Adaptive large model parameter fine tuning method and system based on end-cloud collaboration

Abstract

The invention relates to a self-adaptive large model parameter fine tuning method and system based on end-cloud collaboration. The method comprises the steps that an edge end and a cloud end are arranged to train a large model, an end-cloud collaborative architecture is formed to achieve distributed processing of model training, and the efficiency of model fine adjustment is improved by means of the instant data processing capacity of the edge end and powerful computing resources of the cloud end; a self-adaptive parameter optimization mechanism is introduced, an updating strategy of model parameters is dynamically adjusted according to the performance of the model at different ends, and efficient utilization of the model parameters is ensured; through a key exchange protocol, encryption operation and a data confusion technology, the security and privacy of data in a model fine tuning process are ensured; the dynamic adjustment strategy of model fine adjustment adjusts the training load and resource allocation of the model in real time according to the performance feedback of each end in the end-cloud collaborative architecture, and can flexibly adapt to different operation environments and task requirements.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a method and system for fine-tuning parameters of an adaptive large model based on end-cloud collaboration. Background Art

[0002] With the development of society, people's living standards are constantly improving, and there are many ways to process data. In the field of artificial intelligence, large models (LLMs) have attracted much attention due to their excellent natural language processing capabilities and are widely used in various tasks such as text understanding, language translation, and question answering. However, when LLMs are applied to specific tasks, they usually need to be fine-tuned to adapt to the specific data distribution and task requirements. Existing fine-tuning methods rely on a large amount of labeled data and computing resources, which has some limitations in practical applications.

[0003] Specifically, the existing technologies mainly face three core problems: First, the existing fine-tuning methods usually require large-scale data sets to train models, which brings challenges in data privacy and security; in application scenarios such as healthcare and financial services, data is often sensitive and cannot be used directly for model training. Second, the existing fine-tuning methods have high requirements for computing resources. Training large models requires a lot of computing power, which usually requires the use of high-performance GPU clusters, which are not popular among many institutions and individual developers. In addition, each fine-tuning consumes a lot of electricity and time, which is impractical in resource-limited environments. Furthermore, the existing end-cloud collaborative solutions have not fully considered the efficient use of model parameters and cross-device data privacy protection in the fine-tuning of LLMs. The existing solutions mainly focus on the distributed training and knowledge distillation technology of the model, but ignore the refined management of model parameters and the strict protection of data privacy, resulting in inefficiency and privacy leakage risks in practical applications.

[0004] In summary, the traditional large model parameter fine-tuning method has low adjustment efficiency and problems of unreliable and unsafe data during the adjustment process. Summary of the invention

[0005] Based on this, in order to solve the above technical problems, an adaptive large model parameter fine-tuning method and system based on end-cloud collaboration is provided, which can achieve a more efficient, safer and more reliable model fine-tuning process.

[0006] An adaptive large model parameter fine-tuning method based on end-cloud collaboration, the method comprising:

[0007] Collecting large model training data through the edge end, preprocessing the large model training data, performing preliminary training of the large model according to the preprocessed data, and obtaining a preliminary large model; identifying key frames through the edge end, and obtaining prediction results based on the key frames and the preliminary large model;

[0008] The encryption key is exchanged between the edge and the cloud through a key exchange protocol; on the edge, a symmetric encryption algorithm is used to encrypt the transmission data to be transmitted to the cloud, and data obfuscation technology is introduced on the basis of data encryption; on the edge, a message authentication code is used to verify the integrity of the transmission data;

[0009] The cloud collects performance feedback from each of the edge terminals, evaluates the training load of each of the edge terminals according to the performance feedback, dynamically adjusts resource allocation of each of the edge terminals according to the training load, and updates the model fine-tuning strategy of each of the edge terminals based on the resource allocation result;

[0010] Receiving the encrypted transmission data transmitted from the edge end through the cloud, decrypting the transmission data to obtain the key frame and the prediction result; fine-tuning the parameters of the preliminary large model based on the key frame according to the model fine-tuning strategy, and optimizing the parameters of the fine-tuned preliminary large model based on the prediction result to obtain an optimized large model;

[0011] The optimized large model is evaluated according to performance evaluation indicators, and parameter update is triggered according to the evaluation results; the performance indicators of the optimized large model in the cloud and the edge are collected and monitored in real time, and the model parameter update strategy is dynamically adjusted according to the performance indicators to determine the model parameter update frequency; the large model parameters are adjusted based on the model parameter update strategy and the model parameter update frequency.

[0012] In one embodiment, the edge end preprocesses the large model training data, performs preliminary training of the large model according to the preprocessed data, and obtains a preliminary large model, including:

[0013] At the edge end, performing preprocessing operations of denoising, normalization, and data enhancement on the large model training data to obtain preprocessed data;

[0014] Performing a feature extraction operation on the preprocessed data to extract model training features;

[0015] The model training features are used to perform preliminary training of the large model to obtain a preliminary large model.

[0016] In one embodiment, identifying a key frame through the edge end, and obtaining a prediction result based on the key frame and the preliminary large model includes:

[0017] A characteristic value threshold is set at the edge end to determine the characteristic value corresponding to the model training feature;

[0018] Based on the eigenvalue threshold, taking the frame with the largest eigenvalue change as the key frame;

[0019] The edge end obtains model parameters according to the preliminary large model, and generates prediction results based on the model parameters and the key frames.

[0020] In one embodiment, the method further comprises:

[0021] The edge end uses a message authentication code to verify the integrity of the transmission data, and after the verification is passed, sends the transmission data to the cloud;

[0022] The cloud receives the transmission data and verifies the transmission data using the same key as the edge to obtain an integrity verification result.

[0023] In one embodiment, the cloud collects performance feedback from each edge terminal, including:

[0024] The cloud end collects the training loss value, processing time, and model accuracy of the preliminary large model in each of the edge ends;

[0025] The cloud obtains a performance feedback set of each of the edge terminals based on the training loss value, processing time, and model accuracy.

[0026] In one embodiment, the cloud evaluates the training load of each edge terminal according to the performance feedback, including:

[0027] The cloud determines the number of edge terminals based on the connection relationship with each edge terminal, and obtains the training loss value and processing time of each edge terminal according to the performance feedback;

[0028] The cloud end evaluates the training load of each of the edge ends based on the training loss value and the processing time.

[0029] In one embodiment, the cloud updates the model fine-tuning strategy of each edge terminal based on the resource allocation result, including:

[0030] The cloud end adjusts the model learning rate based on the resource allocation result and the performance feedback;

[0031] The cloud end collects the updated model parameters of each of the edge ends;

[0032] The cloud end updates the model fine-tuning strategy of each of the edge ends according to the model learning rate and the model parameters.

