Ground battlefield threat target detection method based on federated learning

Abstract

The invention relates to the field of federated learning and ground target intelligent perception, in particular to a ground battlefield threat target detection method based on federated learning. According to the technical scheme, the ground battlefield threat target detection method based on federated learning comprises the following steps that S1, clients select adaptive YOLOv8 detection models according to the computing power of the clients, a YOLOv8n model is selected for low-computing-power equipment, a YOLOv8s model is selected for medium-computing-power equipment, and a YOLOv8m model is selected for high-computing-power equipment; s2, the server sets a time window Tmax, only collects client data which completes training and uploads parameters in the time window Tmax, and screens effective clients through an indicator function; and S3, the server integrates the client knowledge through multi-level knowledge distillation and fuses the multi-client knowledge to generate an enhanced global model. Effective coping with equipment heterogeneity is achieved, the overall participation degree and performance are improved, and overall optimization in a heterogeneous environment is achieved by allowing the client to dynamically select YOLOv8 models of different scales according to the computing power of the client.

Description

Technical Field

[0001] This invention relates to the fields of federated learning and intelligent perception of ground targets, and in particular to a method for detecting threat targets on the ground battlefield based on federated learning. Background Technology

[0002] In military battlefield environments, various edge devices (such as drones, reconnaissance vehicles, and individual soldier terminals) need to detect threat targets (such as enemy vehicles and weapons) in real time. However, the computing power of these battlefield edge devices varies significantly: high-performance devices can run complex detection models (such as YOLOv8 and Faster R-CNN), while low-power devices, limited by computing power, can only deploy lightweight models. Traditional federated learning methods require all clients to use the same model structure, which prevents resource-constrained devices from effectively participating in training or forces them to sacrifice detection accuracy. Furthermore, battlefield threat targets are highly dynamic, requiring models to have the ability to respond quickly and update in real time, a requirement that the fixed training cycle and static aggregation mechanism of traditional federated learning cannot meet. Existing technical solutions typically separate target detection from tasks such as distance estimation, resulting in computational redundancy, response delays, and the use of active detection devices easily exposes targets. Although multi-task learning attempts to integrate multi-attribute perception, its complex structure and high computational redundancy make it difficult to adapt to resource-constrained battlefield devices. Finally, traditional centralized training requires sharing sensitive battlefield data, which contradicts the needs of distributed battlefield deployment and poses a significant risk of privacy breaches.

[0003] Traditional methods have significant limitations when facing the heterogeneous computing capabilities of edge devices and the highly dynamic nature of threat targets in battlefield environments: First, federated learning frameworks requiring a unified model structure make it difficult for resource-constrained devices to effectively participate in collaborative training, resulting in a tradeoff between model detection accuracy and real-time performance; Second, static aggregation mechanisms and fixed training cycles are insufficient to meet the battlefield environment's demand for rapid model adaptation, causing detection lag; Third, existing technologies often separate target detection from ranging tasks, leading to system redundancy, response delays, and the vulnerability of targets to active detection; Fourth, the data sharing mechanisms required for centralized training fundamentally contradict the requirements of distributed deployment and data privacy protection on the battlefield. Summary of the Invention

[0004] This invention proposes a ground battlefield threat target detection method based on federated learning, which solves the problem that the traditional federated learning framework in the prior art is unable to effectively coordinate heterogeneous edge devices on the battlefield to achieve high-precision, low-latency threat target detection and rapid adaptation to dynamic environments while protecting data privacy.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A ground battlefield threat target detection method based on federated learning includes the following steps: S1. Each client selects an appropriate YOLOv8 detection model based on its own computing power: low-power devices use the YOLOv8n model, medium-power devices use the YOLOv8s model, and high-power devices use the YOLOv8m model. The client loads a local battlefield threat dataset containing infrared and radar multimodal data, with the data distribution following a non-independent and identically distributed pattern. The client executes the standard YOLOv8 training process, including Mosaic data augmentation, and uses the original YOLO loss function, which includes classification loss, localization loss, and confidence loss, for training. After each round of training, the client generates local model parameters wi and records the training time Ti. S2. The server sets a time window Tmax and only collects data from clients that complete training and upload parameters within the time window Tmax, via an indicator function. Valid clients are selected; the server selects the client model with the largest parameter size from the uploaded models as the teacher model according to the priority of YOLOv8m>YOLOv8s>YOLOv8n, and the rest are used as student models. S3. The server integrates client knowledge through multi-level knowledge distillation, and generates an enhanced global model by fusing knowledge from multiple clients. The knowledge distillation includes: extracting feature maps from layers 5 / 7 / 9 of the teacher network backbone, aligning features using KL divergence; aligning the classification logic output using binary cross-entropy loss; aligning bounding box prediction results based on IoU loss; and then calculating the total loss function to update the global model. S4. The server generates a customized model based on the client device specifications. For high-performance devices, the complete distilled detection model is retained, while for low-performance devices, channel pruning technology is used to lightweight the backbone network of the YOLOv8n model. The server distributes the adapted model parameters to each client through an encrypted channel. After receiving the model, the client performs local inference verification. S5. The client calculates the local loss function value and key indicators and feeds them back to the server. The server aggregates the global loss function of all clients. Repeat steps S1 to S4 until the model converges or reaches the preset number of training rounds.

