| 摘 要: 针对复杂施工现场的数据隐私和实时安全检测问题,提出了一种针对施工现场隐私保护方法研究的联邦知识蒸馏学习框架FedSafety,集成 MPI架构与 RSA-OAEP/AES-256-GCM 混合加密机制,结合多种教师知识蒸馏策略,实现隐私保护的 YOLOv7联邦化部署。在包含17840张非独立同分布(Non-IID)施工图像(涵盖昼夜/遮挡场景)的5个区域数据集上验证表明:经30轮联邦训练后,模型准确值 mAP@0.5达到集中式训练的96.7 %,通信开销降低63.4%,边缘端推理速度稳定在49.7frame/s,验证了该框架在高危作业环境中隐私保护与高效协同的平衡能力,为工业边缘智能提供安全可靠的技术路径。 |
| 关键词: 联邦学习 施工现场隐私保护 混合加密 YOLOv7 知识蒸馏 |
|
中图分类号: TP301
文献标识码: A
|
| 基金项目: 国家自然科学基金项目资助(62371256) |
|
| Research on Privacy Protection Methods at Construction Sites Based on Federated Knowledge Distillation |
|
JIANG Zhefan, JIANG Yunyu, WU Yuanhao, WANG Dazhi, XIAO Jian
|
(College of Integrated Circuit Science and Engineering, Nanjing University of Posts and Telecommunications, Nanjing 210023, China)
1754054595@qq.com; 1255003271@qq.com; 1915583851@qq.com; 1538566617@qq.com; xiaoj@njupt.edu.cn
|
| Abstract: To address the challenges of data privacy and rea-l time safety detection in complex construction sites,this paper proposes a federated knowledge distillation learning framework called FedSafety for the study of privacy protection methods at construction sites. The framework integrates the MPI architecture with a hybrid encryption mechanism combining RSA-OAEP and AES-256-GCM, along with multiple teacher knowledge distillation strategies, to achieve a privacy-preserving federated deployment of YOLOv7. Experimental validation on a dataset comprising 17 840 Non-IID (Non-independent and identically distributed)construction images (covering day/night and occluded scenarios) from five regions shows that after 30 rounds of federated training, the model achieves an mAP @ 0. 5 accuracy of 96.7% compared to centralized training, reduces communication overhead by 63.4% , and maintains a stable inference speed of 49.7 frame/s on edge devices. These results validate the framework’s ability to balance privacy protection and efficient collaboration in high-risk work environments, providing a secure and reliable technical pathway for industrial edge intelligence. |
| Keywords: Federated Learning construction site privacy protection hybrid encryption YOLOv7 knowledge distillation |