[0033] In one embodiment, the real-time monitoring collects the performance indicators of the optimized large model on the cloud and the edge, dynamically adjusts the model parameter update strategy according to the performance indicators, and determines the model parameter update frequency, including:

[0034] Through the performance monitoring function, the accuracy and loss value of the optimized large model on the cloud and the edge are monitored and collected respectively, and the accuracy and loss value are used as performance indicators of the large model;

[0035] Constructing a multi-factor trade-off model, inputting the performance indicator into the multi-factor trade-off model, and dynamically adjusting the model parameter update strategy through the multi-factor trade-off model;

[0036] Based on the performance index, obtain the performance change amount and performance change threshold of the optimized large model; determine the frequency control coefficient;

[0037] The model parameter update frequency is determined according to the performance change amount, the performance change threshold, and the frequency control coefficient.

[0038] In one embodiment, the cloud fine-tunes parameters of the preliminary large model based on the key frames according to the model fine-tuning strategy, and optimizes parameters of the fine-tuned preliminary large model based on the prediction results to obtain an optimized large model, including:

[0039] The cloud determines the learning rate corresponding to the preliminary large model according to the model fine-tuning strategy, and obtains the gradient data of the loss function to the model parameters in the preliminary large model;

[0040] The cloud uses the key frame to fine-tune the parameters of the preliminary large model according to the learning rate and gradient data, and uses the prediction result as a verification condition to optimize the parameters of the fine-tuned preliminary large model;

[0041] The cloud end globally optimizes the preliminary large model according to the key frames and the prediction results to obtain an optimized large model.

[0042] An adaptive large model parameter fine-tuning system based on end-cloud collaboration, the system comprising:

[0043] The edge end is used to collect large model training data, pre-process the large model training data, perform preliminary training of the large model according to the pre-processed data, and obtain a preliminary large model; identify key frames through the edge end, and obtain prediction results based on the key frames and the preliminary large model;

[0044] The encryption key is exchanged between the edge and the cloud through a key exchange protocol; the edge is also used to: use a symmetric encryption algorithm to encrypt the transmission data to be transmitted to the cloud, and introduce data obfuscation technology on the basis of data encryption; use a message authentication code on the edge to verify the integrity of the transmission data;

[0045] The cloud end is used to collect performance feedback of each of the edge ends, evaluate the training load of each of the edge ends according to the performance feedback, dynamically adjust the resource allocation of each of the edge ends according to the training load, and update the model fine-tuning strategy of each of the edge ends based on the resource allocation result;

[0046] The cloud end is further used to receive the encrypted transmission data transmitted from the edge end, decrypt the transmission data to obtain the key frame and the prediction result; according to the model fine-tuning strategy, fine-tune the parameters of the preliminary large model based on the key frame, and optimize the parameters of the fine-tuned preliminary large model based on the prediction result to obtain an optimized large model;

[0047] The control end is used to evaluate the optimized large model according to the performance evaluation index, and trigger parameter update according to the evaluation result; monitor and collect the performance index of the optimized large model in the cloud and the edge in real time, dynamically adjust the model parameter update strategy according to the performance index, and determine the model parameter update frequency; adjust the large model parameters based on the model parameter update strategy and model parameter update frequency.

[0048] The above-mentioned adaptive large model parameter fine-tuning method and system based on end-cloud collaboration, by setting up the edge and cloud to train the large model, constitutes an end-cloud collaborative architecture to realize distributed processing of model training, and utilizes the real-time data processing capabilities of the edge and the powerful computing resources of the cloud to improve the efficiency of model fine-tuning; introduces an adaptive parameter optimization mechanism to dynamically adjust the update strategy of model parameters according to the performance of the model on different ends, thereby ensuring the efficient use of model parameters; through key exchange protocols, encryption operations, and data obfuscation technology, the security and privacy of data during model fine-tuning are ensured; the dynamic adjustment strategy of model fine-tuning adjusts the training load and resource allocation of the model in real time according to the performance feedback of each end in the end-cloud collaborative architecture, and can flexibly adapt to different operating environments and task requirements. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 This is an application environment diagram of an adaptive large model parameter fine-tuning method based on end-cloud collaboration in one embodiment;

[0050] Figure 2 It is a flowchart of a method for fine-tuning parameters of an adaptive large model based on end-cloud collaboration in one embodiment;

[0051] Figure 3 It is a structural block diagram of an adaptive large model parameter fine-tuning system based on end-cloud collaboration in one embodiment;

[0052] Figure 4 A schematic diagram of a network structure of a terminal-cloud collaborative architecture in an embodiment;

[0053] Figure 5 A schematic diagram of the structure of an adaptive parameter optimization mechanism in one embodiment;

[0054] Figure 6 A schematic diagram of a privacy protection data transmission protocol for ensuring data security and privacy in one embodiment;

[0055] Figure 7 A schematic diagram of the structure of a model fine-tuning dynamic adjustment strategy in an embodiment;

[0056] Figure 8 A schematic diagram of a program flow for fine-tuning large model parameters under a device-cloud collaborative architecture in one embodiment;

[0057] Fig. 9 A flowchart of an adaptive large model parameter fine-tuning method based on end-cloud collaboration applied to an intelligent customer service system in an embodiment;

[0058] Fig.10 A flowchart of an adaptive large model parameter fine-tuning method based on end-cloud collaboration applied to a smart home control system in an embodiment;

[0059] Fig.11 A flowchart of an adaptive large model parameter fine-tuning method based on end-cloud collaboration applied to an intelligent traffic management system in an embodiment;

[0060] Fig.12 The present invention is a flow chart of an adaptive large model parameter fine-tuning method based on end-cloud collaboration applied to an industrial automation control system in one embodiment. DETAILED DESCRIPTION

[0061] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0062] The adaptive large model parameter fine-tuning method based on end-cloud collaboration provided in the embodiment of the present application can be applied to Figure 1 In the application environment shown. Figure 1As shown, the application environment includes a cloud 110, an edge 120, and a control end 130, wherein the cloud 110, the edge 120, and the control end 130 are interconnected and communicated. The edge 120 collects the large model training data, pre-processes the large model training data, performs preliminary training of the large model based on the pre-processed data, and obtains a preliminary large model; the edge 120 identifies the key frames, and obtains the prediction results based on the key frames and the preliminary large model; the edge 120 exchanges encryption keys with the cloud 110 through the key exchange protocol; on the edge 120, a symmetric encryption algorithm is used to encrypt the transmission data to be transmitted to the cloud 110, and data obfuscation technology is introduced on the basis of data encryption; a message authentication code is used on the edge 120 to verify the integrity of the transmission data; the cloud 110 collects the performance feedback of each edge 120, evaluates the training load of each edge 120 based on the performance feedback, and dynamically adjusts each edge 120 based on the training load. resource allocation, and updating the model fine-tuning strategy of each edge terminal 120 based on the resource allocation result; receiving the encrypted transmission data transmitted from the edge terminal 120 through the cloud 110, decrypting the transmission data to obtain key frames and prediction results; according to the model fine-tuning strategy, fine-tuning the parameters of the preliminary large model based on the key frames, and optimizing the parameters of the fine-tuned preliminary large model based on the prediction results to obtain the optimized large model; the control terminal 130 evaluates the optimized large model according to the performance evaluation index, and triggers the parameter update according to the evaluation result; real-time monitoring and collection of the performance indicators of the optimized large model in the cloud 110 and the edge terminal 120, dynamically adjusting the model parameter update strategy according to the performance indicators, and determining the model parameter update frequency; adjusting the large model parameters based on the model parameter update strategy and the model parameter update frequency.