[0006] Furthermore, step S1 specifically includes: S11. The central server creates a basic target detection model based on the improved YOLOv8 architecture, which includes a backbone network that uses the C2f module to enhance multi-scale feature extraction capabilities, a feature fusion network that optimizes small target detection by combining FPN and PAN, and a decoupled detection head that separates the classification and localization branches and introduces DFL loss. S12. The client selects the YOLOv8 model size based on its own computing power; S13. The client loads a local private dataset. The data distribution follows a non-independent and identically distributed pattern, containing the main category data and a small amount of other category data. S14. Process the local dataset by first removing invalid data, then performing standardization and data augmentation to simulate complex battlefield scenarios; S15. Each client executes the standard YOLOv8 training process, trains the model locally using private data, uses the original YOLO loss function, generates updated local model parameters wi in each training round, and records the training time Ti for each round.

[0007] Furthermore, step S2 specifically includes: S21. The central server sets a global time window Tmax and only collects client data that has completed training and uploaded parameters within the time window Tmax; S22. Calculate the set of valid participating clients. ,in This represents the sum of the number of data samples from all clients that completed their updates within the time window, where ni represents the amount of local data on the client. S23. The server selects teacher models from the uploaded models according to the model size priority.

[0008] Furthermore, step S3 specifically includes: S31. Normalize the channel activation graph from the teacher network into a soft probability graph corresponding to the key features extracted from layers 5, 7, and 9 of the backbone layer; S32. The server performs feature-level knowledge distillation to extract feature maps of the teacher model and student model in the backbone network layer, and aligns the student model features using KL divergence. S33. Use binary cross-entropy distillation loss to transfer knowledge from the classification logic of the teacher model to the classification logic of the student model; S34. Using loss distillation based on crossover ratio to transfer the bounding box prediction expertise of the teacher model to the student model; S35. A total loss function that integrates the original YOLOv8 loss and the distillation loss is used to achieve global model updates.

[0009] Furthermore, the KL divergence alignment formula for feature-level knowledge distillation described in step S32 is: Where the temperature coefficient is 0.8, Φ(·) is the Softmax normalization function, C is the number of channels, and W·H is the spatial dimension. , These are the activation values ​​at position i in channel c of the teacher and student models, respectively.

[0010] Furthermore, the binary cross-entropy distillation loss function described in step S33 is: ,in, and These are the binary classification scores of the student and teacher models obtained using the Sigmoid function, respectively. wij is the distillation weight calculated based on the importance of the samples, aiming to focus on extracting important samples.

[0011] Furthermore, the loss distillation function based on the cross-union ratio described in step S34 is: ,in It is the IoU value between the i-th anchor bounding box predicted by the teacher and the student.

[0012] Furthermore, the total loss function described in step S35 is: , where LYOLO is the original loss function of YOLOv8, λ is the balance coefficient, λ=0.5, α1=0.3, α2=0.2, which adjust the tradeoff between YOLOv8 loss and distillation loss, where α1 and α2 are hyperparameters that balance the contributions of classification and localization distillation losses, respectively.

[0013] Furthermore, step S4 specifically includes: S41. The server generates a customized model based on the client device specifications. High-performance devices retain the complete distilled model, while low-performance devices use channel pruning to generate a lightweight version. S42. The server distributes the adapted model parameters through an encrypted channel; after receiving the new model, the client performs local inference testing and verification, calculates mAP@0.5 and latency using the input validation set, and triggers a retraining mechanism if the accuracy drops by more than 5% or the latency exceeds the limit.

