《计算机应用》唯一官方网站 ›› 2026, Vol. 46 ›› Issue (8): 2555-2566.DOI: 10.11772/j.issn.1001-9081.2025070920
金亚东1,2,3, 范菁1,2,3(
), 郗恩康1,2,3, 董华1,2,3, 俞浩1,2,3, 孙伊航1,2,3
收稿日期:2025-08-11
修回日期:2025-10-22
接受日期:2025-10-29
发布日期:2025-11-05
出版日期:2026-08-10
通讯作者:
范菁
作者简介:金亚东(1998—),男,云南曲靖人,硕士研究生,CCF会员,主要研究方向:联邦学习、空天地一体化网络基金资助:
Yadong JIN1,2,3, Jing FAN1,2,3(
), Enkang XI1,2,3, Hua DONG1,2,3, Hao YU1,2,3, Yihang SUN1,2,3
Received:2025-08-11
Revised:2025-10-22
Accepted:2025-10-29
Online:2025-11-05
Published:2026-08-10
Contact:
Jing FAN
About author:JIN Yadong, born in 1998, M. S. candidate. His research interests include federated learning, air-space-ground integrated network.Supported by:摘要:
空天地一体化网络(SAGIN)作为支撑全球数据传输的综合性网络发挥着关键作用,而它高度动态异构的网络特性导致节点各数据呈现统计异构性,具体表现为非独立同分布(Non-IID)特性。联邦学习(FL)作为分布式机器学习,利用分布式数据训练模型,而FL对SAGIN中的Non-IID数据进行模型训练时,会损害每个节点上全局模型的泛化。针对这个问题,提出个性化联邦学习智能边缘聚合(FedIEA)方法,通过个性化联邦学习(PFL)捕获节点模型的全局模型中的所需信息。FedIEA方法的核心是一个智能边缘聚合(IEA)模块,它可以智能地将下载的全局模型和边缘模型聚合至每个节点上的边缘目标,以在每次迭代中训练之前的初始化本地模型。为了评估FedIEA方法的有效性,在MNIST、FashionMNIST和CIFAR-10这3个数据集上进行对比实验,结果表明:与联邦平均(FedAvg)、FedProx(Federated Proximal)、通过对数几率校准的联邦学习(FedLC)等典型算法相比,FedIEA在迪利克雷(Dirichlet)不同异构设置下的测试准确率都高于FedAvg等基准算法,在FashionMNIST和CIFAR-10数据集上基于卷积神经网络(CNN)模型比FedAvg分别高出25.30和41.28个百分点;在3个数据集上与9个最先进的PFL算法相比,FedIEA基于CNN和深度神经网络(DNN)模型的测试准确率都具有显著的优势。
中图分类号:
金亚东, 范菁, 郗恩康, 董华, 俞浩, 孙伊航. 面向空天地一体化网络的个性化联邦学习智能边缘聚合方法[J]. 计算机应用, 2026, 46(8): 2555-2566.
Yadong JIN, Jing FAN, Enkang XI, Hua DONG, Hao YU, Yihang SUN. Personalized federated learning intelligent edge aggregation method for space-air-ground integrated networks[J]. Journal of Computer Applications, 2026, 46(8): 2555-2566.
| 数据集 | 分类 | 类别数 | 样本数 | 图像的尺寸 |
|---|---|---|---|---|
| MNIST | 图像 | 10 | 7 000 | 28×28 |
| FashionMNIST | 图像 | 10 | 7 000 | 28×28 |
| CIFAR-10 | 图像 | 10 | 6 000 | 32×32 |
表1 3个基准数据集
Tab. 1 Three benchmark datasets
| 数据集 | 分类 | 类别数 | 样本数 | 图像的尺寸 |
|---|---|---|---|---|
| MNIST | 图像 | 10 | 7 000 | 28×28 |
| FashionMNIST | 图像 | 10 | 7 000 | 28×28 |
| CIFAR-10 | 图像 | 10 | 6 000 | 32×32 |
| 算力等级 | 对应节点类型 | 本地训练轮次 | 本地批次大小 |
|---|---|---|---|
| 高算力 | BS、LEO卫星 | 5 | 32 |
| 中算力 | 低空UAV | 3 | 20 |
| 低算力 | IoT | 1 | 10 |
表2 不同算力等级节点下的性能稳定性
Tab. 2 Performance stability under nodes of different computing power levels
| 算力等级 | 对应节点类型 | 本地训练轮次 | 本地批次大小 |
|---|---|---|---|
| 高算力 | BS、LEO卫星 | 5 | 32 |
| 中算力 | 低空UAV | 3 | 20 |
| 低算力 | IoT | 1 | 10 |
| 任务类型 | 数据集名称 | 类别数 | 总样本数 | 数据维度 |
|---|---|---|---|---|
| 遥感图像解译 | UC Merced Land Use | 21 | 2 100 | 256×256×3 |
| IoT状态检测 | IoTID20 | 8 | 11 200 | 41(特征维度) |
| 卫星通信流量异常检测 | Satellite Traffic2024 | 5 | 8 000 | 16(时序维度) |
表3 SAGIN典型任务数据集
Tab. 3 SAGIN typical task datasets
| 任务类型 | 数据集名称 | 类别数 | 总样本数 | 数据维度 |
|---|---|---|---|---|
| 遥感图像解译 | UC Merced Land Use | 21 | 2 100 | 256×256×3 |
| IoT状态检测 | IoTID20 | 8 | 11 200 | 41(特征维度) |
| 卫星通信流量异常检测 | Satellite Traffic2024 | 5 | 8 000 | 16(时序维度) |
| 带宽 | 参数/MHz | 带宽 | 参数/MHz |
|---|---|---|---|
| BS-UAV带宽 | 20.0 | UAV-LEO带宽 | 10.0 |
| BS-LEO带宽 | 37.5 | UAV-GEO带宽 | 5.0 |
| BS-GEO带宽 | 25.0 |
表4 模拟参数
Tab. 4 Simulation parameters
| 带宽 | 参数/MHz | 带宽 | 参数/MHz |
|---|---|---|---|
