Federal cooperative training optimization method and system for project achievement reward classification

Abstract

The invention relates to a federal cooperative training optimization method and system for project achievement reward classification, a client extracts image features of a project achievement by using an image encoder and a text encoder of a CLIP model, and converts a reward classification label into a natural language template to generate a semantic vector, thereby realizing unified spatial alignment of visual and text features; the attention adapter screens task related features; knowledge distillation is adopted to migrate knowledge of the teacher model to the student model, and classification loss and distillation loss are jointly optimized; the server aggregates only the adapter parameters of the clients to update the global model. Compared with the prior art, the classification accuracy is improved by more than 5%, the cross-client generalization performance is improved by 8%, meanwhile, the convergence speed is increased by 20%, and the federal learning efficiency and the model performance are remarkably improved; scenes such as project achievement reward classification can be effectively supported, privacy and confidentiality of local client data are ensured, and low-delay and high-precision distributed reasoning is realized.

Description

A Federated Collaborative Training Optimization Method and System for Project Outcome Reward Classification Technical Field

[0001] This invention relates to the field of project achievement reward classification technology, and in particular to a federated collaborative training optimization method and system for project achievement reward classification. Background Technology

[0002] Accurately classifying and rewarding the achievements of enterprises in the process of technological innovation projects presents numerous challenges. Due to privacy and security restrictions, project data from various departments and teams cannot be centrally shared, leading to a skewed distribution of project results and consequently affecting the generalization performance of models. Currently, many machine learning tasks follow the principle of "data locality," meaning that data from various participants is distributed across different devices or institutions, and the original data cannot be directly shared. How to achieve efficient collaborative training while protecting privacy has become one of the core challenges in the field of artificial intelligence.

[0003] Federated Learning (FL), as a distributed machine learning paradigm, optimizes the global model by aggregating local model updates from multiple clients, effectively alleviating the data silo problem. With the strict implementation of data privacy regulations (such as China's Data Security Law), the demand for federated learning in critical infrastructure is becoming increasingly urgent. For example, in the task of classifying enterprise science and technology project achievement awards, the contradiction between real-time requirements and data sensitivity is particularly prominent, necessitating a solution with low communication overhead and high robustness.

[0004] Traditional methods (such as gradient compression) can reduce transmission volume, but high compression ratios (e.g., above 90%) can lead to a significant drop in model performance. Lightweight adapter solutions based on CLIP (contrastive language-image pre-trained) models can reduce communication costs, but lack a knowledge distillation mechanism, failing to fully utilize the potential value of the pre-trained model. Traditional federated learning (such as FedAvg and FedProx) and simple CLIP + federated learning (FedCLIP) have significant shortcomings in non-IID data scenarios (where data distribution varies greatly among clients). Client-side local models are prone to overfitting local data distributions, leading to decreased generalization performance after global aggregation. Furthermore, their cross-modal feature alignment capability is weak, making it difficult to constrain the consistency of image features through textual semantics, further exacerbating the negative impact of distribution offset.

[0005] How to combine the advantages of knowledge distillation and CLIP models to improve the efficiency and generalization ability of federated learning, as well as its robustness to non-IID data, while ensuring privacy, has become a key issue that urgently needs to be addressed. Summary of the Invention

[0006] The purpose of this invention is to provide a federated collaborative training optimization method and system for project achievement reward classification, which reduces communication and computing costs while improving model generalization ability and personalized adaptation performance, thereby realizing project achievement reward classification and distributed collaborative reasoning.