[0014] Furthermore, the specific method for the server to aggregate the global loss function in step S5 is as follows: , where K is the total number of clients participating in the federated learning process, ni is the number of data samples of the i-th client, and N represents the total number of samples.

[0015] The positive effects of this invention are: To effectively address device heterogeneity and improve overall participation and performance, this approach allows clients to dynamically select YOLOv8 models of different scales (such as YOLOv8n, YOLOv8s, and YOLOv8m) based on their computing power. By utilizing dynamic client filtering based on time windows and a knowledge distillation strategy that leverages strong models to support weak ones, devices with varying capabilities can effectively participate in federated learning and ultimately obtain high-precision customized models that match their hardware, thus achieving overall performance optimization in heterogeneous environments.

[0016] To improve model update speed and environmental adaptability, a global time window (Tmax) is set on the server to filter fast-responding clients, and efficient multi-level knowledge distillation is combined to perform rapid model aggregation, which significantly shortens the cycle of federated learning rounds, enabling the model to adapt to battlefield changes more quickly and improving the real-time performance and agility of the detection system.

[0017] To achieve efficient and covert multi-task perception, this method enhances the model's multi-task perception capabilities (such as detection and localization) through an improved YOLOv8 architecture (e.g., introducing DFL loss to improve localization accuracy) and a distillation process that integrates knowledge from multiple clients. Inference is based on passive sensor data (infrared, radar), avoiding the exposure risks of active ranging. Furthermore, model customization (e.g., lightweight model pruning) ensures efficient operation on edge devices, achieving covert, accurate, and efficient comprehensive threat target perception.

[0018] While ensuring data privacy, knowledge fusion is achieved. Based on the federated learning framework, data from each client does not need to be uploaded to the server, fundamentally protecting sensitive local battlefield data. Through knowledge distillation technology, knowledge distributed across different clients is extracted and integrated into the global model, achieving "the model moves while the data remains still." This not only strictly protects data privacy but also breaks down data silos, significantly improving the detection accuracy and generalization ability of the global model. Detailed Implementation

[0019] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. 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 of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0020] Example A ground battlefield threat target detection method based on federated learning includes the following steps: S1. Each client selects an appropriate YOLOv8 detection model based on its computing power: low-power devices use the lightweight YOLOv8n model, medium-power devices use the YOLOv8s model, and high-power devices use the full YOLOv8m model. The client loads a local battlefield threat dataset (containing multimodal data such as infrared and radar). The data distribution follows a non-independent identically distributed (Non-IID) pattern, meaning each client's data is primarily composed of the main class data, with a small number of samples from other classes. The client performs local training using the standard YOLOv8 training process, including input image preprocessing, Mosaic data augmentation (randomly cropping, scaling, and stitching multiple images to simulate complex scenes), and training using the original YOLO loss function (including classification loss, localization loss, and confidence loss). After each training round, the client generates local model parameters wi and records the training time Ti. S2. The server sets a time window Tmax (default 100 seconds) and only collects data from clients that complete training and upload parameters within the time window Tmax (via an indicator function). (Filter valid clients), calculate the total amount of data from valid participating clients. The server selects the client model with the largest parameter size from the uploaded models as the teacher model, and the rest as student models, according to the priority of YOLOv8m > YOLOv8s > YOLOv8n. S3. The server integrates client knowledge through multi-level knowledge distillation, and generates an enhanced global model by fusing knowledge from multiple clients. It performs knowledge distillation and integration through feature alignment (extracting feature maps from layers 5 / 7 / 9 of the teacher network backbone and aligning student model features using KL divergence), classification alignment (aligning classification logic outputs using binary cross-entropy loss and weighting distillation weights wij for important samples), and localization alignment (aligning bounding box prediction results based on IoU loss and calculating the IoU similarity between teacher and student prediction boxes). Then, it calculates the total loss function to update the global model and generates an enhanced global model by fusing knowledge from multiple clients. S4. The server generates a customized model based on the client device specifications. For high-performance devices, the complete distilled detection model is retained; for low-performance devices, channel pruning technology is used to lightweight the detection model. The server distributes the adapted model parameters to each client through an encrypted channel. After receiving the model, the client performs local inference verification to ensure that the detection accuracy and real-time performance meet the standards.

