期刊名称：International Journal of Advanced Computer Science and Applications(IJACSA)
印刷版ISSN：2158-107X
电子版ISSN：2156-5570
出版年度：2021
卷号：12
期号：2
页码：564-569
DOI：10.14569/IJACSA.2021.0120270
出版社：Science and Information Society (SAI)
摘要：In biometric applications, deep neural networks have presented significant improvements. However, when presenting carefully designed input training data known as adversarial examples, their output is severely reduced. These types of attacks are termed as adversarial attacks, and any biometric security system is greatly affected by these attacks. In the proposed work, an effective defensive mechanism has been developed against adversarial attacks which are introduced in iris images. The proposed defensive mechanism is following the concept of wavelet domain processing and it investigates the mid and high frequency components of wavelet domain components. Based on this, the model reproduces the various denoised copies of input iris images. The proposed strategies are intended to denoise each sub-band of the wavelet domain and assess the sub-bands most likely to be affected by the adversary using the reconstruction error measured for each sub-band. We test the effectiveness of the proposed adversarial protection mechanism against various attack methods and analyzed the results with other state of the art defense approaches.