[0063] In one embodiment, Figure 2 As shown, a method for fine-tuning parameters of an adaptive large model based on end-cloud collaboration is provided, including the following steps:

[0064] Step 202, collect the big model training data through the edge end, pre-process the big model training data, perform preliminary training of the big model based on the pre-processed data, and obtain a preliminary big model; identify the key frames through the edge end, and obtain the prediction results based on the key frames and the preliminary big model.

[0065] In this embodiment, the edge and cloud are set up to train the large model, forming an end-cloud collaborative architecture to achieve distributed processing of model training. The end-cloud collaborative architecture allows the edge to use its real-time data processing capabilities to perform preliminary model training and transmit the intermediate results to the cloud; the cloud uses its powerful computing resources to further process and optimize these results; the end-cloud collaborative architecture not only improves the efficiency of model fine-tuning, but also provides a basis for data privacy protection.

[0066] Specifically, in this embodiment, the distributed processing of model training is realized through the end-cloud collaborative architecture, with preliminary model training performed on the edge and further optimization performed on the cloud. This architecture effectively utilizes the real-time data processing capabilities of the edge and the powerful computing resources of the cloud, significantly improving the efficiency of model fine-tuning. In the standard language model fine-tuning task, using the WikiText-103 dataset, the adaptive large model parameter fine-tuning method based on end-cloud collaboration provided in this embodiment saves about 10% of the training time compared to the LORA method, and improves the average accuracy by about 3%.

[0067] In the edge-cloud collaborative architecture, preliminary model training on the edge is one of the key steps to achieve efficient model fine-tuning. In this embodiment, a lightweight model training module can be deployed on the edge to perform preliminary model training using real-time data processing capabilities, thereby capturing the characteristics and patterns of local data.

[0068] Specifically, in one embodiment, a method for adaptive large model parameter fine-tuning based on end-cloud collaboration is provided, which may also include data processing and preliminary model training processes. The specific processes include: at the edge, performing denoising, normalization, and data enhancement preprocessing operations on the large model training data to obtain preprocessed data; performing feature extraction operations on the preprocessed data to extract model training features; and using the model training features to perform preliminary training of the large model to obtain a preliminary large model.

[0069] Among them, at the edge, the first step is to preprocess the collected data to improve the efficiency and accuracy of model training. The preprocessing steps include denoising, normalization, data enhancement, etc. The preprocessed data is passed through the feature extraction module to extract features useful for model training. Suppose the input data is Where n is the number of samples, d is the feature dimension of each sample, and feature extraction can be expressed as: F = φ(X); where F is the extracted feature and φ is the feature extraction operation.

[0070] Next, at the edge, using the extracted features F, the edge performs preliminary training on the model. The goal of the preliminary training is to minimize the difference between the model prediction and the actual label. Let the model parameter be θ and the loss function be Then the optimization problem of the initial training can be expressed as: Among them, Y is the actual label, θ edge are the parameters of the edge model after training.

[0071] In one embodiment, an adaptive large model parameter fine-tuning method based on end-cloud collaboration is provided, which may also include a process of determining key frames and generating prediction results. The specific process includes: a eigenvalue threshold is set at the edge to determine the eigenvalue corresponding to the model training feature; based on the eigenvalue threshold, the frame with the largest eigenvalue change is taken as the key frame; the edge obtains model parameters based on the preliminary large model, and generates prediction results based on the model parameters and key frames.

[0072] During the initial training process, the edge will identify key frames in the data, which contain the most important information for model prediction; key frames can be determined by setting a threshold to select the frame with the largest feature value change. The prediction result is based on the key frames and the model parameters θ obtained by training. edge To generate. Let the key frame set be K, then the prediction result It can be expressed as: Here, g is the prediction function of the model.

[0073] Step 204, exchange encryption keys between the edge and the cloud through a key exchange protocol; on the edge, use a symmetric encryption algorithm to encrypt the transmission data to be transmitted to the cloud, and introduce data obfuscation technology based on data encryption; on the edge, use a message authentication code to verify the integrity of the transmission data.

[0074] In the adaptive large model parameter fine-tuning strategy of end-cloud collaboration, data security and privacy are crucial. In order to ensure the security and privacy of data during the model fine-tuning process, a privacy-preserving data transmission protocol is adopted. The protocol combines encryption technology and data obfuscation technology to ensure the security of data during transmission, effectively preventing the risk of data leakage and unauthorized access, and aims to ensure the security and privacy of data during the model fine-tuning process.

[0075] First, before encrypted data transmission, it is necessary to securely exchange encryption keys between the edge and the cloud. In this embodiment, the Diffie-Hellman key exchange protocol is used to implement the encryption key exchange process: Assume that the edge and the cloud generate a key pair (d e ,D e ) and (d c ,D c ), then by exchanging public keys and calculating the common key, the encryption key can be obtained: K enc =f agreedkey (D e ,d c )=f agreedkey (D c ,d e ).

[0076] Data encryption is a key technology to ensure that data is not accessed by unauthorized persons during transmission. This paper uses a symmetric encryption algorithm to encrypt data transmitted from the edge to the cloud. Assume that the original data is D raw , the encryption key is K enc , then the encryption process can be expressed as: D enc =f encrypt (D raw ,K enc ).

[0077] In this embodiment, in order to further enhance the privacy of data, data obfuscation technology is introduced on the basis of data encryption. The obfuscation technology masks the actual content of the data by adding random noise. Assume that the obfuscation operation is f confusion (·), then the obfuscated data is expressed as: D confused =f confusion (D enc ).

[0078] In one embodiment, a method for fine-tuning parameters of an adaptive large model based on end-cloud collaboration is provided, which may also include a process for data integrity verification. The specific process includes: the edge end uses a message authentication code to verify the integrity of the transmitted data, and after the verification is passed, the transmitted data is sent to the cloud; the cloud end receives the transmitted data and uses the same key as the edge end to verify the transmitted data to obtain an integrity verification result.

[0079] Specifically, in the process of data transmission, in addition to confidentiality, the integrity of the data must also be guaranteed, and a message authentication code (MAC) is used to verify the integrity of the data. Suppose the key used by the sender to calculate the MAC is K mac , then the data integrity verification can be expressed as: MAC send =f MAC (D confused ,K mac ). After the cloud receives the data, it will use the same key for verification: ifMAC send ==MAC recv , then data is intact; MAC recv =f MAC (D confused ,K mac ).

