基于生成式扩散模型的人脸图像隐私保护算法

打开文本图片集

中图分类号：TP391.4 文献标识码：A 文章编号：2096-4706（2025）17-0170-08

Abstract：Aiming at the problems that thefacial information in images is proneto privacy leakage，traditional methods leave traces andcanberestored，and Deep Learning-based methods havedificultyin balancing privacyprotectionand image usability，afacial privacyprotectionalgorithmbasedonthegenerativeDifusionModelisproposed.Adynamicfeaturegated mechanism is introduced，and Gated Convolution is used to adaptively mask the key privacyareas.An Eficient Channel AttentionGatedConvolutionblock isproposedtoimprovethefeatureextractionabilityandaresidualstructureisintroducedto acceleratetheconvergenceofthemodel.ExperimentsshowthatcomparedwiththerDeepLearning-basedmethods，thismethod achieves beterresults intermsofboth privacyprotectionand imageusabilityindicators.Therefore，the proposed facial image privacy protection algorithm based on the generative Diffusion Modelis effective and more advantageous.

KeyWords：DiffusionModel; Gated Convolution;EficientChannelAttention Mechanism; facial privacy protection

0 引言

随着信息技术的飞速发展，各种类型数据的规模呈现指数级增长，而其中的隐私问题也随之日益凸显并引起关注。（剩余13351字）