| BS-UAV带宽 | 20.0 | UAV-LEO带宽 | 10.0 |
| BS-LEO带宽 | 37.5 | UAV-GEO带宽 | 5.0 |
| BS-GEO带宽 | 25.0 |
| 算法 | MNIST | FashionMNIST | CIFAR-10 | |||
|---|---|---|---|---|---|---|
| CNN | DNN | CNN | DNN | CNN | DNN | |
| FedAvg[ | 87.21 | 67.81 | 71.10 | 66.61 | 45.03 | 32.02 |
| FedProx[ | 87.28 | 85.03 | 72.75 | 76.29 | 45.24 | 33.22 |
| FedLC[ | 87.19 | 83.80 | 72.77 | 74.30 | 45.07 | 31.94 |
| FedNTD[ | 81.10 | 98.89 | 70.05 | 75.78 | 41.68 | 37.28 |
| FedPer[ | 99.03 | 99.34 | 95.20 | 96.08 | 86.15 | 75.20 |
| Ditto[ | 99.52 | 99.89 | 96.21 | 96.93 | 86.28 | 75.21 |
| FedFomo[ | 99.54 | 99.89 | 96.18 | 96.93 | 86.07 | 75.03 |
| FedAMP[ | 99.57 | 99.87 | 96.20 | 96.93 | 86.30 | 75.23 |
| APPLE[ | 99.28 | 99.89 | 96.19 | 96.89 | 86.23 | 70.82 |
| FedCP[ | 99.45 | 99.82 | 96.03 | 96.72 | 86.16 | 74.68 |
| GPFL[ | 99.18 | 99.65 | 95.69 | 96.53 | 84.54 | 74.52 |
| PFL-DA[ | 99.62 | 99.82 | 95.45 | 96.94 | 86.29 | 75.20 |
| FedAS[ | 99.55 | 99.82 | 96.19 | 96.75 | 86.27 | 74.79 |
| FedIEA | 99.80 | 99.90 | 96.40 | 96.88 | 86.31 | 75.25 |
表5 不同算法在3个数据集上的测试准确率比较 (%)
Tab. 5 Comparison of test accuracies of different algorithms on three datasets
| 算法 | MNIST | FashionMNIST | CIFAR-10 | |||
|---|---|---|---|---|---|---|
| CNN | DNN | CNN | DNN | CNN | DNN | |
| FedAvg[ | 87.21 | 67.81 | 71.10 | 66.61 | 45.03 | 32.02 |
| FedProx[ | 87.28 | 85.03 | 72.75 | 76.29 | 45.24 | 33.22 |
| FedLC[ | 87.19 | 83.80 | 72.77 | 74.30 | 45.07 | 31.94 |
| FedNTD[ | 81.10 | 98.89 | 70.05 | 75.78 | 41.68 | 37.28 |
| FedPer[ | 99.03 | 99.34 | 95.20 | 96.08 | 86.15 | 75.20 |
| Ditto[ | 99.52 | 99.89 | 96.21 | 96.93 | 86.28 | 75.21 |
| FedFomo[ | 99.54 | 99.89 | 96.18 | 96.93 | 86.07 | 75.03 |
| FedAMP[ | 99.57 | 99.87 | 96.20 | 96.93 | 86.30 | 75.23 |
| APPLE[ | 99.28 | 99.89 | 96.19 | 96.89 | 86.23 | 70.82 |
| FedCP[ | 99.45 | 99.82 | 96.03 | 96.72 | 86.16 | 74.68 |
| GPFL[ | 99.18 | 99.65 | 95.69 | 96.53 | 84.54 | 74.52 |
| PFL-DA[ | 99.62 | 99.82 | 95.45 | 96.94 | 86.29 | 75.20 |
| FedAS[ | 99.55 | 99.82 | 96.19 | 96.75 | 86.27 | 74.79 |
| FedIEA | 99.80 | 99.90 | 96.40 | 96.88 | 86.31 | 75.25 |
| 不同数据集上的测试准确率 | |||
|---|---|---|---|
| MNIST | FashionMNIST | CIFAR-10 | |
| 10 | 86.48 | 71.19 | 45.64 |
| 20 | 86.99 | 72.00 | 46.58 |
| 40 | 86.84 | 71.32 | 47.06 |
| 60 | 87.08 | 72.33 | 47.60 |
| 80 | 87.29 | 74.26 | 48.46 |
| 100 | 87.97 | 75.58 | 51.84 |
表6 超参数s不同取值对FedIEA算法测试准确率的影响 (%)
Tab. 6 Impact of different values of hyperparameter s on test accuracy of FedIEA algorithm
| 不同数据集上的测试准确率 | |||
|---|---|---|---|
| MNIST | FashionMNIST | CIFAR-10 | |
| 10 | 86.48 | 71.19 | 45.64 |
| 20 | 86.99 | 72.00 | 46.58 |
| 40 | 86.84 | 71.32 | 47.06 |
| 60 | 87.08 | 72.33 | 47.60 |
| 80 | 87.29 | 74.26 | 48.46 |
| 100 | 87.97 | 75.58 | 51.84 |
| 不同数据集上的测试准确率/% | |||
|---|---|---|---|
| MNIST | FashionMNIST | CIFAR-10 | |
| 6 | 99.84 | 99.77 | 97.56 |
| 5 | 99.82 | 99.72 | 97.45 |
| 4 | 99.83 | 99.80 | 97.50 |
| 3 | 99.85 | 99.79 | 97.58 |
| 2 | 99.86 | 99.81 | 97.59 |
| 1 | 87.29 | 74.26 | 48.46 |
表7 超参数p不同取值对FedIEA算法测试准确率的影响
Tab. 7 Impact of different values of hyperparameter p on test accuracy of FedIEA algorithm
