selection and Model Distribution Based on the dynamic scheduling strategy of the meta-validation engine, the server prioritizes clients with similar data distribution to the meta-validation set and good historical training performance, and selects the client set accordingly. The size of the set is C times the total number of clients; The current global model Distributed to selected clients via a secure, encrypted communication channel.

[0049] 2. Local client parallel processing Each client From local dataset Medium sampling of 5-10 meta-tasks ; For each meta-task Execute the MAML algorithm: Copy global model parameters Get a local copy ; In support set Calculate loss The cross-entropy loss function is used. Execute inner loop parameter update: The inner loop iterates 1-3 times; In query set Calculate the element loss ; A first-order approximation is used, meaning that when calculating the meta-gradient, the second derivative of the inner loop gradient with respect to the meta-parameters is ignored, and only the first derivative is retained. This reduces computational complexity and improves training efficiency. The meta-gradient is calculated as follows: ; Aggregating meta-gradients from multiple tasks: ,in The number of meta-tasks; Privacy processing of gradients: First, differential privacy noise is added, then gradient clipping is performed. ,in Set the cropping threshold, with a value between 0.1 and 1; add Gaussian noise: , where σ is the noise standard deviation, with a value of 0.01-0.1, and then homomorphic encryption is performed using the public key in HSM.

[0050] 3. Security Aggregation and Global Updates The server calculates a weighted average using a secure aggregation algorithm based on secret sharing in encrypted form: ; Decrypt the aggregated gradients using the private key in the HSM and update the global model: .

[0051] 4. Meta-verification and Dynamic Adjustment Evaluate next-generation global models using the meta-validation dataset. The performance of the algorithm is evaluated by calculating classification accuracy and cross-entropy loss.

[0052] In one feasible implementation, the meta-validation dataset is derived from samples contributed by each client after anonymization at a ratio of 1:10, with a size of 1000-5000 samples, covering all mainstream archive categories, and the data distribution is heterogeneous from the local data on the client.

[0053] The training strategy for the next round can be dynamically adjusted based on performance trends, such as adjusting the learning rate and the proportion of clients selected.

[0054] (3) Detailed implementation of the Model Fine-Tuner (MAML) algorithm 1. Model copying and initialization: Copying global parameters Copy to your local workspace to get Ensure that the initial state of local parameters is consistent with that of global parameters; 2. Inner loop forward propagation: in the support set The model calculates predicted probabilities from the input sample features, compares these probabilities with the true labels, and calculates the loss using the cross-entropy loss function. ; 3. Inner loop backpropagation: Based on the loss value, the gradient is calculated using the stochastic gradient descent (SGD) algorithm. This reflects the degree to which parameters affect the loss; 4. Inner loop parameter update: By controlling the parameter update magnitude through the learning rate α, the model can quickly adapt to the current meta-task; 5. Outer loop forward propagation: This involves updating the parameters... Applied to query sets Calculate the meta-loss between the predicted results and the true labels. Evaluate the model's adaptability; 6. Backpropagation in the outer loop: The elementary gradient is calculated using a first-order approximation. Capture the impact of parameters on cross-task generalization ability; 7. Meta-gradient aggregation: The meta-gradients of multiple meta-tasks are averaged to obtain the final gradient g_k, reducing the impact of noise from a single task on the gradient.

[0055] (4) Classification reasoning and online adaptive reasoning For the new file X, the trained global model can be used directly. For classification, after inputting file features, the model outputs category probabilities, selects the category with the highest probability as the classification result, and the inference accuracy is no less than 90%. Supports online few-sample adaptation: Using a small number of user-provided samples, 1-10 support sets as adaptation data, the internal loop adaptation process is executed locally on the client, iterating 1-3 times. The adapted model is then used for classification. The entire process is completed locally without uploading new samples, protecting the privacy of newly provided user samples. The classification accuracy can be improved by 5%-15% after adaptation.

[0056] (5) Specific implementation of the course learning scheduler To simulate the human learning process, the model is first trained on simple tasks, and then gradually transitioned to complex tasks, thereby improving the stability and convergence speed of model training. The specific implementation of the course learning scheduler is as follows: The difficulty is measured using two metrics: inter-class distance and sample entropy. Inter-class distance uses cosine distance to calculate the similarity between features of samples from different classes. The smaller the distance, the greater the difficulty in class differentiation. Sample entropy is calculated based on feature distribution. The higher the entropy value, the more complex the feature distribution of the samples, and the greater the difficulty of the task. Scheduling logic: Initialize a difficulty threshold, and gradually increase the threshold as the training rounds increase, transitioning from simple tasks to complex tasks.

[0057] In one feasible embodiment, the inter-class distance threshold is 0.3 and the sample entropy threshold is 0.5. Every 10 rounds, the inter-class distance threshold is increased by 0.05 and the sample entropy threshold is increased by 0.03. Simple tasks include large inter-class distance and low sample entropy, such as the classification of common disease medical records in medical archives. Complex tasks are considered to include small inter-class distance and high sample entropy, such as the classification of rare disease medical records.

[0058] (6) Specific processing flow of the privacy processing unit 1. Differential privacy processing Gradient clipping: Calculate the L2 norm of the gradient. If the norm exceeds the clipping threshold C (ranging from 0.1 to 1), the gradient is clipped proportionally until the norm equals C, as shown in the formula. To prevent individual abnormal gradients from affecting the global model; Add Gaussian noise: based on differential privacy budget Calculate the standard deviation of noise Add a normally distributed gradient to the clipped gradient. Gaussian noise, to obtain This ensures the privacy of gradients.

[0059] 2. Homomorphic encryption: Using the public key generated in HSM, each element in the gradient vector is encrypted based on the national cryptographic algorithm SM9 to obtain the ciphertext gradient, ensuring that even if the gradient is stolen during transmission, it cannot be decrypted to obtain the original gradient information.