[0007] To achieve the above objectives, the technical solution of the present invention is as follows: A project outcome reward classification method based on knowledge distillation and federated learning, comprising the following steps: S1. The client participating in the training uses the image encoder of the pre-trained CLIP model to extract visual features of the project outcome image, and uses its text encoder to convert the reward classification label into a natural language template to generate a semantic vector, so that the visual features and the semantic vector are aligned in a unified semantic space; S2. Based on the visual features and semantic vectors extracted in step S1, a subset of task discriminative features related to the project outcome reward category is selected through the attention adapter deployed on the client; S3. Using the feature subset as input, the CLIP model is defined as the teacher model, the local classifier deployed on the client is defined as the student model, knowledge distillation is used to transfer the knowledge of the teacher model to the student model, and the classification loss and knowledge distillation loss are jointly optimized to complete the local model training including the attention adapter parameters; S4. The server only aggregates the attention adapter parameters in the clients participating in the training, processes and generates new attention adapter parameters and distributes them to all clients.

[0008] Furthermore, in S2, the selection of feature subsets is based on the cross-modal cosine similarity normalization result. The selection weights are dynamically adjusted by the cosine distance between visual features and text semantic vectors. The formula for calculating the cosine similarity is as follows: In the formula, The visual features output by the CLIP model image encoder are represented by T, and the semantic vector output by the CLIP model text encoder is represented by cosine similarity. As an input bias of the attention adapter.

[0009] Furthermore, in S2, the attention adapter includes a linear input layer, a cross-modal fusion layer, a gated activation function GELU-Gate, a distillation-guided Softmax, and a linear output layer.

[0010] Furthermore, in S3, the temperature parameter t=2 for knowledge distillation, the balance parameter between distillation loss and classification loss is 0.5, and gradient pruning mechanism is used to avoid gradient conflict between distillation loss and task loss, and the gradient magnitude of distillation loss is adjusted through knowledge distillation loss function.

[0011] Furthermore, in S3, the number of training epochs E=1, and the total number of communication epochs R=200; the Adam optimizer is used, with a learning rate range of 5× Up to 5× .

[0012] Furthermore, in S4, the server aggregation strategy includes: selecting 50% of the clients to participate in training in each round, and weighting the attention adapter parameters uploaded by each participating client based on the number of local samples to generate new attention adapter parameters.

[0013] A project outcome reward classification system based on knowledge distillation and federated learning for the above method includes: multiple clients configured to execute S1–S3.

[0014] The server is configured to run S4.

[0015] Furthermore, the client includes: a multimodal feature extraction module, used to extract visual features of project result images using the image encoder of the pre-trained CLIP model, and use its text encoder to convert reward classification labels into natural language templates to generate semantic vectors, so that visual features and semantic vectors are aligned in a unified semantic space; an attention adapter module, used to dynamically generate attention weights and filter task discriminative feature subsets related to the project result reward category based on the cross-modal cosine similarity between the visual features and semantic vectors; and a local training module, used to take the feature subset as input, define the CLIP model as the teacher model, define the local classifier deployed on the client as the student model, use knowledge distillation to transfer knowledge from the teacher model to the student model, jointly optimize the classification loss and knowledge distillation loss, and complete the local model training including the attention adapter parameters.

[0016] Furthermore, the server includes: a global model distribution module for distributing initial or updated attention adapter parameters to each client; a parameter aggregation module for receiving attention adapter parameters uploaded by clients participating in this round of training, and performing a weighted average based on the number of samples from each client to generate updated attention adapter parameters; and a model synchronization module for sending the updated attention adapter parameters to the clients to support the next round of federated training iterations.

[0017] The advantages of this invention are: 1. It achieves efficient collaborative training while ensuring data privacy. Each enterprise client does not need to upload sensitive original project result images or labels, but only needs to upload lightweight attention adapter parameters (accounting for less than 5% of the entire model) to the server, which greatly reduces communication overhead (by about 90%), while meeting the data localization requirements of regulations such as the Data Security Law.

[0018] 2. Significantly improves model performance and generalization ability. By freezing the pre-trained CLIP model as the teacher, and combining cross-modal semantic alignment (converting "company-level / provincial-level / industry association awards" into natural language templates) and knowledge distillation mechanisms, the student model can effectively inherit CLIP's multimodal prior knowledge. Experiments show that compared with existing methods, this approach improves classification accuracy by more than 5% on non-independent and identically distributed (Non-IID) data, cross-client generalization performance by 8%, and robustness by 12%.

