大数据和人工智能等新技术的涌现，使数字媒体安全受到了新的威胁与挑战。传统的数字图像溯源与取证技术已不足以应对新兴的媒体篡改与内容生成手段。因此本项目将利用深度学习技术，对数字图像溯源分析与取证的基础理论与关键技术开展以下研究：（1）针对社交媒体环境中，因多重操作引入图像噪声，进而造成的数字图像设备及操作难以溯源的问题，设计权值卷积单元，研究神经网络轻量化策略，构建高效简便的面向多分类任务的深度神经网络；（2）针对基于深度学习的篡改图像取证问题中存在严重的训练样本失衡问题，探讨交叉熵损失函数的优化方向及基于迁移学习的预训练策略，设计融合双域特征的双流神经网络，以快速准确地判断图像内容的真实性并实现篡改区域精细化定位；（3）针对由深度网络生成的虚假图像取证技术的新需求，在生成分布的变分推断理论基础上，将预训练的深度神经网络与变分自编码器级联，研究一种具有准确性与普适性的生成图像取证方法。

With the emergence of new technologies such as big data and artificial intelligence, the security of digital media has been threatened and challenged. Traditional digital image traceability and forensics technology is not enough to deal with the emerging media tampering and content generation. Therefore, this project will use deep learning technology to carry out the following research on the basic theory and key technologies of digital image traceability analysis and forensics: (1) In social media environments, multiple manipulations will introduce noise to the image, leading to the problem that it is difficult to expose the source and the history of the image. Thus, we will design a weight convolution unit, do research on the strategy of efficient neural networks, in order to create a simple but efficient deep neural network for multiclass classification tasks. (2) Focusing on the imbalance problem of training samples in the tampered image forensics, we will discuss how to optimize the cross entropy loss and how to set the pre-train strategy based on transfer learning. Besides, the two stream neural network will be designed to combine features in different domains, in order to determine the authenticity of the image and achieve refined locating of tampered areas. (3) Inspired by the strong demand for the forensics technologies of fake images generated by deep networks, through connecting a pre-trained deep neural network with a variational auto-encoder, a new accurate and universal generated image forensics technology will be developed based on the variational inference of the generated distribution.