基于卷积痕迹挖掘的GAN生成假脸图片检测

Research on GAN-Generated False Face Image Detection Based on Convolution Trace Mining

虚假人脸的生成与篡改带来的安全威胁已引起广泛关注。针对目前大部分研究需分别对不同GAN(生成对抗网络)生成的虚假图片训练检测模型,很难提取通用特征检测虚假图片,导致模型泛化能力不足的问题,提出了一种基于虚假图片生成过程中转置卷积层造假痕迹挖掘的GAN生成人脸虚假图片检测模型。首先基于虚假图片像素局部相关性和对比损失函数改进图片特征向量提取框架,再利用粒子群算法改进最大期望算法构成EM-PSO(最大期望-粒子群)算法优化特征向量求解过程,进而获取模型在图片RGB三通道计算得到的特征向量。通过支持向量机和图片特征向量实现虚假图片检测。实验结果表明:在由FFHQ真实人脸数据、StyleGAN和StyleGAN2生成的假脸数据构成的模型训练数据集上,模型检测准确率最高可达94.25%,AUC值可达0.99,模型检测准确率明显优于VGG16模型在此数据集上的检测准确率,由此验证了该模型的有效性。

The threat caused by the generation and tampering of fake faces has attracted widespread attention.But most of the current research needs to train detection models respectively for the fake images generated by different GANs(Generative Adversarial Networks)since they are hard to find the common discriminative features for judging the fake images,resulting in insufficient model generalization ability.A GAN-generated fake face images detection model based on forensics trace mining in transposed convolution layer is proposed for the insufficient generalization ability of existing models.Firstly,the image feature vector extraction framework is optimized based on the principle of GAN-generated image pixel local correlation and contrastive loss function.And then the particle optimization swarm algorithm and the maximum expectation algorithm are used to compose the EM-PSO(maximum expectation-particle swarm)algorithm,which will optimize model solution process and obtain the feature vector from RGB three-channel in convolution process.Finally,support vector machine uses feature vectors to detect fake images.Experiments were performed on data set composed of FFHQ real face data and fake face data generated by StyleGAN and StyleGAN2.The results demonstrate that the detection accuracy of the proposed model reaches 94.25%,and the AUC value reaches 0.99.The detection accuracy of proposed model is superior to the VGG16 model,verifying the effectiveness of the proposed model.