[0080] By adopting a privacy-preserving data transmission protocol combined with encryption technology and data obfuscation technology, the security of data during transmission is guaranteed; the risk of data leakage and unauthorized access is effectively prevented, and user trust in the system is enhanced.

[0081] In step 206, the cloud collects performance feedback from each edge terminal, evaluates the training load of each edge terminal based on the performance feedback, dynamically adjusts the resource allocation of each edge terminal based on the training load, and updates the model fine-tuning strategy of each edge terminal based on the resource allocation result.

[0082] In order to further improve the efficiency and performance of model fine-tuning, a dynamic adjustment strategy for model fine-tuning is proposed, which can adjust the model's training load and resource allocation according to the model's performance and resource conditions on different terminals.

[0083] In one embodiment, a method for adaptive large model parameter fine-tuning based on end-cloud collaboration is provided, which may also include a process in which the cloud collects performance feedback from each edge end. The specific process includes: the cloud collects the training loss value, processing time, and model accuracy of the preliminary large model in each edge end; the cloud obtains a performance feedback set of each edge end based on the training loss value, processing time, and model accuracy.

[0084] First, the cloud needs to collect performance feedback from the edge, including indicators such as model accuracy, loss value, and processing time. The collection of performance feedback can be expressed by the formula: in, represents the performance feedback set, represents the accuracy of the edge model, Represents the training loss value of the edge model, T edge Indicates the processing time of the edge model.

[0085] In one embodiment, a method for adaptive large model parameter fine-tuning based on end-cloud collaboration is provided, which may also include a process in which the cloud evaluates the training load of each edge end. The specific process includes: the cloud determines the number of edge ends based on the connection relationship with each edge end, and obtains the training loss value and processing time of each edge end according to performance feedback; the cloud evaluates the training load of each edge end based on the training loss value and processing time.

[0086] Specifically, the cloud evaluates the training load of each edge based on the collected performance feedback. Training load evaluation can help the cloud allocate resources reasonably to avoid overloading or underloading of some edge. Training load evaluation can be expressed as: in, represents the average training load, n is the number of edges, and T edge,i They represent the training loss value and processing time of the i-th edge respectively.

[0087] The cloud can dynamically adjust the resource allocation of each edge based on the training load evaluation results. Resource allocation optimization aims to maximize the model training efficiency while ensuring that the resources of each edge are used reasonably. Resource allocation optimization can be performed using the formula: in, represents the optimized resource allocation plan, is the amount of resources allocated to the i-th edge.

[0088] In one embodiment, a method for adaptive large model parameter fine-tuning based on end-cloud collaboration is provided, which may also include a process in which the cloud updates the model fine-tuning strategy of each edge end. The specific process includes: the cloud adjusts the model learning rate based on resource allocation results and performance feedback; the cloud collects the updated model parameters of each edge end; the cloud updates the model fine-tuning strategy of each edge end according to the model learning rate and model parameters.

[0089] According to the resource allocation optimization results, the cloud updates the model fine-tuning strategy of each edge terminal, including adjusting the batch size, learning rate, number of training rounds and other parameters of model training. The model fine-tuning strategy update can be expressed as: in, represents the updated model parameters of the i-th edge, α new,i It is the learning rate adjusted based on the current resource allocation and performance feedback. Different from the initial model training, this is the update of edge model parameters triggered by the cloud after the initial training. The dynamic adjustment strategy of model fine-tuning is an iterative execution process. The cloud periodically collects performance feedback, evaluates training load, optimizes resource allocation, and updates the model fine-tuning strategy.

[0090] In this embodiment, the dynamic adjustment strategy of model fine-tuning is provided to adjust the training load and resource allocation of the model according to the performance feedback of each end in the end-cloud collaborative architecture. This strategy makes the model fine-tuning process more flexible and can adapt to different operating environments and task requirements. If applied in the field of smart home control, it can automatically adjust the working status of home appliances according to the user's living habits and preferences to improve the user experience. In other words, through the dynamic adjustment strategy of model fine-tuning, the efficiency and flexibility of parameter fine-tuning of large language models under the end-cloud collaborative architecture can be achieved; this strategy can adapt to different operating environments and task requirements, and significantly improves the performance and adaptability of model fine-tuning.

[0091] Step 208, receiving the encrypted transmission data transmitted from the edge through the cloud, decrypting the transmission data to obtain key frames and prediction results; according to the model fine-tuning strategy, fine-tuning the parameters of the preliminary large model based on the key frames, and optimizing the parameters of the fine-tuned preliminary large model based on the prediction results to obtain the optimized large model.

[0092] In one embodiment, a method for adaptive large model parameter fine-tuning based on end-cloud collaboration is provided, which may also include a process for further processing and optimizing the model. The specific process includes: the cloud determines the learning rate corresponding to the preliminary large model according to the model fine-tuning strategy, and obtains the gradient data of the loss function in the preliminary large model to the model parameters; the cloud uses key frames to fine-tune the parameters of the preliminary large model according to the learning rate and gradient data, and uses the prediction results as verification conditions to optimize the parameters of the fine-tuned preliminary large model; the cloud globally optimizes the preliminary large model according to the key frames and prediction results to obtain the optimized large model.

[0093] In this embodiment, after receiving the encrypted intermediate results, the cloud first decrypts them to obtain the original key frames and prediction results. The cloud model further processes and optimizes them based on this information. Specifically, it fine-tunes the model based on the key frames and uses the prediction results of the edge as verification conditions to optimize the model parameters. Let the cloud model parameters be θ cloud , then the fine-tuning process can be expressed as: Among them, α is the learning rate, is the loss function L cloud For θ cloud The cloud also performs global optimization of the model based on key frames and prediction results to ensure the generalization ability of the model on different edge terminals, which involves updating model parameters to minimize the overall loss function. This loss function takes into account the prediction errors of all edges:

[0094] The combination of preliminary model training on the edge and further processing and optimization on the cloud enables efficient fine-tuning of large language models under the edge-cloud collaborative architecture. This collaborative working mode fully utilizes the real-time data processing capabilities of the edge and the powerful computing resources of the cloud, improves the efficiency of model fine-tuning, and also provides a basis for data privacy protection.

[0095] Step 210, evaluate the optimized large model according to the performance evaluation indicators, and trigger parameter update according to the evaluation results; monitor and collect the performance indicators of the optimized large model in the cloud and edge in real time, dynamically adjust the model parameter update strategy according to the performance indicators, and determine the model parameter update frequency; adjust the large model parameters based on the model parameter update strategy and model parameter update frequency.

