面向实际场景的人工智能脆弱性分析

Vulnerability Analysis of Artificial Intelligence in Real World

人工智能技术广泛应用于自动驾驶、无人机、机器人等自主无人系统,是实现场景感知、情报获取、辅助决策等复杂功能的重要支撑。因此,研究人工智能技术的脆弱性和本身安全性问题引起了越来越多的关注。对抗机器学习(adversarial machine learning)是机器学习和计算机安全领域的交叉学科,是人工智能算法普遍面临的挑战之一。文中以实际场景下的人工智能安全性为出发点,梳理了对抗样本发展的起源,形成的机理以及发展脉络。首先从攻击、防御两个方面探究各种方法的原理和优缺点;其次,在分析研究经典算法和适用场景的基础上,研究了在实际场景下智能技术面临的脆弱性和挑战;最后,针对在图像、语音、网络和软件应用等不同领域中所面临的挑战和未来发展趋势做了进一步的分析和展望。

Artificial intelligence is widely used in autonomous unmanned systems such as autonomous driving,unmanned aerial vehicles,robots,etc,which is an important support to realize complex functions such as scene perception,intelligence acquisition,assistant decision-making and so on.Therefore,more and more attention has been paid for the vulnerability and security of artificial intelligence technology.Adversarial machine learning is an interdisciplinary subject in the field of machine learning and computer security.It is one of the challenges that artificial intelligence algorithms are facing.On the basis of the security of artificial intelligence in the actual scene,we introduce the origin,principle and development of generating adversarial examples.Firstly,we explore the principles,advantages and disadvantages of attack and defense.Secondly,based on the analysis of classic algorithms and real world,the vulnerability and challenges faced in the actual scene are studied.Finally,we make a further study on the challenges and future development trend in different fields such as image,voice,network and software application.