| 不同数据集上的测试准确率/% | |||
|---|---|---|---|
| MNIST | FashionMNIST | CIFAR-10 | |
| 6 | 99.84 | 99.77 | 97.56 |
| 5 | 99.82 | 99.72 | 97.45 |
| 4 | 99.83 | 99.80 | 97.50 |
| 3 | 99.85 | 99.79 | 97.58 |
| 2 | 99.86 | 99.81 | 97.59 |
| 1 | 87.29 | 74.26 | 48.46 |
| 算法 | MNIST | FashionMNIST | CIFAR-10 | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| β=0.2 | β=0.4 | β=0.6 | β=0.8 | β=0.2 | β=0.4 | β=0.6 | β=0.8 | β=0.2 | β=0.4 | β=0.6 | β=0.8 | |
| FedAvg | 87.21 | 89.95 | 90.34 | 91.82 | 71.10 | 75.67 | 77.82 | 79.32 | 45.03 | 47.10 | 49.08 | 50.97 |
| FedProx | 87.28 | 89.64 | 90.82 | 91.98 | 72.75 | 75.67 | 77.59 | 80.01 | 45.24 | 47.74 | 49.06 | 51.91 |
| FedLC | 87.19 | 90.03 | 90.52 | 91.73 | 72.77 | 75.58 | 76.31 | 80.13 | 45.07 | 47.14 | 50.32 | 52.00 |
| FedNTD | 81.10 | 81.02 | 83.39 | 86.21 | 70.05 | 79.80 | 82.97 | 85.82 | 41.68 | 42.15 | 43.78 | 45.38 |
| FedPer | 99.03 | 96.52 | 93.93 | 90.93 | 95.20 | 89.87 | 87.87 | 85.87 | 86.15 | 75.18 | 69.77 | 65.76 |
| Ditto | 99.52 | 95.73 | 93.94 | 90.94 | 96.21 | 90.92 | 87.92 | 86.93 | 86.28 | 75.30 | 70.28 | 63.32 |
| FedFomo | 99.54 | 96.92 | 91.92 | 90.93 | 96.18 | 90.91 | 88.91 | 85.91 | 86.07 | 76.07 | 70.08 | 66.07 |
| FedAMP | 99.57 | 96.84 | 92.94 | 90.94 | 96.20 | 89.92 | 86.92 | 84.92 | 86.30 | 75.29 | 69.29 | 63.31 |
| APPLE | 99.28 | 95.82 | 92.85 | 91.85 | 96.19 | 90.81 | 87.81 | 85.81 | 86.23 | 73.81 | 68.81 | 60.89 |
| FedCP | 99.45 | 96.90 | 93.92 | 90.92 | 96.03 | 90.87 | 88.86 | 86.86 | 86.16 | 75.75 | 68.74 | 61.79 |
| GPFL | 99.18 | 96.87 | 95.92 | 91.94 | 95.69 | 89.65 | 87.67 | 85.64 | 84.54 | 77.40 | 70.20 | 67.74 |
| PFL-DA | 99.62 | 94.94 | 93.94 | 91.14 | 95.45 | 90.93 | 88.93 | 86.93 | 86.29 | 76.31 | 69.31 | 60.33 |
| FedAS | 99.55 | 95.95 | 92.94 | 90.93 | 96.19 | 90.80 | 89.94 | 86.80 | 86.27 | 75.33 | 68.35 | 61.36 |
| FedIEA | 99.80 | 96.94 | 93.95 | 91.96 | 96.40 | 90.95 | 88.86 | 86.96 | 86.31 | 77.43 | 70.95 | 65.16 |
表8 3个数据集在不同β异构程度下14种算法的测试准确率 ( %)
Tab. 8 Test accuracies of 14 algorithms on three datasets under different heterogeneity levels
| 算法 | MNIST | FashionMNIST | CIFAR-10 | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| β=0.2 | β=0.4 | β=0.6 | β=0.8 | β=0.2 | β=0.4 | β=0.6 | β=0.8 | β=0.2 | β=0.4 | β=0.6 | β=0.8 | |
| FedAvg | 87.21 | 89.95 | 90.34 | 91.82 | 71.10 | 75.67 | 77.82 | 79.32 | 45.03 | 47.10 | 49.08 | 50.97 |
| FedProx | 87.28 | 89.64 | 90.82 | 91.98 | 72.75 | 75.67 | 77.59 | 80.01 | 45.24 | 47.74 | 49.06 | 51.91 |
| FedLC | 87.19 | 90.03 | 90.52 | 91.73 | 72.77 | 75.58 | 76.31 | 80.13 | 45.07 | 47.14 | 50.32 | 52.00 |
| FedNTD | 81.10 | 81.02 | 83.39 | 86.21 | 70.05 | 79.80 | 82.97 | 85.82 | 41.68 | 42.15 | 43.78 | 45.38 |
| FedPer | 99.03 | 96.52 | 93.93 | 90.93 | 95.20 | 89.87 | 87.87 | 85.87 | 86.15 | 75.18 | 69.77 | 65.76 |
| Ditto | 99.52 | 95.73 | 93.94 | 90.94 | 96.21 | 90.92 | 87.92 | 86.93 | 86.28 | 75.30 | 70.28 | 63.32 |
| FedFomo | 99.54 | 96.92 | 91.92 | 90.93 | 96.18 | 90.91 | 88.91 | 85.91 | 86.07 | 76.07 | 70.08 | 66.07 |
| FedAMP | 99.57 | 96.84 | 92.94 | 90.94 | 96.20 | 89.92 | 86.92 | 84.92 | 86.30 | 75.29 | 69.29 | 63.31 |