[0096] In one embodiment, a method for adaptive large model parameter fine-tuning based on end-cloud collaboration is provided, which may also include a process of introducing an adaptive parameter optimization mechanism, and the specific process includes: through a performance monitoring function, respectively monitoring and collecting the accuracy and loss value of the optimized large model on the cloud and edge, and using the accuracy and loss value as performance indicators of the large model; constructing a multi-factor trade-off model, inputting the performance indicator into the multi-factor trade-off model, and dynamically adjusting the model parameter update strategy through the multi-factor trade-off model; obtaining the performance change amount and performance change threshold of the optimized large model based on the performance indicator; determining the frequency control coefficient; determining the model parameter update frequency according to the performance change amount, performance change threshold, and frequency control coefficient.

[0097] The introduction of the adaptive parameter optimization mechanism aims to achieve efficient use and dynamic adjustment of model parameters. This mechanism intelligently adjusts the update strategy of model parameters according to the performance of the model on the edge and cloud to adapt to different data characteristics and task requirements.

[0098] In this embodiment, the introduced adaptive parameter optimization mechanism mainly includes triggering conditions for parameter updates, real-time performance monitoring, multi-factor trade-off updates, and frequency control of parameter updates.

[0099] Specifically, parameter updates are not performed periodically, but are determined based on the actual performance of the edge model. (usually recall rate), when the model performance does not meet the preset standard, the parameter update is triggered. The trigger condition can be expressed as: Among them, τ is the preset performance threshold.

[0100] Real-time performance monitoring is the basis of the adaptive parameter optimization mechanism. By monitoring key indicators such as the accuracy and loss value of the model at the edge and in the cloud, the performance of the model can be understood in real time. and They represent the performance indicators of edge and cloud models respectively. The real-time performance monitoring process can be expressed as: in, Represents a performance monitoring function, which is used to collect and analyze model performance data.

[0101] In addition to performance evaluation, the update of model parameters can also take into account multiple factors, including data characteristics, task requirements, and resource constraints. The adaptive parameter optimization mechanism can dynamically adjust the parameter update strategy by establishing a multi-factor trade-off model.

[0102] The frequency of parameter updates is determined by the rate of change of model performance and the degree of change of data characteristics. By controlling the update frequency, the stability of model training and adaptability to new data can be balanced. The update frequency control can be expressed as: Among them, fupdate is the update frequency, γ is the control coefficient, is the performance change, τ change is the threshold for performance change.

[0103] In this embodiment, based on the end-cloud collaborative architecture, an adaptive parameter optimization mechanism is designed to dynamically adjust the update strategy of model parameters according to the performance of the model on the edge and the cloud; through this dynamic adjustment, the model can more accurately adapt to different data characteristics and task requirements, thereby optimizing the utilization efficiency of model parameters. In other words, the introduced adaptive parameter optimization mechanism dynamically adjusts the update strategy of model parameters according to the performance of the model on different ends; this dynamic adjustment ensures the efficient utilization of model parameters and avoids waste of resources. And through the adaptive parameter optimization mechanism and the privacy-preserving data transmission protocol, unnecessary data transmission is reduced and communication overhead is reduced; through the end-cloud collaborative architecture and adaptive parameter optimization, the model can more accurately adapt to different data characteristics and task requirements.

[0104] It should be understood that, although the various steps in the above-mentioned flow chart are displayed in sequence according to the indication of the arrows, these steps are not necessarily executed in sequence according to the order indicated by the arrows. Unless there is a clear explanation in this article, the execution of these steps is not strictly limited in order, and these steps can be executed in other orders. Moreover, at least a part of the steps in the above-mentioned flow chart may include a plurality of sub-steps or a plurality of stages, and these sub-steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these sub-steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a part of the sub-steps or stages of other steps.

[0105] In one embodiment, Figure 1 , Figure 3 As shown, an adaptive large model parameter fine-tuning system based on end-cloud collaboration is provided, including a cloud end 110, an edge end 120 and a control end 130, wherein:

[0106] The edge end 120 is used to collect and preprocess the large model training data, perform preliminary training of the large model based on the preprocessed data, and obtain a preliminary large model; the key frames are identified through the edge end 120, and the prediction results are obtained based on the key frames and the preliminary large model;

[0107] The encryption key is exchanged between the edge terminal 120 and the cloud terminal 110 through a key exchange protocol; the edge terminal 120 is also used to: use a symmetric encryption algorithm to encrypt the transmission data to be transmitted to the cloud terminal 110, and introduce data obfuscation technology on the basis of data encryption; use a message authentication code on the edge terminal 120 to verify the integrity of the transmission data;

[0108] The cloud 110 is used to collect performance feedback from each edge terminal 120, evaluate the training load of each edge terminal 120 according to the performance feedback, dynamically adjust the resource allocation of each edge terminal 120 according to the training load, and update the model fine-tuning strategy of each edge terminal 120 based on the resource allocation result;

[0109] The cloud 110 is further used to receive the encrypted transmission data transmitted from the edge 120, decrypt the transmission data to obtain key frames and prediction results; fine-tune the parameters of the preliminary large model based on the key frames according to the model fine-tuning strategy, and optimize the parameters of the fine-tuned preliminary large model based on the prediction results to obtain an optimized large model;

[0110] The control end 130 is used to evaluate the optimized large model according to the performance evaluation indicators, and trigger parameter updates according to the evaluation results; monitor and collect the performance indicators of the optimized large model in the cloud 110 and the edge end 120 in real time, dynamically adjust the model parameter update strategy according to the performance indicators, and determine the model parameter update frequency; adjust the large model parameters based on the model parameter update strategy and model parameter update frequency.

[0111] In one embodiment, the edge end 120 is also used to perform preprocessing operations such as denoising, normalization, and data enhancement on the large model training data to obtain preprocessed data; perform feature extraction operations on the preprocessed data to extract model training features; and use the model training features to perform preliminary training of the large model to obtain a preliminary large model.

[0112] In one embodiment, the edge end 120 is also used to set a eigenvalue threshold to determine the eigenvalue corresponding to the model training feature; based on the eigenvalue threshold, the frame with the largest eigenvalue change is used as the key frame; the model parameters are obtained according to the preliminary large model, and the prediction results are generated based on the model parameters and the key frames.

[0113] In one embodiment, the edge terminal 120 is also used to use a message authentication code to verify the integrity of the transmitted data. After the verification is passed, the transmitted data is sent to the cloud 110; the cloud 110 is also used to receive the transmitted data and use the same key as the edge terminal 120 to verify the transmitted data to obtain an integrity verification result.

[0114] In one embodiment, the cloud 110 is also used to collect the training loss value, processing time, and model accuracy of the preliminary large model in each edge terminal 120; based on the training loss value, processing time, and model accuracy, a performance feedback set of each edge terminal 120 is obtained.

[0115] In one embodiment, the cloud 110 is also used to determine the number of edge terminals 120 based on the connection relationship with each edge terminal 120, and obtain the training loss value and processing time of each edge terminal 120 according to performance feedback; based on the training loss value and processing time, the training load of each edge terminal 120 is evaluated.

