SecYOLO: 一种用于行为检测的隐私保护多方YOLO框架

陈泽鹏; 陈兰香; 朱益钊; 许胜民

引用本文：

陈泽鹏,陈兰香,朱益钊,许胜民.SecYOLO: 一种用于行为检测的隐私保护多方YOLO框架[J].信息安全学报,已采用 [点击复制]
Chen Zepeng,Chen Lanxiang,Zhu Yizhao,Xu Shengmin.SecYOLO: a privacy-preserving YOLO framework tai-lored for behavioral detection[J].Journal of Cyber Security,Accept [点击复制]

本文已被：浏览 47次下载 0次
SecYOLO: 一种用于行为检测的隐私保护多方YOLO框架
陈泽鹏¹, 陈兰香¹, 朱益钊², 许胜民³
0 字体:加大+\|默认\|缩小-
(1.福建师范大学计算机与网络空间安全学院;2.澳门城市大学;3.福建师范大学)

摘要:

实时目标检测框架YOLO(You Only Look Once)凭借其高效的推理速度与优异的检测精度，已在安防监控、自动驾驶等行为识别任务中取得了显著成效。在实际应用中，行为检测模型训练过程涉及来自多个数据提供方并包含个人生物特征及行为细节的海量敏感图像数据。由于现有隐私保护机制不完善，难以有效应对模型训练过程中的隐私泄露风险。随着YOLO在安防监控等隐私敏感领域的快速部署，其配套的隐私保护机制研究却相对滞后，这种研究与应用之间的不协调加剧了对个人敏感信息保护的风险。对此，我们提出了SecYOLO，一种用于行为检测的隐私保护多方YOLO框架。具体而言，我们给出了一个新的设计，突破了YOLO模型多方协同训练的瓶颈，通过设计多数据拥有者与两个不勾结服务器组成的系统模型，允许多数据拥有者在不进行数据共享情况下协同完成模型训练，并采用线性秘密共享技术对原始图像敏感信息进行保护。在安全性设计方面，本方案采用半诚实攻击模型，并给出了严格的安全性证明。最后，本文通过实验对比分析，详细评估了SecYOLO的性能优势。结果表明，本方案在实现隐私安全的前提下，检测精度上与原始 YOLO 框架表现相当；在训练阶段引入了合理的计算开销，而在推理阶段几乎不引入额外开销，保持了与 YOLO 相同的预测效率。

关键词: SecYOLO 隐私保护行为检测秘密共享安全多方计算

DOI：

投稿时间：2025-06-11修订日期：2026-01-01

基金项目:

SecYOLO: a privacy-preserving YOLO framework tai-lored for behavioral detection

Chen Zepeng¹, Chen Lanxiang¹, Zhu Yizhao², Xu Shengmin³

(1.College of Computer and Cyber Security, Fujian Normal University;2.City University of Macau;3.Fujian Normal University)

Abstract:

Real-time object detection framework You Only Look Once (YOLO), known for its high inference speed and excel-lent detection accuracy, has achieved remarkable success in behavior recognition tasks such as surveillance and autonomous driving. In practice, training behavior detection models often involves massive amounts of sensitive image data containing individuals’ biometric features and behavioral details. However, due to the lack of robust privacy-preserving mechanisms, existing solutions are insufficient to mitigate privacy leakage risks during the model training phase. While YOLO is being rapidly deployed in privacy-sensitive domains like security surveil-lance, its corresponding privacy protection research significantly lags behind. This imbalance between deployment and privacy research exacerbates the risks to personal data security. In this work, we propose SecYOLO, a priva-cy-preserving multi-party YOLO framework for behavior detection. Specifically, we design a system model con-sisting of multiple data owners and two non-colluding servers, enabling collaborative training without data sharing. We leverage linear secret sharing to protect the sensitive information in raw images. In terms of security, our design is based on the semi-honest adversarial model, and we provide a related security analysis. Finally, we conduct ex-tensive experiments to evaluate the performance of SecYOLO. Results show that our framework achieves compara-ble detection accuracy to the original YOLO while ensuring strong privacy protection. Although it introduces mod-erate overhead during training, it introduces minimal computation cost in inference as standard YOLO, thereby maintaining the same prediction efficiency.

Key words: SecYOLO privacy-preserving behavioral detection secret sharing secure multi-party computation