[0021] S5. The client calculates the local loss function value and feeds it back to the server. The server aggregates the global loss function of all clients. Repeat steps S1 to S4 until the model converges (e.g., the mAP@0.5 metric stabilizes) or the preset number of training rounds is reached.

[0022] In step S1, local model training and client initialization are performed, and the specific steps are as follows: S11. The central server creates a basic target detection model based on the improved YOLOv8 architecture. The backbone network uses the C2f module to enhance multi-scale feature extraction capabilities to adapt to complex battlefield environments. The feature fusion network optimizes the detection of small targets by combining FPN (Feature Pyramid) and PAN (Path Aggregation Network). The decoupled detection head separates the classification and localization branches and introduces DFL (Distribution Focal Loss) loss to improve bounding box accuracy.

[0023] S12. The central server sets the federated learning rules and dynamic client selection strategy. The client selects the YOLOv8 model size according to its own computing power (number of CPU cores, memory capacity, GPU computing power). Low computing power devices select the YOLOv8n model; medium computing power devices select the YOLOv8s model; and high computing power devices select the YOLOv8m model.

[0024] S13. The client loads a local private dataset. The data distribution follows a non-independent identically distributed (Non-IID) pattern and includes the main category data as well as a small amount of other category data.

[0025] S14. Process the local dataset by first removing invalid data (blurred images, radar signals with sensor noise interference), then performing standardization and data augmentation to simulate complex battlefield scenarios.

[0026] S15. Each client executes the standard YOLOv8 training process (including Mosaic data augmentation), trains the model locally using private data, uses the original YOLO loss function (including classification loss, localization loss and confidence loss), generates updated local model parameters wi for each training round, and records the training time Ti for each round.

[0027] In step S2, the model parameters are uploaded to the server. The server selects the model with the largest parameter size from the uploaded models as the teacher model. The specific steps are as follows: S21. The central server sets a global time window Tmax (default is 100 seconds), and only collects data from clients that complete training and upload parameters within the time window Tmax (via an indicator function). (Filter valid clients), where Tmax represents the global time window set by the server.

[0028] S22. Calculate the set of valid participating clients. ,in This represents the sum of the number of data samples from all clients that completed their updates within the time window, where ni represents the amount of local data on the client.

[0029] S23. The server selects the client model with the largest parameter size from the uploaded models as the teacher model according to the priority of YOLOv8m>YOLOv8s>YOLOv8n, and the rest as student models.

[0030] In step S3, the server integrates client knowledge through multi-level knowledge distillation. The specific steps are as follows: S31. First, the channel activation graph from the teacher network is normalized into a soft probability graph corresponding to the key features extracted from specific backbone layers (i.e., layers 5, 7, and 9). These layers were chosen because of their fundamental role in feature fusion, effectively encapsulating the essential features of the original image data.

[0031] S32. The server performs feature-level knowledge distillation, extracting feature maps of the teacher and student models at the backbone network layers (layers 5 / 7 / 9). Kullback-Leibler (KL) divergence calculation is related to the student network, emphasizing key prediction regions of the model. This step aims to refine the learning of channel-level features in the student model. Student model features are aligned using KL divergence (temperature coefficient is 0.8), as shown in the formula: Where Φ(·) is the Softmax normalization function, C is the number of channels, and W·H is the spatial dimension. , These are the activation values ​​at position i in channel c of the teacher and student models, respectively.

[0032] S33. Binary Cross-Entropy Distillation Loss (BCE Loss) is used to transfer knowledge from the classification logic of the teacher model to the classification logic of the student model (temperature coefficient of 1), aligning the teacher model's classification logic (Sigmoid output) with the student model. The binary classification distillation loss is designed to address cross-testing protocol inconsistencies by treating the classification logit graph as multiple binary classification graphs. The formula for calculating the binary cross-entropy distillation loss function is as follows: ,in, and These are the binary classification scores of the student and teacher models obtained using the Sigmoid function, respectively. wij is the distillation weight calculated based on the importance of the samples, aiming to focus on extracting important samples.

[0033] S34. Loss distillation based on Intersection over Union (IoU) is used to transfer the bounding box prediction expertise of the teacher model to the student model. The IoU similarity between the teacher and student predicted boxes is calculated. The loss function based on IoU is calculated as follows: ,in It is the IoU value between the i-th anchor bounding box predicted by the teacher and the student. The goal is to minimize the loss based on intersection-union (IoU) to improve the student model's ability to accurately delineate object boundaries.

