对抗制
计算机科学
面部识别系统
欺骗攻击
面子(社会学概念)
深层神经网络
人工智能
深度学习
计算机安全
付款
模式识别(心理学)
机器学习
万维网
社会科学
社会学
作者
Jiahui Yang,Yushi Cheng,Xiaoyu Ji,Wenyuan Xu
标识
DOI:10.1145/3573428.3573621
摘要
Face recognition systems are widely used in various security-crucial applications such as financial payments, device unlocking, and personnel access. With the rapid development of deep learning, face recognition systems nowadays are usually based on deep neural networks (DNNs). However, recent studies have shown that DNN-based face recognition algorithms are vulnerable to adversarial example attacks and thus may suffer from real-world threats. In this paper, we propose Adv-Sticker, a physical adversarial attack against face recognition systems leveraging a small printed sticker. By optimizing both the attack region and the adversarial sticker, we manage to reduce the size of the sticker to 3*3 cm and make it robust across various environmental conditions. Evaluation on four commonly-used face recognition algorithms (Facenet, Mobile-Facenet, Ir152, and Irse50) shows that Adv-Sticker can physically spoof face recognition systems with an overall attack success rate of 96.9% for the dodging attack, and 70.1% for the impersonation attack.
科研通智能强力驱动
Strongly Powered by AbleSci AI