Deceiving Face Recognition Neural Network with Samples Generated by Deepfool
Jingsong Xue, Yang Yu, Dongsheng Jing · Journal of Physics Conference Series · 2019
Abstract Image-based identity authentication systems have been extensively used recently. However, it holds the risk of illegal use of the images for authentication by others, which will cause a severe threat to personal privacy and security protection. The concept of adversarial samples, which is often utilized in the field of deep learning to deceive classification models, has become more and more popular, and has accumulated lots of significant efforts. Meanwhile, previous work mainly focused on interpreting images by features rather than by the structural characteritics of the human face, which made the deception of face recognition model ineffective. To solve the problem, face recognition neural network deceiving method that is based on Deepfool algorithm is proposed. For a specific face recognition neural network, white-box attack is used to generate adversarial samples, and Euclidean distance is utilized to optimize adversarial samples to obtain face photos with misleading attributes. The proposed approach can be used for privacy and security protection. Experiments on several face testing data sets verify effeteness of our proposed approach.