[0019] 3. Balancing training efficiency and resource conservation. Employing a local training strategy with E=1 and a communication design with R=200 rounds, coupled with adaptive learning rate and gradient pruning, not only accelerates convergence (reducing the number of rounds required to reach 90% accuracy by 20%), but also avoids local overfitting. A lightweight attention adapter dynamically selects task-relevant high-frequency features, further reducing computational burden.

[0020] 4. The system architecture is flexible and highly scalable. The server only aggregates adapter parameters, while the client retains full inference capabilities, making it suitable for various enterprise distributed scenarios. This solution not only solves the problems of data silos and distribution offsets in the classification of science and technology project rewards, but also provides a general technical paradigm for multimodal federated learning in other privacy-sensitive fields. Attached Figure Description

[0021] Figure 1 is a schematic diagram of the federated collaborative training optimization method for project achievement reward classification in the embodiment; Figure 2 is a schematic diagram of client-side local training and server-side parameter aggregation in the embodiment; Figure 3 is the calculation process of knowledge distillation loss in the embodiment; Figure 4 is the feature extraction and fusion process of encoder and attention adapter in CLIP model in the embodiment; Figure 5 is a comparison of the generalization performance of the method in the embodiment with FedAvg, FedProx, FedCLIP and other methods; Figure 6 is a performance comparison of the method (CLIP + federated learning + knowledge distillation) and FedCLIP (CLIP + federated learning) in the enterprise science and technology project reward classification task in the embodiment. Detailed Implementation

[0022] 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.

[0023] This embodiment first proposes a client-side local training and server-side parameter aggregation process for a federated collaborative training optimization method for project achievement reward classification, as shown in Figure 2. This process includes: T1. Client initialization: downloading the global model (including the frozen CLIP encoder and initial attention adapter parameters); T2. Local client updating attention adapter parameters and training, including feature extraction, feature selection, and knowledge distillation; T3. Parameter upload: uploading only a subset of attention adapter parameters from the local client to the server.

[0024] In Figure 2, Di represents the local data of the i-th client, Wi represents the attention adapter parameters of the i-th client, and Ui represents the attention adapter parameters uploaded by the i-th client to the server. Preferably, this embodiment uses knowledge distillation and the CLIP model to obtain key features of the local data using the local client adapter, trains the local model, obtains a better global aggregation model, and trains and updates the local model through a federated learning framework to improve the model's generalization and personalization capabilities.

[0025] Specifically, this embodiment provides a federated collaborative training optimization method for classifying project achievement awards, as shown in Figure 1, which mainly includes the following steps: S1. The client participating in the training uses the image encoder and text encoder of the pre-trained CLIP model to extract image features of enterprise science and technology project achievement data through cross-modal contrastive learning, and converts the award classification labels (company-level, provincial-level, and industry association awards) into natural language templates to generate semantic vectors, achieving unified spatial alignment of visual and text features, and providing high-quality, semantically rich initial features for subsequent tasks; S2. Based on the visual features and semantic vectors extracted in step S1, a lightweight attention adapter deployed on the client filters high-frequency feature subsets of company-level awards, provincial-level awards, and industry association awards, removing redundant or noisy features (such as backgrounds, irrelevant icons, etc.) that are unrelated to the award classification, significantly reducing... The subsequent model input dimension improves training efficiency and generalization ability; S3. Using a feature subset as input, the CLIP model is defined as the teacher model, and the local classifier deployed on the client is defined as the student model. Knowledge distillation is used to transfer the knowledge of the teacher model to the student model, and the classification loss and knowledge distillation loss are jointly optimized to complete the local model training including the attention adapter parameters, thereby improving the local classification accuracy; S4. The server only aggregates the attention adapter parameters of the clients participating in the training, processes and generates new attention adapter parameters, and distributes them to all clients to update the global model, realizing knowledge collaboration under privacy protection and significantly reducing communication overhead.

