面向GANs合成人脸图像的鉴别方法研究

随着互联网及信息技术飞速发展，网络数据呈爆炸式增长，由于技术滥用和监管不力，合成人脸图像充斥网络且效果逼真，对人眼和自动识别系统发起严峻挑战，造成诸多法律、伦理和社会问题。项目以GANs合成人脸图像为研究对象，针对鉴别过程中存在特征属性难以刻画、基准样本不足、对抗环境未知、存在新生成模型攻击等问题，分别在“大样本”、“少样本”和“零样本”情况下依次从特征感知、学习行为、模式理解三个层面展开工作：（1）针对合成人脸图像细粒度特征不易被表征问题，研究基于神经元激活响应及其紧致表示的特征感知与学习方法；（2）针对基准样本不足的“少样本”学习问题，结合网络数据分析与学习特点，从学习行为层面探索基于流形正则化的在线半监督学习策略构建方法；（3）针对未知新生成模型的“零样本”学习问题，从模式理解层面挖掘基于树状神经网络架构的无监督自主学习模式与方法。旨在为合成人脸图像鉴别方法研究拓展思路。

The network data is exploding with the remarkable development of the Internet and Information Technology. Due to technology abuse and lack of supervision, the synthetic face images flood the network with realistic effect, which poses severe challenges to human eyes and automatic identification systems and causes many legal, ethical and social issues. Taken GANs-oriented synthetic face images as research objects, aiming at problems such as the difficulty of characterizing features, insufficient reference samples, unknown adversarial environment, and the existence of newly generated model attacks in the identification process, the works of feature perception, learning behavior and pattern understanding under the conditions of large sample, few-shot and zero-shot learning are carried out respectively in this project. The research content includes: (1) To address the problem that fine-grained features of synthetic face images are not easy to be characterized, a feature sensing and learning method based on neuron activation response and its compact representation is studied. (2) For the few-shot learning with insufficient reference samples, the online semi-supervised neural network learning strategies based on manifold regularization is constructed. (3) For the zero-shot learning problem of newly generated models, an unsupervised autonomous learning model based on tree-like neural network architecture is discovered. In the end, we expect to provide some new solutions for synthetic face image identification.