[0116] In one embodiment, the cloud 110 is also used to adjust the model learning rate based on resource allocation results and performance feedback; collect updated model parameters of each edge terminal 120; and update the model fine-tuning strategy of each edge terminal 120 according to the model learning rate and model parameters.

[0117] In one embodiment, the control end 130 is also used to monitor and collect the accuracy and loss value of the optimized large model on the cloud 110 and the edge end 120 through a performance monitoring function, and use the accuracy and loss value as performance indicators of the large model; construct a multi-factor trade-off model, input the performance indicator into the multi-factor trade-off model, and dynamically adjust the model parameter update strategy through the multi-factor trade-off model; obtain the performance change amount and performance change threshold of the optimized large model based on the performance indicator; determine the frequency control coefficient; determine the model parameter update frequency according to the performance change amount, performance change threshold, and frequency control coefficient.

[0118] In one embodiment, the cloud 110 is also used to determine the learning rate corresponding to the preliminary large model according to the model fine-tuning strategy, and obtain the gradient data of the loss function in the preliminary large model to the model parameters; according to the learning rate and gradient data, the preliminary large model is fine-tuned using key frames, and the prediction results are used as verification conditions to optimize the parameters of the fine-tuned preliminary large model; the preliminary large model is globally optimized according to the key frames and prediction results to obtain an optimized large model.

[0119] In one embodiment, a network structure of edge-cloud collaborative architecture is provided to achieve real-time data processing at the edge and deep model optimization at the cloud to improve the fine-tuning efficiency and accuracy of large language models. Figure 4 As shown, it can specifically include: data preprocessing and feature extraction, preliminary model training, key frame determination and prediction result generation, and further processing and optimization on the cloud, wherein: data preprocessing and feature extraction are responsible for preprocessing and feature extraction of the original data collected at the edge to enhance the model's ability to understand the data; preliminary model training mainly uses the extracted features to perform preliminary model training at the edge to minimize prediction errors; key frame determination and prediction result generation are responsible for identifying key frames from the preliminary training process and generating prediction results based on these key frames; further processing and optimization on the cloud specifically receives the intermediate results of the edge on the cloud, and performs further processing and optimization to improve the global performance of the model.

[0120] like Figure 4As shown in the figure, arrows represent data flow and information transmission in each part. The output of data preprocessing and feature extraction is initially trained through preliminary model training to generate key frames and prediction results; these results are sent to the cloud through a secure data transmission protocol for further processing and optimization, where the cloud model is fine-tuned and globally optimized based on this information; the optimization results of further processing and optimization in the cloud are fed back to data preprocessing and feature extraction and preliminary model training to guide further training and feature extraction on the end side.

[0121] like Figure 5 As shown, in one embodiment, the structure of the adaptive parameter optimization mechanism is provided to achieve efficient utilization and dynamic adjustment of model parameters. The structure of the adaptive parameter optimization mechanism is as follows Figure 5 As shown, it can specifically include: performance evaluation triggering, real-time performance monitoring, multi-factor trade-off updating, and parameter update frequency control, among which: performance evaluation triggering is responsible for performance evaluation of the edge model and decides whether to trigger parameter update according to preset standards. It is the entry point of the entire adaptive parameter optimization mechanism; real-time performance monitoring continuously monitors the performance indicators of edge and cloud models to provide real-time data support for parameter updates; multi-factor trade-off updating comprehensively considers performance evaluation results and other factors (such as data characteristics, task requirements, and resource constraints) to dynamically adjust the model parameter update strategy; parameter update frequency control specifically controls the frequency of parameter updates according to the rate of change of model performance and the degree of change of data characteristics to balance the stability and adaptability of model training.

[0122] like Figure 5 As shown in the figure, arrows are used to represent data flow and information transmission in each part. The evaluation results triggered by the performance evaluation trigger the multi-factor trade-off update strategy and affect the parameter update frequency control; the real-time performance monitoring data supports the decision-making process of the multi-factor trade-off update and feeds back to the parameter update frequency control to adjust the update frequency; the update strategy of the multi-factor trade-off update guides the parameter update frequency control to achieve dynamic adjustment of parameter updates; finally, the outputs of all modules are summarized to achieve efficient and accurate optimization of model parameters.

[0123] In one embodiment, a privacy protection data transmission protocol diagram for ensuring data security and privacy is provided, such as Figure 6As shown in the figure, the protocol combines encryption technology and data obfuscation technology to ensure the security of data during transmission. Specifically, it includes: data encryption module, data obfuscation unit, secure key exchange mechanism, and data integrity verification unit, among which: the data encryption module encrypts the data on the edge device to ensure the security of the data before transmission; the data obfuscation unit further obfuscates the encrypted data to increase the privacy of the data and prevent potential data leakage; the secure key exchange mechanism is responsible for securely exchanging the encryption keys required between the edge device and the cloud server; the data integrity verification unit verifies the data integrity after receiving the data in the cloud to ensure that the data has not been tampered with during transmission.

[0124] like Figure 6 As shown, the output of the edge device points to the data encryption module through an arrow, indicating the encryption process of the original data; the output of the data encryption module points to the data obfuscation unit, indicating the data obfuscation process; the secure key exchange mechanism is connected to the data encryption module and the data obfuscation unit through dotted arrows, indicating that the key exchange provides support for the encryption and obfuscation process; the output of the data obfuscation unit points to the cloud server through an arrow, indicating the transmission of encrypted and obfuscated data; after the cloud server receives the data, the data integrity is verified by the data integrity verification unit.

[0125] In one embodiment, a structure of a dynamic adjustment strategy for model fine-tuning is provided, aiming to achieve high efficiency and adaptability of model fine-tuning under a terminal-cloud collaborative architecture. Figure 7 As shown, the structure includes: performance feedback collection, training load evaluation, resource allocation optimization, and model fine-tuning strategy update, among which: performance feedback collection is responsible for collecting performance feedback from the edge, including indicators such as model accuracy, loss value, and processing time, to provide data support for dynamic adjustment strategies; training load evaluation evaluates the training load of each edge based on performance feedback to ensure the rationality of resource allocation; resource allocation optimization dynamically adjusts the resource allocation of each edge to maximize model training efficiency and ensure the rational use of resources; model fine-tuning strategy update updates the model fine-tuning strategy of each edge based on the resource allocation optimization results, including parameters such as batch size, learning rate, and number of training rounds.

[0126] like Figure 7 As shown in the figure, each part represents data flow and information transmission through arrows; the output of performance feedback collection drives training load evaluation, guides resource allocation optimization, and triggers model fine-tuning strategy update; the result of training load evaluation affects the decision of resource allocation optimization, and is fed back to the model fine-tuning strategy update to adjust the fine-tuning strategy; the resource allocation optimization plan guides the model fine-tuning strategy update; finally, the output of all modules is summarized to achieve dynamic adjustment of model fine-tuning.