[0060] This invention also provides a computer-readable storage medium storing a program thereon, which, when executed by a processor, implements an AI large model file classification method as described in the above embodiments.

[0061] The computer-readable storage medium can be an internal storage unit of any data processing device described in any of the foregoing embodiments, such as a hard disk or memory. The computer-readable storage medium can also be an external storage device of any data processing device, such as a plug-in hard disk, SmartMediaCard (SMC), SD card, or FlashCard equipped on the device. Furthermore, the computer-readable storage medium can include both internal storage units and external storage devices of any data processing device. The computer-readable storage medium is used to store the computer program and other programs and data required by the data processing device, and can also be used to temporarily store data that has been output or will be output.

[0062] This invention provides a method and apparatus for classifying large AI model archives, which enables distributed clients to improve the model's adaptability to new tasks through meta-learning without leaking local data, and has high reliability.

[0063] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0064] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for classifying large-scale AI model archives, characterized in that, When executed in a distributed system comprising a central server and multiple clients, the following steps are included: S1: The central server initializes the model parameters of the global AI model and distributes the initial model parameters to each client; S2: Update the parameters of the global AI large model through multiple rounds of iteration, including the following sub-steps; S21: The central server selects a portion of clients based on a preset strategy and sends the encrypted current global model parameters to the selected clients. S22: The selected client trains the received model parameters locally based on local private archive data and generates model update information; S23: The client performs differential privacy and homomorphic encryption on the model update information in sequence, generates secure model update information, and uploads it to the central server; S24: The central server performs security aggregation calculations on the security model update information received from multiple clients and updates the global model parameters; S3: Each client loads the trained global AI model, performs classification inference on the newly collected archive data locally, and outputs the archive classification results.

2. The AI ​​large model archive classification method according to claim 1, characterized in that, The local training and updating in S22 includes sampling from local private archive data to construct multiple few-shot meta-learning tasks. Each meta-learning task includes a support set for rapid model adaptation and a query set for performance validation. For each meta-learning task, the following sub-steps are included: S221: Copy the global model parameters to obtain a local model copy; S222: Train the local model copy on the support set in an inner loop and update the model parameters using gradient descent; S223: Calculate the loss value of the adapted model parameters on the query set as the meta-loss, and calculate the model gradient based on the meta-loss; S224: Aggregate the model gradients from multiple meta-learning tasks to obtain the final model update information.

3. The AI ​​large model archive classification method according to claim 1, characterized in that, It also includes course learning strategies, which evaluate the difficulty level of meta-learning tasks based on inter-class distance and sample entropy metrics; in the early stages of training, low-difficulty meta-learning tasks are selected for training first, and as the training rounds increase, the difficulty threshold is gradually increased to introduce high-difficulty meta-learning tasks.

4. The AI ​​large model archive classification method according to claim 1, characterized in that, The privacy protection processing steps in S23 include the following steps: S231: Perform gradient clipping on the model update information, restrict the gradient norm to a preset range, and then add random noise that conforms to a Gaussian distribution; S232: Use a homomorphic encryption algorithm to encrypt the model update information that has undergone differential privacy processing, and generate secure model update information that can be aggregated and calculated in ciphertext.

5. The AI ​​large model archive classification method according to claim 1, characterized in that, It also includes online few-shot adaptation, which includes the following steps: ST1: When the client receives fewer than the preset number of labeled samples provided by the user for a new category of archives; ST2: Using the labeled samples as the support set, perform local inner loop training based on the current global model to quickly adapt and obtain a personalized model for the new category; ST3: Uses a personalized model to perform classification reasoning on new category files, with the entire process completed locally without uploading the original data.

6. The AI ​​large model archive classification method according to claim 1, characterized in that, The distributed training step further includes: the central server using the meta-validation dataset to evaluate the performance of the updated global model, and dynamically adjusting the client selection strategy and learning rate parameters for the next round of training based on the evaluation results.

7. An AI large-scale model file classification device, characterized in that, Includes a central server device and multiple client devices connected via a network; The central server device includes: A model management module used for storing, maintaining, and updating global AI model parameters; A secure aggregation module for securely aggregating and calculating encrypted model update information uploaded by multiple clients; A scheduling management module for managing the distributed training process, including client selection strategies and training round control; An encrypted communication module for enabling secure data transmission with client devices; The client device includes: A local training module used to perform model training based on local private archive data and generate model update information; A privacy protection module for performing privacy protection processing on model update information, including differential privacy units and homomorphic encryption units; A classification reasoning module used to classify archival data using a trained model; A local storage module used for encrypted storage of local private archive data and model parameters.

8. The AI ​​large-scale model file classification device according to claim 7, characterized in that, The client's local training module also includes a meta-learning engine for training the model based on a few-shot meta-learning task. The meta-learning engine includes: a task building unit for sampling and constructing the meta-learning task from local data; a model adaptation unit for performing inner loop training on the support set; and a meta-optimization unit for calculating the meta-loss and generating gradients on the query set.

9. The AI ​​large-scale model file classification device according to claim 8, characterized in that, It also includes a course learning scheduling unit, used to schedule the training order of meta-learning tasks according to a progressive difficulty strategy, the course learning scheduling unit comprising: A difficulty assessment subunit based on inter-class distance and sample entropy to calculate task difficulty; A scheduling control subunit that dynamically adjusts the difficulty threshold based on the training progress.

10. The AI ​​large-scale model file classification device according to claim 7, characterized in that, The client also includes a multimodal archive acquisition interface for acquiring different types of archive data, including: Document scanning interface for acquiring digitized images of paper archives; A data import interface for receiving electronic archive files; A voice capture interface used to record voice files and convert them into text data.

Images