面向图像视频取证的机器学习综述

A Survey of Deep Learning in Image and Video Forensics

近年来,随着机器学习技术,特别是深度学习技术的飞速发展,使得一般人也能够生成非常逼真的高质量造假图像和视频。这给社会和个人带来了极大的风险,也引起了世界各国相关部门以及学术界的高度重视。针对图像和视频的篡改技术和取证技术是相互对抗相互促进的矛盾双方。机器学习技术的飞速发展,同样地也触发了图像/视频取证技术的跨越式演化。本文对近年来,特别是过去三年面向图像/视频取证的机器学习技术的飞速发展现状进行了综述,展示了基于传统人工构造特征以及端到端的图像视频取证机器学习方法,并探讨了不同检测技术的优缺点,重点对Deepfake换脸视频的取证技术以及基于深度学习的取证与反取证的对抗进行了介绍。对现有的科研工作进行了科学的归类。最后对其未来的发展趋势进行了展望,旨在为后续学者的研究进一步推动图像/视频取证的机器学习技术提供指导。

In recent years, with the rapid development of machine learning technologies, especially deep learning technologies, even normal people can produce vivid and high-quality forged images and videos, which introduces great risk to our society and brings great attention of governments and scholars. Image/video forgery technologies and the corresponding forensics technologies are the two aspects in a contradiction. Also with the rapid development of machine learning technologies, the evolution of image/video forensics technologies are ongoing. In this essay, the latest development of image/video forensics oriented machine learning technologies is surveyed. The machine learning methods based on traditional handcrafted features and end to end methods are introduced. We discussed the advantages and disadvantages of different detection technologies, focusing on forensics technologies targeted at Deepfake face-transplant videos and deep learning based confrontation between forensics and counter forensics. The existing scientific research work has been scientifically classified. In the end, this essay further outlines future research directions, aiming to provide guidance for follow-up scholars to further promote the machine learning technology of image/video forensics.