[0127] In one embodiment, Figure 8 As shown in the figure, the program flow of fine-tuning large model parameters under the end-cloud collaborative architecture is provided. The flowchart shows the complete process from data preprocessing to model training, performance evaluation, parameter optimization, and the final model fusion and update. Specifically, it mainly includes: data preprocessing steps, local model training steps, performance evaluation steps, cloud model optimization steps, parameter optimization strategy decision steps, model fusion and update steps, among which, data preprocessing steps: executed on the edge device, responsible for data cleaning, standardization and feature engineering; local model training steps: performed on the edge device, using local data to perform preliminary training on the model; performance evaluation steps: real-time monitoring and evaluation of model performance to provide a basis for subsequent optimization; cloud model optimization steps: performed on the cloud server, further optimizing the model based on the performance feedback of the edge; parameter optimization strategy decision steps: dynamically determine the parameter update strategy based on the performance evaluation results and resource status; model fusion and update steps: executed on the cloud server, responsible for fusing and updating the optimized model parameters to the edge device.

[0128] like Figure 8 As shown, the data preprocessing step points to the local model training step, indicating that local model training is performed after data preprocessing; the local model training step points to the performance evaluation step, indicating that performance evaluation is performed after training; the outputs of the performance evaluation step point to the cloud model optimization step and the parameter optimization strategy decision step respectively, indicating that the performance evaluation results are used for both cloud model optimization and parameter optimization strategy decision; the outputs of the cloud model optimization step and the parameter optimization strategy decision step point to the model fusion and update step, indicating that the optimized model parameters are fused and updated; the output of the model fusion and update step points to the edge device, indicating that the updated model parameters are transmitted back to the edge device.

[0129] In one embodiment, for example, this embodiment applies the adaptive large model parameter fine-tuning method based on end-cloud collaboration to the intelligent customer service system to improve the response speed and service quality of the customer service system. Fig. 9 The specific implementation process is as follows:

[0130] User request reception: The intelligent customer service system receives user service requests in various forms such as text, voice or image;

[0131] Request preprocessing: The system preprocesses user requests, including language recognition, semantic analysis, etc., to extract key information;

[0132] End-cloud collaborative processing: The edge performs preliminary request processing, such as simple question-answer matching, and transmits complex requests to the cloud;

[0133] Adaptive parameter fine-tuning: The cloud dynamically adjusts model parameters based on the type of user requests and historical data to optimize the response strategy;

[0134] Privacy-preserving data transmission: During data transmission, privacy-preserving protocols are used to ensure the security of user data.

[0135] Intelligent response generation: The cloud generates intelligent responses based on the fine-tuned model and sends them back to the edge;

[0136] Response feedback and optimization: The intelligent customer service system adjusts the response strategy based on user feedback and performs iterative optimization of the model.

[0137] In one embodiment, this embodiment applies the adaptive large model parameter fine-tuning method based on end-cloud collaboration to the smart home control system to achieve intelligent management of the home environment. Fig.10 The specific implementation process is as follows:

[0138] Environmental monitoring data collection: Smart home systems collect various data about the home environment through sensors, such as temperature, humidity, light, etc.

[0139] Data preprocessing: The system preprocesses the collected environmental data, including data cleaning and standardization, to ensure data quality;

[0140] End-cloud collaborative decision-making: The edge makes preliminary environmental control decisions based on real-time data and transmits complex decisions to the cloud;

[0141] Adaptive parameter optimization: The cloud dynamically adjusts control parameters based on historical data and user preferences to optimize the home environment;

[0142] Privacy-preserving data transmission: During data transmission, privacy-preserving protocols are used to ensure the security of family data;

[0143] Intelligent control execution: The cloud generates intelligent control instructions and sends them to the edge for execution, such as adjusting air conditioning temperature, lighting brightness, etc.

[0144] System performance evaluation and optimization: The smart home system conducts performance evaluation based on execution results and user feedback, and continuously optimizes control strategies.

[0145] In one embodiment, this embodiment applies the adaptive large model parameter fine-tuning method based on end-cloud collaboration to the intelligent traffic management system to improve the efficiency and safety of traffic management. Fig.11 The specific implementation process is as follows:

[0146] Traffic data collection: Real-time collection of traffic flow and vehicle information data through sensors and cameras deployed at major traffic arteries;

[0147] Data preprocessing: preprocess the collected raw traffic data, including data cleaning, format conversion, etc., to ensure data quality;

[0148] End-cloud collaborative analysis: The edge performs preliminary analysis on real-time traffic data, identifies events such as traffic jams and accidents, and transmits key information to the cloud;

[0149] Adaptive parameter fine-tuning: The cloud dynamically adjusts the parameters of the traffic management model based on the traffic event information received to optimize the traffic flow control strategy;

[0150] Privacy-preserving data transmission: During the data transmission process, a privacy-preserving protocol is used to ensure the privacy information security of traffic participants;

[0151] Traffic control command generation: Traffic control commands, such as traffic light adjustment and vehicle diversion, are generated in the cloud and sent to the edge for execution;

[0152] System performance evaluation and optimization: Evaluate system performance based on traffic conditions and control effects, and continuously optimize models and control strategies.

[0153] In one embodiment, this embodiment applies the adaptive large model parameter fine-tuning method based on end-cloud collaboration to industrial automation control systems to improve production efficiency and product quality. Fig.12 The specific implementation process is as follows:

[0154] Production data collection: Real-time collection of production process data through sensors and monitoring equipment deployed on the production line;

[0155] Data preprocessing: preprocess the collected production data, including data cleaning and standardization, to ensure data quality;

[0156] End-cloud collaborative decision-making: The edge performs preliminary analysis on real-time production data, identifies production anomalies, and exchanges information with the cloud.

[0157] Adaptive parameter fine-tuning: The cloud dynamically adjusts the parameters of the production control model based on the analysis results of the edge to optimize the production process;

[0158] Privacy-preserving data transmission: During the data transmission process, privacy-preserving protocols are used to ensure the security of enterprise production data;

[0159] Production control command generation: The cloud generates production control commands, such as machine start and stop, speed adjustment, etc., and sends them to the edge for execution;

[0160] System performance evaluation and optimization: Evaluate system performance based on production efficiency and product quality, and continuously optimize models and control strategies.

[0161] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0162] The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the attached claims.

