首页|车联网中基于有向无环图区块链的个性化联邦互蒸馏学习方法

车联网中基于有向无环图区块链的个性化联邦互蒸馏学习方法

扫码查看
联邦学习(FL)作为一种分布式训练方法,在车联网(IoV)中得到了广泛应用.区别于传统机器学习,FL允许智能网联车辆(CAVs)通过共享模型而非原始数据来协同训练全局模型,从而保护CAV隐私和数据安全.为了提升联邦学习模型精度,降低通信开销,该文首先提出一种基于有向无环图(DAG)区块链和CAVs的IoV架构,分别负责全局模型共享和本地模型训练.其次,设计了一种基于DAG区块链的异步联邦互蒸馏学习(DAFML)算法在本地同时训练教师和学生模型,教师模型的专业级网络结构可取得更高精度,学生模型的轻量级网络结构可降低通信开销,并采用互蒸馏学习使教师模型和学生模型从互相转移的软标签中学习知识以更新模型.此外,为了进一步提高模型精度,基于全局训练轮次和模型精度设定个性化权值来调节互蒸馏占比.仿真结果表明,DAFML算法在模型精度和蒸馏比率方面优于其他比较算法.
Direct Acyclic Graph Blockchain-based Personalized Federated Mutual Distillation Learning in Internet of Vehicles
Federated Learning(FL)emerges as a distributed training method in the Internet of Vehicle(IoV),allowing Connected and Automated Vehicles(CAVs)to train a global model by exchanging models instead of raw data,protecting data privacy.Due to the limitation of model accuracy and communication overhead in FL,in this paper,a Directed Acyclic Graph(DAG)blockchain-based IoV is proposed that comprises a DAG layer and a CAV layer for model sharing and training,respectively.Furthermore,a DAG blockchain-based Asynchronous Federated Mutual distillation Learning(DAFML)algorithm is introduced to improve the model performance,which utilizes a teacher model and a student model to mutual distillation in the local training.Specifically,the teacher model with a professional network could achieve higher model accuracy,while the student model with a lightweight network could reduce the communication overhead in contrast.Moreover,to further improve the model accuracy,the personalized weight based on global epoch and model accuracy is designed to adjust the mutual distillation in the model updating.Simulation results demonstrate that the proposed DAFML algorithm outperforms other benchmarks in terms of the model accuracy and distillation ratio.

Federated Learning(FL)Mutual distillationDirect Acyclic Graph(DAG)Personalized weight

黄晓舸、吴雨航、尹宏博、梁承超、陈前斌

展开 >

重庆邮电大学通信与信息工程学院 重庆 400065

联邦学习 互蒸馏 有向无环图 个性化权值

国家自然科学基金国家自然科学基金重庆市自然科学基金重庆市自然科学基金

6237108262001076CSTB2023NSCQ-MSX0726cstc2020jcyjmsxmX0878

2024

电子与信息学报
中国科学院电子学研究所 国家自然科学基金委员会信息科学部

电子与信息学报

CSTPCD北大核心
影响因子:1.302
ISSN:1009-5896
年,卷(期):2024.46(7)