摘要
近年来随着基于神经网络深度学习的二维人体姿态估计技术迅速发展,同时也带来了隐私风险。在服务器端进行推理时,未经授权的第三方可能会窃取并滥用用户的性别、面部特征等敏感信息。针对这一挑战,提出了一种基于人类视觉认知与神经网络模型差异的保护策略。在对输入图像进行离散小波变换后,训练与推断阶段选取高频分量,从而有效隐藏可识别的视觉细节。实验表明,方法在保证高精度和高召回率的同时,显著增强了对视觉信息的保护,实现了更优的性能和隐私平衡。
In recent years,the rapid development of two-dimensional human pose estimation technology based on neural network deep learning has also brought about privacy risks.When inference is performed on the server side,unauthorized third parties may steal and misuse sensitive information such as users'gender and facial features.To address this challenge,a protection strategy based on the differences between human visual cognition and neural network models is proposed.After performing discrete wavelet transform on the input image,high-frequency components are selected during both training and inference stages to effectively hide recognizable visual details.Experiments show that this method significantly enhances the protection of visual information while ensuring high accuracy and recall rate,achieving better performance and privacy balance.
作者
吴晶
WU Jing(Northeast Petroleum University,Daqing Heilongjiang 163318,China)
出处
《佳木斯大学学报(自然科学版)》
2025年第7期33-36,共4页
Journal of Jiamusi University:Natural Science Edition
关键词
隐私保护
神经网络
计算机视觉
离散小波变换
人体姿态估计
privacy protection
neural network
computer vision
discrete wavelet transform
human pose estimation
