基于I-ConvNeXt的GAN生成人脸图像鉴别方法
GAN-generated Fake Images Recognition Based on Improved ConvNeXt
肖梦思 1吴建斌 1涂雅蒙 1袁林锋2
作者信息
- 1. 华中师范大学物理科学与技术学院,湖北 武汉 430079
- 2. 中船重工武汉船舶通信研究所,湖北 武汉 430079
- 折叠
摘要
为鉴别社交网络中人脸图像的真假,在ConvNeXt基础上提出一种针对生成式对抗网络(Generative Adversarial Networks,GAN)生成人脸图像的鉴别方法.该方法以ConvNeXt网络结构为主体,利用人脸图像的颜色特征和空间纹理特征,采用多颜色空间多通道组合输入(Multichannel Input,MCI),扩大网络的学习范围;同时引入通道注意力机制和空间注意力机制来凸显真假人脸图像在颜色分量和纹理特征上的差异,进而实现生成人脸图像和真实人脸图像的检测与识别.实验结果表明,使用改进后的ConvNeXt(Improved ConvNeXt,I-ConvNeXt)网络结构对GAN生成人脸图像的识别准确率达到了99.405%,与原ConvNeXt算法相比,平均准确率提高了1.455个百分点.该结果验证了所提方案的可行性、合理性.
Abstract
In order to distinguish the authenticity of face images in social networks,a recognition method based on ConvNeXt for face image generated by Generative adversarial networks(GAN)is proposed.The ConvNeXt network structure is used as the main body,using the color features and spatial texture features of the face image,and multi-channel combination input(Multi-channel Input,MCI)with multi-color space is used to expand the learning range of the network,while channel attention mecha-nism and spatial attention mechanism are introduced to highlight the differences between real and fake face images in color com-ponents and spatial features,and then the detection and recognition of fake face images are achieved.The experimental results show that the recognition accuracy of face images generated by GAN with improved ConvNeXt(I-ConvNeXt)network structure reaches 99.405%,with an average accuracy improvement of 1.455 percentage points compared with the original ConvNeXt algo-rithm.The results validate the feasibility and reasonableness of the proposed scheme.
关键词
生成式对抗网络/注意力机制/颜色特征/生成人脸/多通道输入Key words
generative adversarial network/attention mechanism/color features/generated face image/multi-channel input引用本文复制引用
基金项目
国家自然科学基金资助项目(U1736121)
中央高校基本科研业务费专项资金资助项目(CCNU22JC024)
出版年
2024
国家科技期刊平台
NSTL
万方数据