| APPLE | 99.28 | 95.82 | 92.85 | 91.85 | 96.19 | 90.81 | 87.81 | 85.81 | 86.23 | 73.81 | 68.81 | 60.89 |
| FedCP | 99.45 | 96.90 | 93.92 | 90.92 | 96.03 | 90.87 | 88.86 | 86.86 | 86.16 | 75.75 | 68.74 | 61.79 |
| GPFL | 99.18 | 96.87 | 95.92 | 91.94 | 95.69 | 89.65 | 87.67 | 85.64 | 84.54 | 77.40 | 70.20 | 67.74 |
| PFL-DA | 99.62 | 94.94 | 93.94 | 91.14 | 95.45 | 90.93 | 88.93 | 86.93 | 86.29 | 76.31 | 69.31 | 60.33 |
| FedAS | 99.55 | 95.95 | 92.94 | 90.93 | 96.19 | 90.80 | 89.94 | 86.80 | 86.27 | 75.33 | 68.35 | 61.36 |
| FedIEA | 99.80 | 96.94 | 93.95 | 91.96 | 96.40 | 90.95 | 88.86 | 86.96 | 86.31 | 77.43 | 70.95 | 65.16 |
| 算法 | MNIST | FashionMNIST | CIFAR-10 |
|---|---|---|---|
| FedAvg | 0.976 2 | 0.929 6 | 0.810 6 |
| FedProx | 0.977 8 | 0.925 8 | 0.842 2 |
| FedLC | 0.979 1 | 0.929 7 | 0.843 7 |
| FedNTD | 0.959 0 | 0.919 3 | 0.808 4 |
| FedPer | 0.999 1 | 0.998 1 | 0.989 5 |
| Ditto | 0.999 7 | 0.999 7 | 0.996 3 |
| FedFomo | 0.999 7 | 0.999 0 | 0.993 3 |
| FedAMP | 0.999 7 | 0.999 0 | 0.993 4 |
| APPLE | 0.996 0 | 0.990 3 | 0.989 2 |
| FedCP | 0.995 7 | 0.994 0 | 0.993 2 |
| GPFL | 0.997 3 | 0.997 1 | 0.991 6 |
| PFL-DA | 0.999 0 | 0.998 9 | 0.992 3 |
| FedAS | 0.999 7 | 0.999 1 | 0.993 5 |
| FedIEA | 0.999 8 | 0.998 5 | 0.998 4 |
表9 默认实验设置下14种算法的AUC指标
Tab. 9 AUC metrics of 14 algorithms under default experimental settings
| 算法 | MNIST | FashionMNIST | CIFAR-10 |
|---|---|---|---|
| FedAvg | 0.976 2 | 0.929 6 | 0.810 6 |
| FedProx | 0.977 8 | 0.925 8 | 0.842 2 |
| FedLC | 0.979 1 | 0.929 7 | 0.843 7 |
| FedNTD | 0.959 0 | 0.919 3 | 0.808 4 |
| FedPer | 0.999 1 | 0.998 1 | 0.989 5 |
| Ditto | 0.999 7 | 0.999 7 | 0.996 3 |
| FedFomo | 0.999 7 | 0.999 0 | 0.993 3 |
| FedAMP | 0.999 7 | 0.999 0 | 0.993 4 |
| APPLE | 0.996 0 | 0.990 3 | 0.989 2 |
| FedCP | 0.995 7 | 0.994 0 | 0.993 2 |
| GPFL | 0.997 3 | 0.997 1 | 0.991 6 |
| PFL-DA | 0.999 0 | 0.998 9 | 0.992 3 |
| FedAS | 0.999 7 | 0.999 1 | 0.993 5 |
| FedIEA | 0.999 8 | 0.998 5 | 0.998 4 |
| 算法 | MNIST | FashionMNIST | CIFAR-10 | |||
|---|---|---|---|---|---|---|
| 目标时间/s | 通信流量/MB | 目标时间/s | 通信流量/MB | 目标时间/s | 通信流量/MB | |
| FedAvg | 5 642.88 | 93.31 | 6 479.67 | 93.31 | 5 648.98 | 140.81 |
| FedProx | 7 215.00 | 137.74 | 7 747.27 | 137.74 | 6 415.76 | 207.86 |
| FedLC | 5 663.54 | 93.35 | 5 122.56 | 93.35 | 5 789.52 | 140.85 |
| FedNTD | 8 384.00 | 95.53 | 8 330.14 | 95.53 | 8 098.56 | 144.46 |
| FedPer | 3 535.75 | 93.29 | 2 479.50 | 93.29 | 5 276.50 | 140.79 |
| Ditto | 2 423.07 | 249.32 | 2 626.65 | 182.18 | 3 184.80 | 274.91 |
| FedFomo | 3 796.25 | 226.62 | 5 039.05 | 226.62 | 4 913.46 | 341.97 |
| FedAMP | 2 423.07 | 137.74 | 1 896.49 | 137.74 | 2 478.60 | 207.86 |
| APPLE | 38 176.50 | 182.18 | 41 618.61 | 182.18 | 40 615.90 | 274.91 |
| FedCP | 9 636.12 | 495.91 | 10 587.81 | 495.99 | 12 674.98 | 528.17 |
| GPFL | 9 521.60 | 179.08 | 15 062.70 | 179.08 | 9 604.92 | 226.58 |
| PFL-DA | 1 651.86 | 91.91 | 835.58 | 91.91 | 3 240.12 | 140.27 |
| FedAS | 2 739.60 | 150.30 | 7 402.08 | 191.05 | 11 150.14 | 230.28 |
| FedIEA | 1 624.26 | 91.20 | 7 877.20 | 93.91 | 2 460.52 | 140.21 |
