多色彩通道特征融合的GAN合成图像检测方法

GAN Synthetic Image Detection Using Fused Features in the Multi-Color Channels

当前,生成对抗网络(Generative Adversarial Networks,GAN)合成的逼真图像难以识别,严重危害国家网络安全及社会稳定.与此同时,多数基于深度神经网络模型设计的检测器需要大规模训练样本,且存在模型可解释度不高、泛化性能差等问题.为了克服上述亟待解决的关键性难题,本文提出一种多色彩通道特征融合的GAN合成图像检测方法.首先,探索分析真实自然图像和GAN合成图像在不同色彩空间相邻像素之间的差异,并设计差异度量算法,完成色彩通道选择.其次,利用图像像素间的高度相关性,在八个方向上通过二阶马尔可夫链对相邻像素之间的差分数组进行建模,提取差分像素邻接矩阵特征.最后,利用上述特征,设计一种简单且高效的集成分类器完成GAN合成图像的检测任务.在基于StyleGAN模型合成的伪造人脸数据集中,所提出方法的检测准确率高达100.00%;在小样本训练约束条件下,正负样本对数仅仅为2时,检测准确率高达99.65%;在单类样本训练约束条件下,正样本数仅仅为50时,检测准确率高达92.84%.在基于更先进的StyleGAN2和PGGAN模型合成的伪造场景数据集中,所提出方法的检测准确率达到99.96%以上.以上大量实验表明,本文所提出的方法明显优于比较的GAN合成图像检测方法.本文方法已经开源:https://github.com/cyxcyx559/ccss.

Currently,it is very difficult to identify the images synthesized by generative adversarial networks(GAN),which severely poses the threat on national cyber security and social stability.Meanwhile,most classifiers based on deep neural networks require large-scale samples for training,where the problems such as low model interpretability and poor generalization performance are less addressed.To overcome the limitations,we propose to design the ensemble classifier using fused features in the multi-color channels.First of all,by studying the discrimination of adjacent pixels in the multi-color channels between natural and GAN synthetic images,the difference metric is designed based on the correla⁃tion of adjacent pixels,in order to select the optimal color channels.Secondly,by utilizing the highly-correlated relation⁃ship among pixels,the difference array between adjacent pixels are modeled through a second-order Markov chain along eight directions,and meanwhile the subtractive pixel adjacency matrix features are successfully extracted.Finally,based on the extracted features,a simple but efficient detector for identifying GAN synthetic images is constructed.In the image da⁃taset synthesized by the StyleGAN model,the results show that the accuracy of the proposed detector can reach 100.00%.It can also identify GAN synthetic images very well when the pair number of positive and negative training samples is 2(99.65%accuracy)or only 50 positive training samples are provided(92.84%accuracy).The accuracy can also reach more than 99.96%in the image dataset synthesized by StyleGAN2 and PGGAN models.Numerous experiments show that the proposed method in this paper is better than the compared forensic methods.Our code is available at https://github.com/cyxcyx559/ccss.