对抗样本的攻防算法研究

Study on adversarial attacks and defenses for adversarial examples

深度学习技术的深入发展推动了人工智能的崛起。但在深度学习中,对抗样本的存在为人工智能技术带来了极大的安全隐患,制约了深度学习和人工智能的发展与普及。文章主要介绍了对抗样本的概念,阐述了对抗样本的线性解释。之后,分析了几种经典的对抗攻击算法和对抗防御算法。在此基础上,文章根据研究进展分析了经典方法的有效性并提出了未来研究的挑战。

Deep learning is the core of the rise of artificial intelligence.The existence of adversative samples brings great security risks to deep learning and restricts the development and application of deep learning.This paper mainly introduces the concept of adversarial sample and expounds the linear interpretation of adversarial sample.After that, several classical adversarial attack algorithms and adversarial defense algorithms are analyzed.On this basis, the effectiveness of the method is analyzed according to the research progress and the challenges for future research are proposed.