表10 14种算法达到测试准确率所需的目标时间和通信流量
Tab. 10 Target time and communication traffic required for 14 algorithms to achieve test accuracy
| 算法 | MNIST | FashionMNIST | CIFAR-10 | |||
|---|---|---|---|---|---|---|
| 目标时间/s | 通信流量/MB | 目标时间/s | 通信流量/MB | 目标时间/s | 通信流量/MB | |
| FedAvg | 5 642.88 | 93.31 | 6 479.67 | 93.31 | 5 648.98 | 140.81 |
| FedProx | 7 215.00 | 137.74 | 7 747.27 | 137.74 | 6 415.76 | 207.86 |
| FedLC | 5 663.54 | 93.35 | 5 122.56 | 93.35 | 5 789.52 | 140.85 |
| FedNTD | 8 384.00 | 95.53 | 8 330.14 | 95.53 | 8 098.56 | 144.46 |
| FedPer | 3 535.75 | 93.29 | 2 479.50 | 93.29 | 5 276.50 | 140.79 |
| Ditto | 2 423.07 | 249.32 | 2 626.65 | 182.18 | 3 184.80 | 274.91 |
| FedFomo | 3 796.25 | 226.62 | 5 039.05 | 226.62 | 4 913.46 | 341.97 |
| FedAMP | 2 423.07 | 137.74 | 1 896.49 | 137.74 | 2 478.60 | 207.86 |
| APPLE | 38 176.50 | 182.18 | 41 618.61 | 182.18 | 40 615.90 | 274.91 |
| FedCP | 9 636.12 | 495.91 | 10 587.81 | 495.99 | 12 674.98 | 528.17 |
| GPFL | 9 521.60 | 179.08 | 15 062.70 | 179.08 | 9 604.92 | 226.58 |
| PFL-DA | 1 651.86 | 91.91 | 835.58 | 91.91 | 3 240.12 | 140.27 |
| FedAS | 2 739.60 | 150.30 | 7 402.08 | 191.05 | 11 150.14 | 230.28 |
| FedIEA | 1 624.26 | 91.20 | 7 877.20 | 93.91 | 2 460.52 | 140.21 |
| 动态参数组合 | FedAvg[ | FedProx[ | FedAMP[ | FedIEA |
|---|---|---|---|---|
| 38.21 | 39.57 | 78.62 | 82.15 | |
| 41.05 | 42.33 | 80.19 | 84.02 | |
| 45.03 | 45.24 | 86.30 | 86.31 |
表11 动态场景参数下各算法测试准确率 ( %)
Tab. 11 Test accuracies of various algorithms under dynamic scene parameters
| 动态参数组合 | FedAvg[ | FedProx[ | FedAMP[ | FedIEA |
|---|---|---|---|---|
| 38.21 | 39.57 | 78.62 | 82.15 | |
| 41.05 | 42.33 | 80.19 | 84.02 | |
| 45.03 | 45.24 | 86.30 | 86.31 |
| 算法 | 准确率/% | 准确率方差 | ||
|---|---|---|---|---|
| 高算力节点 | 中算力节点 | 低算力节点 | ||
| FedAvg | 48.21 | 42.05 | 35.12 | 28.67 |
| FedProx | 49.15 | 43.22 | 36.89 | 25.33 |
| FedAMP | 87.02 | 84.19 | 79.56 | 9.81 |
| FedIEA | 86.53 | 85.91 | 84.82 | 3.24 |
表12 不同算力节点组上各算法的测试准确率与方差
Tab. 12 Test accuracy and variance of each algorithm on different groups of computing power nodes
| 算法 | 准确率/% | 准确率方差 | ||
|---|---|---|---|---|
| 高算力节点 | 中算力节点 | 低算力节点 | ||
| FedAvg | 48.21 | 42.05 | 35.12 | 28.67 |
| FedProx | 49.15 | 43.22 | 36.89 | 25.33 |
| FedAMP | 87.02 | 84.19 | 79.56 | 9.81 |
| FedIEA | 86.53 | 85.91 | 84.82 | 3.24 |
| 算法 | 遥感图像解译 | IoT设备检测 | 卫星流量检测 | |||
|---|---|---|---|---|---|---|
| 准确率 | AUC | 准确率 | AUC | 准确率 | AUC | |
| FedAvg | 58.21 | 0.821 | 62.33 | 0.856 | 55.19 | 0.803 |
| FedAMP | 83.15 | 0.973 | 87.05 | 0.981 | 81.22 | 0.968 |
| PFL-DA | 84.02 | 0.978 | 88.19 | 0.984 | 82.07 | 0.972 |
| FedIEA | 85.97 | 0.985 | 89.52 | 0.990 | 83.84 | 0.979 |
表13 各算法在SAGIN典型任务下的性能 ( %)
Tab. 13 Performance of each algorithm in typical SAGIN tasks
| 算法 | 遥感图像解译 | IoT设备检测 | 卫星流量检测 | |||
|---|---|---|---|---|---|---|
| 准确率 | AUC | 准确率 | AUC | 准确率 | AUC | |
| FedAvg | 58.21 | 0.821 | 62.33 | 0.856 | 55.19 | 0.803 |
| FedAMP | 83.15 | 0.973 | 87.05 | 0.981 | 81.22 | 0.968 |
| PFL-DA | 84.02 | 0.978 | 88.19 | 0.984 | 82.07 | 0.972 |
| FedIEA | 85.97 | 0.985 | 89.52 | 0.990 | 83.84 | 0.979 |