[0034] S35. The total loss function, which integrates the original YOLOv8 loss with the distillation loss, is given by the following equation: Here, LYOLO is the original loss function of YOLOv8, λ is the balancing coefficient (default coefficients: λ=0.5, α1=0.3, α2=0.2), which modulates the trade-off between the YOLOv8 loss and the distillation loss, where α1 and α2 are hyperparameters that balance the contributions of classification and localization distillation losses, respectively. This method enables YOLOv8 to inherit refined object detection insights from the teacher network, particularly improving its performance in complex detection scenarios and small object recognition.

[0035] In step S4, the server generates a customized student model based on the client's hardware capabilities and deploys and updates the model. The specific steps are as follows: S41. The server generates a customized model based on the client device specifications. High-performance devices retain the complete distilled model, while low-performance devices use channel pruning to generate a lightweight version. Channel pruning is performed on the YOLOv8n backbone network to remove redundant channels.

[0036] S42. The server distributes the adapted model parameters through an encrypted channel to prevent man-in-the-middle attacks. After receiving the new model, the client performs local inference testing and verification, inputting a validation set (such as 100 battlefield images), calculating mAP@0.5 and latency. If the accuracy drops by more than 5% or the latency exceeds the limit, a retraining mechanism is triggered.

[0037] In step S5, the client uploads the local loss value. And key indicators (false positive rate, false negative rate), calculate local loss values ​​(such as detection error rate, positioning deviation), and use the global loss function. To measure the model's overall performance across all client data, a smaller value indicates a more accurate model. The loss for each client is weighted by the amount of data collected. Here, K is assumed to be the total number of clients participating in the federated learning process, ni indicates the number of data samples held by the i-th client, and N represents the total number of samples. The loss value is then fed back to the server, where the central server aggregates the loss values ​​from all clients, adjusts the model parameters, and repeats steps 1 to 5 until the model converges or reaches the maximum number of training rounds.

[0038] The above-described embodiments are detailed and specific, illustrating preferred embodiments of the present invention. They are only used to illustrate the technical ideas and features of the present invention, with the aim of enabling those skilled in the art to understand the content of the present invention and implement it accordingly. However, they are not limited to the present invention, and the patent scope of the present invention cannot be limited by this embodiment alone. That is, any equivalent changes or modifications made to the spirit disclosed in the present invention, without departing from the structure of the present invention, such as local improvements within the system and modifications or transformations between subsystems, are still within the patent scope of the present invention.

Claims

1. A ground battlefield threat target detection method based on federated learning, characterized in that, The method comprises the following steps: S1. Each client selects a suitable YOLOv8 detection model according to its own computing power. Low-power devices select YOLOv8n model, medium-power devices select YOLOv8s model, and high-power devices select YOLOv8m model. The client loads the local battlefield threat dataset containing infrared and radar multi-modal data, and the data distribution follows non-independent and identically distributed. The client executes the standard YOLOv8 training process, including Mosaic data enhancement, and uses the original YOLO loss function containing classification loss, positioning loss and confidence loss for training. After each round of training, the client generates local model parameters w i , and records the training time T i ; S2. The server sets a time window T max , only collects the client data that completes training and uploads parameters within the time window T max , and screens effective clients through the indication function ; the server selects the client model with the largest parameter size as the teacher model and the rest as the student model from the uploaded models in the order of YOLOv8m> YOLOv8s> YOLOv8n. S3. The server integrates the client knowledge through multi-level knowledge distillation, fuses multiple client knowledge to generate an enhanced global model; the knowledge distillation comprises: extracting feature maps of the 5th / 7th / 9th layers of the teacher network backbone layer, and aligning the features through KL divergence; aligning the classification logic output using binary cross-entropy loss; aligning the boundary box prediction result based on IoU loss; and then calculating a total loss function to realize global model updating; S4. The server generates a customized model according to the specifications of the client devices, retains the complete distilled detection model for high-performance devices, and adopts channel pruning technology to perform lightweight processing on the backbone network of the YOLOv8n model for low-performance devices; the server distributes the adapted model parameters to each client through an encrypted channel, and the client performs local inference verification after receiving the model; S5. The client calculates the local loss function value and key indicators and feeds back to the server, and the server aggregates the global loss functions of all clients; steps S1 to S4 are repeated until the model converges or the preset training number of rounds is reached.