[0026] Through steps S1 to S4, under the constraints of not sharing data, not updating large models, and low communication costs, the classification accuracy, personalization adaptability, and cross-client generalization ability of enterprise science and technology project achievement rewards are significantly improved.

[0027] In step S2, the selection of high-frequency feature subsets is based on the cross-modal cosine similarity normalization result. The selection weights are dynamically adjusted by the cosine distance between visual features and text semantic vectors. The formula for calculating the cosine similarity is as follows: In the formula, The visual features output by the CLIP model image encoder are represented by T, and the semantic vector output by the CLIP model text encoder is represented by cosine similarity. As an input bias of the attention adapter, it enhances the task relevance of cross-modal features.

[0028] In step S2, the attention adapter includes a linear input layer, a cross-modal fusion layer, a gated activation function GELU-Gate, a distillation-guided Softmax layer, and a linear output layer. The linear input layer maps the feature dimensions to a low-dimensional space; the cross-modal fusion layer dynamically adjusts feature weights based on the association between images and text, making the selected high-frequency features more aligned with the cross-modal task requirements of enterprise technology project reward classification; the gated activation function GELU-Gate dynamically suppresses noisy features through a gating mechanism; and the Softmax attention weights guide the adapter's attention weights with the global knowledge of the teacher model, enhancing the synergy between the selected high-frequency features and the distillation process.

[0029] Referring to Figure 4, the feature extraction and fusion process of the CLIP encoder and adapter is illustrated. In feature extraction, the CLIP model adopts a dual-tower architecture, including an image encoder and a text encoder. Through contrastive learning, images and text are mapped to a unified semantic space. The image encoder extracts fixed features I, and the text encoder generates semantic features T. The adapter is a lightweight module used to dynamically filter task-relevant subsets from the fixed features of CLIP, reducing communication and computational overhead. The adapter generates attention weights through two-layer linear transformation, Tanh activation function, and Softmax. Key features are extracted from the input image features through attention weights, retaining only the parts most relevant to the current task. The CLIP model acts as the teacher model, providing soft labels, while the adapter acts as the student model, aligning the real labels with the task loss and the soft labels with the distillation loss to align with the teacher output.

[0030] The training of the attention adapter includes optimizing the attention weight vector through backpropagation to dynamically adapt it to non-independent and identically distributed data; the threshold for high-frequency feature selection adopts an adaptive dynamic adjustment mechanism: the initial threshold is set to 0.7 (based on the feature importance distribution of the top 10% of samples in the training set), and after each round of local training, the threshold is updated according to the mean (μ) of the attention weights of the current batch of features: threshold = max (0.5, μ-0.1), ensuring that the selected features contain core information while avoiding overfitting.

[0031] In step S3, the temperature parameter t=2 for knowledge distillation, the balance parameter between distillation loss and classification loss is set to 0.5, and a gradient pruning mechanism (pruning threshold of 1.0) is used to avoid gradient conflicts between distillation loss and task loss, achieving efficient transfer of multimodal knowledge from the teacher model (CLIP) to the student model (local classifier). Compared to using CLIP or knowledge distillation alone, the classification accuracy is improved by more than 5%, and the gradient magnitude of distillation loss is adjusted through the knowledge distillation loss function. In the formula, This represents the knowledge distillation loss; the temperature parameter t controls the smoothness of the soft label, and t=2 can balance the label discrimination and knowledge transfer (experiments show that the soft label is over-sharpened when t=1 and the discrimination is insufficient when t=3). The term is to counteract the gradient decay of the KL divergence at high temperatures (derivation process: assume...). When t increases, and The differences were magnified. (It can keep the gradient magnitude stable within the effective range). It is the Kullback-Leibler divergence (KL divergence), used to measure the difference between two probability distributions; This represents the probability distribution of the teacher model's output; This represents the probability distribution of the student model's output.