Claims

1. An adaptive large model parameter fine-tuning method based on end-cloud collaboration, characterized in that: The method comprises: Collecting large model training data through the edge end, preprocessing the large model training data, performing preliminary training of the large model according to the preprocessed data, and obtaining a preliminary large model; identifying key frames through the edge end, and obtaining prediction results based on the key frames and the preliminary large model; The encryption key is exchanged between the edge and the cloud through a key exchange protocol; on the edge, a symmetric encryption algorithm is used to encrypt the transmission data to be transmitted to the cloud, and data obfuscation technology is introduced on the basis of data encryption; on the edge, a message authentication code is used to verify the integrity of the transmission data; The cloud collects performance feedback from each of the edge terminals, evaluates the training load of each of the edge terminals according to the performance feedback, dynamically adjusts resource allocation of each of the edge terminals according to the training load, and updates the model fine-tuning strategy of each of the edge terminals based on the resource allocation result; Receiving the encrypted transmission data transmitted from the edge end through the cloud, decrypting the transmission data to obtain the key frame and the prediction result; fine-tuning the parameters of the preliminary large model based on the key frame according to the model fine-tuning strategy, and optimizing the parameters of the fine-tuned preliminary large model based on the prediction result to obtain an optimized large model; The optimized large model is evaluated according to performance evaluation indicators, and parameter update is triggered according to the evaluation results; the performance indicators of the optimized large model in the cloud and the edge are collected and monitored in real time, and the model parameter update strategy is dynamically adjusted according to the performance indicators to determine the model parameter update frequency; the large model parameters are adjusted based on the model parameter update strategy and the model parameter update frequency.

2. The method for fine-tuning parameters of an adaptive large model based on end-cloud collaboration according to claim 1 is characterized in that: The edge end preprocesses the large model training data, performs preliminary training of the large model according to the preprocessed data, and obtains a preliminary large model, including: At the edge end, performing preprocessing operations of denoising, normalization, and data enhancement on the large model training data to obtain preprocessed data; Performing a feature extraction operation on the preprocessed data to extract model training features; The model training features are used to perform preliminary training of the large model to obtain a preliminary large model.

3. The method for fine-tuning parameters of an adaptive large model based on end-cloud collaboration according to claim 2 is characterized in that: Identifying key frames through the edge end, and obtaining prediction results based on the key frames and the preliminary large model, including: A characteristic value threshold is set at the edge end to determine the characteristic value corresponding to the model training feature; Based on the eigenvalue threshold, taking the frame with the largest eigenvalue change as the key frame; The edge end obtains model parameters according to the preliminary large model, and generates prediction results based on the model parameters and the key frames.

4. The method for fine-tuning parameters of an adaptive large model based on end-cloud collaboration according to claim 1 is characterized in that: The method further comprises: The edge end uses a message authentication code to verify the integrity of the transmission data, and after the verification is passed, sends the transmission data to the cloud; The cloud receives the transmission data and verifies the transmission data using the same key as the edge to obtain an integrity verification result.

5. The method for fine-tuning parameters of an adaptive large model based on end-cloud collaboration according to claim 1 is characterized in that: The cloud collects performance feedback from each edge terminal, including: The cloud end collects the training loss value, processing time, and model accuracy of the preliminary large model in each of the edge ends; The cloud obtains a performance feedback set of each of the edge terminals based on the training loss value, processing time, and model accuracy.

6. The method for fine-tuning parameters of an adaptive large model based on end-cloud collaboration according to claim 5 is characterized in that: The cloud evaluates the training load of each edge terminal according to the performance feedback, including: The cloud determines the number of edge terminals based on the connection relationship with each of the edge terminals, and obtains the training loss value and processing time of each of the edge terminals according to the performance feedback; The cloud end evaluates the training load of each of the edge ends based on the training loss value and the processing time.

7. The method for fine-tuning parameters of an adaptive large model based on end-cloud collaboration according to claim 6 is characterized in that: The cloud updates the model fine-tuning strategy of each edge terminal based on the resource allocation result, including: The cloud end adjusts the model learning rate based on the resource allocation result and the performance feedback; The cloud end collects the updated model parameters of each of the edge ends; The cloud end updates the model fine-tuning strategy of each of the edge ends according to the model learning rate and the model parameters.

8. The method for fine-tuning parameters of an adaptive large model based on end-cloud collaboration according to claim 1 is characterized in that: The real-time monitoring collects the performance indicators of the optimized large model in the cloud and the edge, dynamically adjusts the model parameter update strategy according to the performance indicators, and determines the model parameter update frequency, including: Through the performance monitoring function, the accuracy and loss value of the optimized large model on the cloud and the edge are monitored and collected respectively, and the accuracy and loss value are used as performance indicators of the large model; Constructing a multi-factor trade-off model, inputting the performance indicator into the multi-factor trade-off model, and dynamically adjusting the model parameter update strategy through the multi-factor trade-off model; Based on the performance index, obtain the performance change amount and performance change threshold of the optimized large model; determine the frequency control coefficient; The model parameter update frequency is determined according to the performance change amount, the performance change threshold, and the frequency control coefficient.

9. The method for fine-tuning parameters of an adaptive large model based on end-cloud collaboration according to claim 1 is characterized in that: The cloud fine-tunes the parameters of the preliminary large model based on the key frame according to the model fine-tuning strategy, and optimizes the parameters of the fine-tuned preliminary large model based on the prediction result to obtain an optimized large model, including: The cloud determines the learning rate corresponding to the preliminary large model according to the model fine-tuning strategy, and obtains the gradient data of the loss function to the model parameters in the preliminary large model; The cloud uses the key frame to fine-tune the parameters of the preliminary large model according to the learning rate and gradient data, and uses the prediction result as a verification condition to optimize the parameters of the fine-tuned preliminary large model; The cloud end globally optimizes the preliminary large model according to the key frames and the prediction results to obtain an optimized large model.

10. An adaptive large model parameter fine-tuning system based on end-cloud collaboration, characterized in that: The system comprises: The edge end is used to collect large model training data, pre-process the large model training data, perform preliminary training of the large model according to the pre-processed data, and obtain a preliminary large model; identify key frames through the edge end, and obtain prediction results based on the key frames and the preliminary large model; The encryption key is exchanged between the edge and the cloud through a key exchange protocol; the edge is also used to: use a symmetric encryption algorithm to encrypt the transmission data to be transmitted to the cloud, and introduce data obfuscation technology on the basis of data encryption; use a message authentication code on the edge to verify the integrity of the transmission data; The cloud end is used to collect performance feedback of each of the edge ends, evaluate the training load of each of the edge ends according to the performance feedback, dynamically adjust the resource allocation of each of the edge ends according to the training load, and update the model fine-tuning strategy of each of the edge ends based on the resource allocation result; The cloud end is further used to receive the encrypted transmission data transmitted from the edge end, decrypt the transmission data to obtain the key frame and the prediction result; according to the model fine-tuning strategy, fine-tune the parameters of the preliminary large model based on the key frame, and optimize the parameters of the fine-tuned preliminary large model based on the prediction result to obtain an optimized large model; The control end is used to evaluate the optimized large model according to the performance evaluation index, and trigger parameter update according to the evaluation result; monitor and collect the performance index of the optimized large model in the cloud and the edge in real time, dynamically adjust the model parameter update strategy according to the performance index, and determine the model parameter update frequency; adjust the large model parameters based on the model parameter update strategy and model parameter update frequency.

Images