基于迁移学习的AI合成人脸图像鉴别研究

Research on AI Synthetic Face Image Identification Based on Transfer Learning

目的人工智能(Artificial Intelligence,AI)生成高质量人脸图像的伪造技术愈发成熟,使得人脸图像的真实性检验面临重大考验。利用一种深度学习的方法对真伪人脸图像进行二分类,以实现对伪造图像的识别。方法提出一种基于迁移学习的方法,构建MobileNetV2网络,保留其在ImageNet数据集上的预训练权值,并对采用FaceSwap技术生成的5274张假脸图像和6650张真脸图像进行辨识。结果迁移模型在测试集上预测的准确度能达到0.94,该网络架构对于真假人脸图像的辨别具有一定的稳健性。结论利用迁移学习的方法能够实现对真伪人脸图像的辨识,在一定程度上对AI合成人脸图像的真实性检验具有借鉴意义。

Objective High-quality face image forgery generated by Artificial Intelligence(AI) technology has become more and more mature, which makes the authenticity testing of face images have a great challenge. This article uses the deep learning approach to classify real and fake face images in order to achieve the identification of fake images. Methods In this paper, we propose a method of transfer learning that we construct MobileNetV2 network and retain its pre-training weight on ImageNet data set. We use 5 274 fake face images produced by FaceSwap technology and 6 650 real face images for identification. Results The migration model gets an accuracy of 0.94 in the test set and has a certain stability to identify real and fake face images. Conclusion Using transfer learning approach can achieve the identification of real and fake face images. To some extent, this method can be used for reference in authenticity testing about AI synthetic face images.