人脸隐私增强计算模型优化方案

An Optimization Method for Face Privacy Enhancement Computation Models

为解决现有图像级人脸隐私保护模型中二元属性翻转问题和图像信息损失过大等问题,本文提出了一种优化方案。通过定义属性概率分数概念并设计新的隐私保护损失函数项,来解决隐私保护模型中的二元属性翻转问题。同时,使用L1距离来约束隐私保护图像与原始图像之间的信息差异,以保证在模型实现有效隐私保护的同时减少信息损失。与现有PrivacyNet方案相比,本方案克服了属性翻转问题和生成图像视觉效果差问题,在多属性隐私保护和实用性方面具有优越性。

To address the issues of binary attribute flipping and excessive image information loss in existing image level facial privacy protection models,this paper proposes an optimization scheme.By defining the concept of attribute probability score and designing a new privacy protection loss function term,the problem of binary attribute flipping in privacy protection models is solved.At the same time,L1 distance is used to constrain the information difference between the privacy preserving image and the original image,to ensure that the model achieves effective privacy protection while reducing information loss.Compared with existing PrivacyNet schemes,this scheme overcomes the problems of attribute flipping and poor visual effects of generated images,and has advantages in multi-attribute privacy protection and practicality.