[0032] Referring to Figure 3, the calculation process of knowledge distillation loss for the federated collaborative training optimization method used for project achievement reward classification is shown. The input consists of original image data X and corresponding text labels y, and feature extraction is performed. The CLIP model, acting as the teacher model, provides semantically rich feature representations and generates soft labels. The adapter, acting as the student model, is a lightweight attention module containing two layers of linear transformation + Tanh activation + Softmax, using attention weights to filter features. During local client training, the task loss aligns the student model's predictions with the true labels, and the distillation loss aligns the soft labels of the student model with those of the teacher model. The client optimizes the distillation loss through backpropagation to improve the student model's ability to absorb teacher knowledge.

[0033] In step S3, the hyperparameter settings for local training include: local training epochs E=1, total communication epochs R=200; the Adam optimizer is used, and the learning rate range is 5× Up to 5× Learning rate range 5× Up to 5× The selection criteria are as follows: Validated through grid search, this range balances the convergence speed of the adapter parameters with the retention of CLIP pre-trained knowledge (learning rate > 5×). At this time, overfitting of the adapter to local data leads to decreased generalization; learning rate <5× (At that time, training convergence is slow).

[0034] In S4, the server aggregation strategy includes: randomly selecting 50% of clients to participate in each round of training, and weighting the adapter parameters according to the sample size of each participating client; the number of adapter parameters is only 5% of the total model. Combined with the cross-modal feature alignment mechanism, the communication overhead is reduced by 90%, while the model generalization performance (cross-client classification accuracy) is improved by 8%.

[0035] This method further selects the features extracted by the encoder based on the adapter in the local client to select more important data features, making the model more expressive. Uploading the adapter parameters helps protect the privacy of the local client data, and uploading only the adapter parameters reduces the number of parameters in the upload and download process, alleviating the communication load problem.

[0036] This embodiment also provides experimental testing of a federated collaborative training optimization method for classifying project outcome rewards.

[0037] Referring to Figure 4, the client collects image data. The image encoder uses a pre-trained ViT-B / 32 model (Vision Transformer) to extract fixed features. The text encoder converts the project achievement reward classification labels (such as "company-level reward", "provincial-level reward", "industry association reward") into natural language templates. The task loss uses cross-entropy to align with the true labels, and the distillation loss uses KL to align with CLIP soft labels. The adapter generates attention weights through two layers of linear transformation, Tanh activation function and Softmax, and extracts key features through the attention weights.

[0038] Referring to Figure 5, the generalization ability is measured by the accuracy of the local client model that was not involved in training. For the relevant dataset, each client's subset is split into 60% training, 20% validation, and 20% testing. The Adam optimizer is used during training, with a learning rate ranging from 5×... Up to 5× The total number of communication rounds was R=200. In the hyperparameter settings, the temperature parameter t=2.0 was used to balance the smoothness of the soft labels, and the task loss and distillation loss balancing parameter was 0.5. The experimental environment was PyTorch, and the hardware was an NVIDIA T4 GPU. The experiments were compared with traditional federated learning methods and CLIP-based methods. Traditional federated learning methods included FedAvg, which aggregates all model parameters, and FedProx, which introduces proximal terms to alleviate client drift. CLIP-based methods included FedCLIP, which trains an adapter based on fixed features. Evaluation metrics included generalization accuracy (measured by the classification accuracy of target clients not involved in training), personalization accuracy (measured by the accuracy of clients involved in training on the local test set), and communication and computation efficiency (including parameter transfer volume, training time, and resource consumption). Experimental results show that knowledge distillation may experience performance fluctuations due to computational overhead in the early stages of small datasets; therefore, the gains from knowledge distillation are more significant with larger datasets. Experimental results show that on the enterprise technology project reward classification dataset (containing 4 types of labels, 100,000 samples, and non-independent and identically distributed data from 5 clients), the performance of this method compared with existing technologies is as follows: Compared with CN112233445A (attention adapter federated learning): classification accuracy is improved by 5.2%, and communication overhead is reduced by 90%; compared with FedCLIP (federated learning + CLIP, without distillation): classification accuracy is improved by 3.8%, and convergence speed is accelerated by 20%; compared with FedDistill (federated learning + knowledge distillation, without CLIP): cross-modal classification accuracy is improved by 6.5%, and robustness to non-independent and identically distributed data is improved by 12%. This method achieves efficient knowledge transfer and resource optimization while protecting privacy, and is suitable for distributed applications in privacy-sensitive fields.