| 算法 | 数据集 | 成员推理攻击 | 准确率 | 性能损失 | |
|---|---|---|---|---|---|
| 无隐私 | 有隐私 | ||||
| FedAvg | MNIST | 45.21 | 67.81 | 62.15 | 8.35 |
| CIFAR-10 | 48.33 | 45.03 | 39.87 | 11.46 | |
| FedAMP | MNIST | 28.15 | 99.87 | 97.02 | 2.85 |
| CIFAR-10 | 30.22 | 86.30 | 82.19 | 4.76 | |
| FedIEA | MNIST | 15.07 | 99.90 | 98.53 | 1.37 |
| CIFAR-10 | 17.23 | 86.31 | 84.62 | 1.96 | |
表14 分级隐私保护机制下的隐私-性能对比 ( %)
Tab. 14 Privacy-performance comparison under hierarchical privacy protection mechanism
| 算法 | 数据集 | 成员推理攻击 | 准确率 | 性能损失 | |
|---|---|---|---|---|---|
| 无隐私 | 有隐私 | ||||
| FedAvg | MNIST | 45.21 | 67.81 | 62.15 | 8.35 |
| CIFAR-10 | 48.33 | 45.03 | 39.87 | 11.46 | |
| FedAMP | MNIST | 28.15 | 99.87 | 97.02 | 2.85 |
| CIFAR-10 | 30.22 | 86.30 | 82.19 | 4.76 | |
| FedIEA | MNIST | 15.07 | 99.90 | 98.53 | 1.37 |
| CIFAR-10 | 17.23 | 86.31 | 84.62 | 1.96 | |
| [1] | McMahan H B, Moore E, Ramage D, et al. Communication-efficient learning of deep networks from decentralized data[J]. Proceedings of Machine Learning Research, 2017, 54: 1273-1282. |
| [2] | Zhang J, Hua Y, Wang H, et al. FedaLA: adaptive local aggregation for personalized federated learning[J]. Proceedings of the AAAI Conference on Artificial Intelligence, 2023, 37(9): 11237-11244. |
| [3] | Li Y, Wen G. Research and practice of financial credit risk management based on federated learning[J]. Engineering Letters, 2023, 31(1): No.28. |
| [4] | Alhafiz F S, Basuhail A A. Non-IID medical imaging data on COVID-19 in the federated learning framework: impact and directions[J]. COVID, 2024, 4(12): 1985-2016. |
| [5] | Xu C, Xiao H, Li Z, et al. Diversified service oriented multimedia intelligent transmission in 6G space-air-ground-ocean heterogeneous network[J]. IEEE Wireless Communications, 2025, 32(2): 130-138. |
| [6] | Fallah A, Mokhtari A, Ozdaglar A. Personalized federated learning with theoretical guarantees: a model-agnostic meta-learning approach[C]// NeurIPS 2020. Red Hook: Curran Associates Inc., 2020: 3557-3568. |
| [7] | Chen H Y, Chao W L. On bridging generic and personalized federated learning for image classification[PP/OL]. V2. arXiv (2022-07-11) [2025-03-08].. |
| [8] | Zhang J, Hua Y, Wang H, et al. FedCP: separating feature information for personalized federated learning via conditional policy[C]// KDD 2023. New York: ACM, 2023: 3249-3261. |
| [9] | Yang L, Huang J, Li W, et al. Personalized federated learning on Non-IID data via group-based meta-learning[J]. ACM Transactions on Knowledge Discovery from Data, 2023, 17(4): No.49. |
| [10] | Huang C, Chen G, Xiao P, et al. Fair resource allocation for hierarchical federated edge learning in space-air-ground integrated networks via deep reinforcement learning with hybrid control[J]. IEEE Journal on Selected Areas in Communications, 2024, 42(12): 3618-3631. |
| [11] | Yang K, Wang Y, Gao X, et al. Communications in space-air-ground integrated networks: an overview[J]. Space: Science and Technology, 2025, 5: No.0199. |
| [12] | 俞浩,范菁,孙伊航,等. 边缘异构下的联邦分割学习优化方法[J]. 计算机应用, 2026, 46(1): 33-42. |
