模式识别实验室知识库

With the development of generative models, new types of fake iris have emerged. Distinguished from traditional spoofing means caused by cosmetic contact lenses, such iris images are realistic and easily accessible, which poses a threat to privacy protection and information security. In this paper, we are the first to study iris forgery detection method that can simultaneously defend against contact lenses based or GAN-generated spoofing attacks. Through multi-model ensemble, we design a simple but effective detection framework. The backbone part of our method consists of three CNN networks, including ResNet-18, EfficientNet-B0 and ConvNeXt-tiny. We conduct experiments on three public iris datasets and a great deal of StyleGAN-generated iris images which are collected by ourselves. The proposed method has been proved to be effective on the detection of various iris forgeries, and it has the state-of-the-art performances.