摘要
针对现有自适应量化联邦学习存在参与者本地模型参数隐私泄露的问题,提出一种适合智能物联网的高效安全的自适应量化联邦学习方案。该方案利用自适应量化技术减少参与者的通信开销,设置两个聚合服务器,将Diffie-Hellman密钥交换协议、秘密共享方案和不经意传输协议相结合,构造一种保护本地模型参数隐私的安全聚合协议,并在合理假设下证明所提方案的安全性。实验结果表明该方案能够获得较高准确率的全局模型,极大减少了参与者的通信开销和隐私保护计算开销,非常适用于智能物联网中资源受限的轻量级物联网设备。
For the problem of participants’private leakage on the model parameters in the existing adaptive quantization fede-rated learning schemes,this paper proposed an efficient and secure adaptive quantization federated learning scheme suitable for artificial intelligence of things(AIoT).This scheme utilized adaptive quantization technology to reduce the communication overhead for participants.It constructed a secure aggregation protocol in two aggregation servers to protect the privacy of local model parameters by combining the Diffie-Hellman key exchange protocol,secret sharing schemes,and oblivious transfer protocols.This paper proved the proposed scheme was secure under reasonable assumptions.The experimental results show that the scheme not only can achieve a global model with high accuracy,but also can significantly reduce communication overhead and computation costs for protecting participants’privacy.This scheme is suitable for resource-constrained,such as lightweight IoT devices in AIoT.
作者
马海英
沈金宇
杨天玲
仇健
王占君
Ma Haiying;Shen Jinyu;Yang Tianling;Qiu Jian;Wang Zhanjun(School of Artificial Intelligence&Computer Science,Nantong University,Nantong Jiangsu 226019,China;School of Mathematics&Statistics,Nantong University,Nantong Jiangsu 226019,China)
出处
《计算机应用研究》
北大核心
2025年第8期2503-2510,共8页
Application Research of Computers
基金
南通市自然科学基金面上项目(JC2023069)
南通大学信息科学技术学院研究生科研与实践创新计划资助项目(NTUSISTPR24_07)。
关键词
联邦学习
隐私保护
自适应量化
秘密共享
不经意传输协议
federated learning(FL)
privacy protection
adaptive quantization
secret sharing
oblivious transfer protocol