[0039] Referring to Figure 6, the experimental results based on the enterprise technology project reward classification dataset (including 4 categories of labels, 100,000 samples, and non-independent identically distributed data from 5 clients) are shown. Comparing the core performance differences between FedCLIP (CLIP + Federated Learning) and this invention (CLIP + Federated Learning + Knowledge Distillation), it is clear that the introduction of knowledge distillation significantly improves the technical performance in non-IID data scenarios (e.g., client 1 mainly uses "company-level reward" samples, and client 2 mainly uses "provincial-level reward" samples). The classification accuracy of enterprise technology project rewards increased from 94.40% to 98.54%, an improvement stemming from the "teacher-student" collaborative mechanism of knowledge distillation, which further absorbs the global semantic knowledge of CLIP, reducing misclassification caused by local data noise or class imbalance. The cross-client generalization accuracy increased from 70.56% to 78.56%. This improvement in generalization performance demonstrates the adaptability of the method in this embodiment in distributed scenarios. The synergy of "local feature selection + global knowledge preservation" allows cross-client feature alignment to be achieved simply by aggregating adapter parameters, significantly improving generalization ability. Robustness to non-IID data improved from 62.19% to 74.19%. In the enterprise technology project reward classification scenario, the data from each client naturally exhibits non-independent and identically distributed characteristics. Due to the lack of global knowledge guidance, FedCLIP's local model is prone to overfitting local data distributions, leading to a sharp drop in performance during cross-client inference. This invention forces the student model to align with CLIP's soft labels through knowledge distillation, using the global semantic knowledge of the teacher model to constrain local training, reducing the impact of distribution shift. The number of rounds required to achieve 90% accuracy at convergence speed was reduced from 150 rounds to 120 rounds. The efficiency improvement stems from the synergy between knowledge distillation and the lightweight adapter. The lightweight attention adapter reduces the computational overhead of local training and parameter transmission, enabling the model to reach the target accuracy in fewer communication rounds, balancing performance and efficiency.

[0040] In addition, this embodiment also proposes a federated collaborative training optimization system for classifying project results rewards, including a server and multiple clients.

[0041] The client includes: a multimodal feature extraction module, used to extract visual features of project result images using the image encoder of a pre-trained CLIP model, and use its text encoder to convert reward classification labels into natural language templates to generate semantic vectors, so that visual features and semantic vectors are aligned in a unified semantic space; an attention adapter module, used to dynamically generate attention weights and filter a subset of task discriminative features related to the project result reward category based on the cross-modal cosine similarity between the visual features and semantic vectors; and a local training module, used to take the feature subset as input, define the CLIP model as the teacher model, define the local classifier deployed on the client as the student model, use knowledge distillation to transfer knowledge from the teacher model to the student model, jointly optimize the classification loss and knowledge distillation loss, and complete the local model training including the attention adapter parameters.

[0042] The server includes: a global model distribution module, used to distribute initial or updated attention adapter parameters to each client; a parameter aggregation module, used to receive attention adapter parameters uploaded by clients participating in this round of training, and perform a weighted average based on the number of samples from each client to generate updated attention adapter parameters; and a model synchronization module, used to distribute the updated attention adapter parameters to the clients to support the next round of federated training iterations.

[0043] The above embodiments are only used to explain the concept of the present invention, and are not intended to limit the protection of the present invention. Any non-substantial modifications made to the present invention using this concept should fall within the protection scope of the present invention.

