Evaluation and Comparison of Five Popular Fake Face Detection Networks [5种流行假脸视频检测网络性能分析和比较]

作者：Gao Y.; Hu Y.^*; Yu Z.; Lin Y.; Liu B.

来源：Journal of Applied Sciences, 2019, 37(5): 590-608.

DOI：10.3969/j.issn.0255-8297.2019.05.002

• Data augmentation
• Dataset partition
• Deep neural network
• Fake face video detection
• Generalization
• Threshold selection

Several fake face detectors based on convolutional neural network (CNN) have been reported to resist the impact of fake faces, but they all face a common problem that the intra-dataset test is generally with high accuracy, but the performance of crossdataset test drops significantly, which indicates low generalization ability. Based on thorough evaluations for five popular fake face detectors including MesoInception-4, MISLnet, ShallowNetV1, Inception-v3 and Xception, this paper completes both intra-dataset test and cross-dataset test on three fake face datasets. In experiment, the effects on generalization ability from of factors, such as dataset partition, data augmentation and threshold selection, are investigated.

• 单位
  华南理工大学