| Yu Hao, Fan Jing, Sun Yihang, et al. Federated split learning optimization method under edge heterogeneity[J]. Journal of Computer Applications, 2026, 46(1): 33-42. | |
| [13] | Li T, Sahu A K, Zaheer M, et al. Federated optimization in heterogeneous networks[EB/OL]. (2018-12-14) [2025-04-28].. |
| [14] | Zhang J, Li Z, Li B, et al. Federated learning with label distribution skew via logits calibration[J]. Proceedings of Machine Learning Research, 2022, 162: 26311-26329. |
| [15] | Lee G, Jeong M, Shin Y, et al. Preservation of the global knowledge by not-true distillation in federated learning[C]// NeurIPS 2022. Red Hook: Curran Associates Inc., 2022: 38461-38474. |
| [16] | Arivazhagan M G, Aggawal V, Singh A K, et al. Federated learning with personalization layers[PP/OL]. V1. arXiv (2019-12-02) [2025-04-28].. |
| [17] | 张珂嘉,方志军,周南润,等. 基于模型预分配与自蒸馏的个性化联邦学习方法[J]. 计算机应用, 2026, 46(1): 10-20. |
| Zhang Kejia, Fang Zhijun, Zhou Nanrun, et al. Personalized federated learning method based on model pre-assignment and self-distillation[J]. Journal of Computer Applications, 2026, 46(1): 10-20. | |
| [18] | Li T, Hu S, Beirami A, et al. Ditto: fair and robust federated learning through personalization[J]. Proceedings of Machine Learning Research, 2021, 139: 6357-6368. |
| [19] | Zhang M, Sapra K, Fidler S, et al. Personalized federated learning with first order model optimization[PP/OL]. V4. arXiv (2021-03-26) [2025-05-10].. |
| [20] | Huang Y, Chu L, Zhou Z, et al. Personalized cross-silo federated learning on Non-IID data[J]. Proceedings of the AAAI Conference on Artificial Intelligence, 2021, 35(9): 7865-7873. |
| [21] | Luo J, Wu S. Adapt to adaptation: learning personalization for cross-silo federated learning[C]// IJCAI 2022. California: IJCAI, 2022: 2166-2173. |
| [22] | Zhang J, Hua Y, Wang H, et al. GPFL: simultaneously learning global and personalized feature information for personalized federated learning[C]// ICCV 2023. Piscataway: IEEE, 2023: 5018-5028. |
| [23] | Shi N, Kontar R A. Personalized federated learning via domain adaptation with an application to distributed 3D printing[J]. Technometrics, 2023, 65(3): 328-339. |
| [24] | Yang X, Huang W, Ye M. FedAS: bridging inconsistency in personalized federated learning[C]// CVPR 2024. Piscataway: IEEE, 2024: 11986-11995. |
| [25] | Zhao H, Ji F, Wang Y, et al. Space-air-ground-sea integrated network with federated learning[J]. Remote Sensing, 2024, 16(9): No.1640. |
| [26] | Han D J, Fang W, Hosseinalipour S, et al. Orchestrating federated learning in space-air-ground integrated networks: adaptive data offloading and seamless handover[J]. IEEE Journal on Selected Areas in Communications, 2024, 42(12): 3505-3520. |
| [27] | Qin Y, Yang Y, Tang F, et al. Differentiated federated reinforcement learning based traffic offloading on space-air-ground integrated networks[J]. IEEE Transactions on Mobile Computing, 2024, 23(12): 11000-11013. |
| [28] | Li L, Zhu L, Li W. Privacy-preserving federated learning for space-air-ground integrated networks: a bi-level reinforcement learning and adaptive transfer learning optimization framework[J]. Sensors, 2025, 25(9): No.2828. |
