【摘要】 在联邦学习中，共享梯度信息可能导致隐私泄露和数据篡改风险。安全聚合协议提出解决方案，但现有方案通信轮次多、计算开销大。研究人员开始研究轻量级隐私保护联邦学习，其中非交互式联邦学习成为研究热点。针对此问题，提出了双服务器下的非交互可验证的联邦学习方案（NVFL）。该方案利用随机哈希编码和线性同态加密技术保护用户模型隐私，实现正确聚合和验证，无需额外通信开销。NVFL方案轻量级、非交互式，适用于资源有限的设备。仿真测试显示，NVFL框架相比其他安全联邦学习框架，训练精度几乎不损失的情况下，节约了34.3%的时间开销。更多还原

【Abstract】 In federated learning,sharing gradient information can lead to risks of privacy leakage and data tampering. Secure aggregation protocols propose solutions,but existing methods involve multiple communication rounds and high computational costs.Researchers have begun exploring lightweight privacy-preserving federated learning,with non-interactive federated learning emerging as a research focus. To address this issue,a non-interactive verifiable federated learning scheme under dual servers(NVFL)is proposed. This scheme utilizes random hash encoding and linear homomorphic encryption techniques to protect user model privacy,achieving correct aggregation and verification without additional communication overhead. The NVFL scheme is lightweight,non-interactive,and suitable for devices with limited computational resources. Simulation tests demonstrate that the NVFL framework,compared to other secure federated learning frameworks,saves 34.3% of time overhead while maintaining almost no loss in training accuracy.更多还原