Claims

1. A federated collaborative training optimization method for classifying project outcome rewards, characterized in that, Includes the following steps: S1. The client participating in the training uses the image encoder of the pre-trained CLIP model to extract the visual features of the project results image, and uses its text encoder to convert the reward classification labels into natural language templates to generate semantic vectors, so that the visual features and semantic vectors are aligned in a unified semantic space. S2. Based on the visual features and semantic vectors extracted in step S1, a subset of task discriminative features related to the project's achievement reward category is filtered through an attention adapter deployed on the client. S3. Using a subset of features as input, the CLIP model is defined as the teacher model, and the local classifier deployed on the client is defined as the student model. Knowledge distillation is used to transfer the knowledge of the teacher model to the student model. The classification loss and knowledge distillation loss are jointly optimized to complete the local model training, including the attention adapter parameters. S4. The server only aggregates the attention adapter parameters from the clients participating in training, processes and generates new attention adapter parameters, and distributes them to all clients.

2. The federated collaborative training optimization method for project outcome reward classification as described in claim 1, characterized in that, In step S2, the selection of feature subsets is based on the cross-modal cosine similarity normalization result. The selection weights are dynamically adjusted by the cosine distance between visual features and text semantic vectors. The formula for calculating the cosine similarity is as follows: In the formula, The visual features output by the CLIP model image encoder are represented by T, and the semantic vector output by the CLIP model text encoder is represented by cosine similarity. As an input bias of the attention adapter.

3. The federated collaborative training optimization method for project outcome reward classification as described in claim 1, characterized in that, In S2, the attention adapter includes a linear input layer, a cross-modal fusion layer, a gated activation function GELU-Gate, a distillation-guided Softmax, and a linear output layer.

4. The federated collaborative training optimization method for project outcome reward classification as described in claim 1, characterized in that, In S3, the temperature parameter t=2 for knowledge distillation, the balance parameter between distillation loss and classification loss is 0.5, and gradient pruning mechanism is used to avoid gradient conflict between distillation loss and task loss, and the gradient magnitude of distillation loss is adjusted through knowledge distillation loss function.

5. The federated collaborative training optimization method for project outcome reward classification as described in claim 1, characterized in that: In S3, the number of training epochs is E=1, and the total number of communication epochs is R=200; the Adam optimizer is used, and the learning rate range is 5× Up to 5× 。 6. The federated collaborative training optimization method for project outcome reward classification as described in claim 1, characterized in that: In S4, the server aggregation strategy includes: selecting 50% of the clients to participate in training in each round, and weighting the uploaded attention adapter parameters according to the number of local samples of each participating client to generate new attention adapter parameters.

7. A federated collaborative training optimization system for classifying project outcome rewards based on the method of any one of claims 1 to 6, characterized in that, include: Multiple clients are configured to execute S1–S3.

8. The server is configured to run S4.

9. The federated collaborative training optimization system for classifying project achievement rewards as described in claim 7, characterized in that, The client includes: a multimodal feature extraction module, used to extract visual features of project result images using the image encoder of a pre-trained CLIP model, and use its text encoder to convert reward classification labels into natural language templates to generate semantic vectors, so that visual features and semantic vectors are aligned in a unified semantic space; an attention adapter module, used to dynamically generate attention weights and filter a subset of task discriminative features related to the project result reward category based on the cross-modal cosine similarity between the visual features and semantic vectors; and a local training module, used to take the feature subset as input, define the CLIP model as the teacher model, define the local classifier deployed on the client as the student model, use knowledge distillation to transfer knowledge from the teacher model to the student model, jointly optimize the classification loss and knowledge distillation loss, and complete the local model training including the attention adapter parameters.

10. The federated collaborative training optimization system for classifying project achievement rewards as described in claim 7, characterized in that, The server includes: a global model distribution module, used to distribute initial or updated attention adapter parameters to each client; a parameter aggregation module, used to receive attention adapter parameters uploaded by clients participating in this round of training, and perform a weighted average based on the number of samples from each client to generate updated attention adapter parameters; and a model synchronization module, used to distribute the updated attention adapter parameters to the clients to support the next round of federated training iterations.

Images