| [29] | Shi W, Li J, Cheng N, et al. Multi-drone 3-D trajectory planning and scheduling in drone-assisted radio access networks[J]. IEEE Transactions on Vehicular Technology, 2019, 68(8): 8145-8158. |
| [30] | Lyu X, Ni W, Tian H, et al. Optimal schedule of mobile edge computing for Internet of Things using partial information[J]. IEEE Journal on Selected Areas in Communications, 2017, 35(11): 2606-2615. |
| [31] | Kanellopoulos S A, Kourogiorgas C I, Panagopoulos A D, et al. Channel model for satellite communication links above 10GHz based on Weibull distribution[J]. IEEE Communications Letters, 2014, 18(4): 568-571. |
| [32] | Zhao M, Chen C, Liu L, et al. Orbital collaborative learning in 6G space-air-ground integrated networks[J]. Neurocomputing, 2022, 497: 94-109. |
| [33] | Zhu Z, Hong J, Zhou J. Data-free knowledge distillation for heterogeneous federated learning[J]. Proceedings of Machine Learning Research, 2021, 139: 12878-12889. |
| [34] | Wu X, Zhang Y, Shi M, et al. An adaptive federated learning scheme with differential privacy preserving[J]. Future Generation Computer Systems, 2022, 127: 362-372. |
| [35] | Li H, Dou X, Tao C, et al. RSI-CB: a large-scale remote sensing image classification benchmark using crowdsourced data[J]. Sensors, 2020, 20(6): No.1594. |
| [36] | Soltani M D, Purwita A A, Zeng Z, et al. Modeling the random orientation of mobile devices: measurement, analysis and LiFi use case[J]. IEEE Transactions on Communications, 2019, 67(3): 2157-2172. |
| [37] | Zhou C, Wu W, He H, et al. Delay-aware IoT task scheduling in space-air-ground integrated network[C]// GLOBECOM 2019. Piscataway: IEEE, 2019: 1-6. |
| [1] | 孙天乐, 曹腾飞. 车联网中基于Stackelberg博弈的联邦学习质量感知激励机制[J]. 《计算机应用》唯一官方网站, 2026, 46(7): 2229-2238. |
| [2] | 钟琪, 张淑芬, 张镇博, 李涛. 基于梯度特征的联邦学习后门防御算法[J]. 《计算机应用》唯一官方网站, 2026, 46(7): 2184-2195. |
| [3] | 任志强, 陈学斌, 屈昌盛. 面向联邦学习中拜占庭攻击的主动防御框架[J]. 《计算机应用》唯一官方网站, 2026, 46(7): 2196-2207. |
| [4] | 董汦楗, 顾瑞春. 融合动态特征对齐与温度感知聚合的联邦学习框架[J]. 《计算机应用》唯一官方网站, 2026, 46(6): 1746-1755. |
| [5] | 俞浩, 范菁, 郗恩康, 金亚东, 董华, 孙伊航. 边缘异构下的高效联邦分割学习框架HEFSL[J]. 《计算机应用》唯一官方网站, 2026, 46(5): 1397-1407. |
| [6] | 姜志, 陈学斌, 罗长银, 甄子业. 联邦学习中改进Kolmogorov-Arnold网络的混合优化框架[J]. 《计算机应用》唯一官方网站, 2026, 46(4): 1023-1033. |
| [7] | 平欢, 夏战国, 刘思诚, 刘奇翰, 李春磊. 基于多层联邦学习的终端数据隐私保护方案[J]. 《计算机应用》唯一官方网站, 2026, 46(3): 830-838. |
| [8] | 王磊, 周文轩, 贾柠晖, 屈志昊. 面向隐私敏感物联网数据的联邦学习双向通信压缩[J]. 《计算机应用》唯一官方网站, 2026, 46(3): 887-898. |
| [9] | 马凯光, 陈学斌, 菅银龙, 王柳, 高远. 基于混合序列模型与联邦类平衡算法的网络入侵检测[J]. 《计算机应用》唯一官方网站, 2026, 46(3): 857-866. |
| [10] | 郗恩康, 范菁, 金亚东, 董华, 俞浩, 孙伊航. 联邦学习在隐私安全领域面临的威胁综述[J]. 《计算机应用》唯一官方网站, 2026, 46(3): 798-808. |
| [11] | 钟琪, 张淑芬, 张镇博, 菅银龙, 景忠瑞. 面向联邦学习的投毒攻击检测与防御机制[J]. 《计算机应用》唯一官方网站, 2026, 46(2): 445-457. |
| [12] | 张珂嘉, 方志军, 周南润, 史志才. 基于模型预分配与自蒸馏的个性化联邦学习方法[J]. 《计算机应用》唯一官方网站, 2026, 46(1): 10-20. |
| [13] | 菅银龙, 陈学斌, 景忠瑞, 钟琪, 张镇博. 联邦学习中基于条件生成对抗网络的数据增强方案[J]. 《计算机应用》唯一官方网站, 2026, 46(1): 21-32. |
| [14] | 俞浩, 范菁, 孙伊航, 金亚东, 郗恩康, 董华. 边缘异构下的联邦分割学习优化方法[J]. 《计算机应用》唯一官方网站, 2026, 46(1): 33-42. |
| [15] | 俞浩, 范菁, 孙伊航, 董华, 郗恩康. 联邦学习统计异质性综述[J]. 《计算机应用》唯一官方网站, 2025, 45(9): 2737-2746. |
| 阅读次数 | ||||||
|
全文 |
|
|||||
|
摘要 |
|
|||||