2. The ground battlefield threat target detection method based on federated learning according to claim 1, characterized in that, Step S1 specifically comprises: S11. The central server creates a basic target detection model based on the improved YOLOv8 architecture, which includes a backbone network using C2f modules to enhance multi-scale feature extraction capability, a feature fusion network optimizing small target detection by combining FPN and PAN, and a decoupled detection head separated by classification and positioning branches and introducing DFL loss; S12. The client selects the YOLOv8 model size according to its own computing power; S13. The client loads the local private data set, and the data distribution follows the non-independent and identically distributed principle, including main category data and a small amount of other category data; S14. The local data set is processed, first invalid data is removed, and then standardized processing and data enhancement are performed to simulate complex battlefield scenarios; S15. Each client executes the standard YOLOv8 training process, trains the model locally with private data, uses the original YOLO loss function, and generates updated local model parameters w each round of training i , and records the training time T each round i .

3. The method of ground battlefield threat target detection based on federated learning according to claim 1, characterized in that, Step S2 specifically comprises: S21. The central server sets a global time window T max , only collects client data that completed training and uploaded parameters within the time window T max ; S22. Calculate the set of effective participating clients where represents the sum of the number of data samples from all clients that completed their updates within the time window, n i represents the amount of data local to the client; S23. The server selects a teacher model from the uploaded models according to the model size priority.

4. The ground battlefield threat target detection method based on federated learning according to claim 1, characterized in that, Step S3 specifically comprises: S31. Normalize the channel activation map from the teacher network to a soft probability map corresponding to the key features extracted from the 5th, 7th and 9th layers of the backbone layer; S32. The server performs feature-level knowledge distillation by extracting feature maps of the teacher model and the student model at the backbone network layer, and aligning the student model features through KL divergence; S33. Adopt binary cross-entropy distillation loss to transfer knowledge from the classification logic of the teacher model to the classification logic of the student model; S34. Adopt IoU-based loss distillation to transfer the boundary box prediction professional knowledge of the teacher model to the student model; S35. Implement global model updating by integrating the total loss function of the original YOLOv8 loss and the distillation loss.

5. The method of ground battlefield threat target detection based on federated learning according to claim 4, characterized in that, The KL divergence alignment formula of the feature-level knowledge distillation in step S32 is: where the temperature coefficient is 0.8, Φ(·) is the Softmax normalization function, C is the number of channels, W·H is the spatial dimension, , are the activation values of the c-th channel and i-th position of the teacher and student model, respectively.

6. The method of ground battlefield threat target detection based on federated learning according to claim 4, characterized in that, The binary cross-entropy distillation loss function in step S33 is: where, and are binary classification scores of student and teacher models obtained using Sigmoid function, respectively, w ij is a distillation weight based on the importance calculation of samples, aiming to focus on extracting important samples.

7. The method of ground battlefield threat target detection based on federated learning according to claim 4, characterized in that, The IoU-based loss distillation function in step S34 is: where is the IoU value between the ith anchor bounding box and the teacher and student predictions.

8. The method of ground battlefield threat target detection based on federated learning according to claim 4, characterized in that, The total loss function in step S35 is: where L YOLO is the original loss function of YOLOv8, λ is a balancing coefficient, λ = 0.5, a1 = 0.3, a2 = 0.2, which adjusts the trade-off between YOLOv8 loss and distillation loss, where a1 and a2 are hyperparameters that balance the contributions of classification and localization distillation losses, respectively.

9. The method of ground battlefield threat target detection based on federated learning according to claim 1, characterized in that, Step S4 specifically comprises: S41. The server generates a customized model according to the specifications of the client devices, retains the complete distilled model for high-performance devices, and adopts channel pruning to generate a lightweight version for low-performance devices; S42. The server distributes the adapted model parameters through an encrypted channel; the client receives the new model and performs a local inference test verification, calculates mAP@0.5 and delay by inputting the verification set, and if the accuracy drops by more than 5% or the delay exceeds the limit, triggers the retraining mechanism.

10. The method of ground battlefield threat target detection based on federated learning according to claim 1, characterized in that, The specific way in which the server aggregates the global loss function in step S5 is as follows: where K is the total number of clients participating in the federated learning process, n; is the number of data samples of the i-th client, and N represents the total number of samples.

Images