Collaborative Research: AF: Small: Exploring the Frontiers of Adversarial Robustness

在人工智能(AI)深度融入我们的生活但网络和物理威胁始终存在的时代，对于面对故意攻击和操纵的算法设计的可靠性、安全性和可信性来说，对对手输入的健壮性至关重要。不幸的是，许多标准的大数据算法容易受到敌意输入的利用，导致人工智能在医疗保健和金融等安全关键领域的广泛部署遭遇重大障碍。该项目的目标是确定和处理对抗性稳健性方面新出现的方向，并制定对抗性投入脆弱性的基本原则。该项目不仅将实现算法设计和数学工具的新元素，还将对值得信赖的人工智能的广泛应用产生立竿见影的影响。研究小组将让研究生参与这一项目，并计划就对抗稳健性和适应性数据分析的新方向举办两次研讨会。此外，该团队将启动与当地一所高中的外联活动，以指导5-12年级的学生。该项目的目标是开发新的大数据算法，该算法对对手的输入具有健壮性。在高层次上，该项目的主要关注点是：1)黑盒流传输环境中的对抗稳健性，其中对手可以访问先前的输出但不能访问算法的内部状态；2)白盒流传输环境中的对抗稳健性，其中对手另外可以访问算法使用的内部状态和先前的随机比特；以及3)有限空间的自适应数据分析。此外，研究团队将在新的攻击模型出现时探索和整合它们，允许不断适应不断变化的挑战。这一奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

In an age where artificial intelligence (AI) is deeply integrated into our lives yet both cyber and physical threats are ever-present, robustness to adversarial input is critically important for ensuring the reliability, security, and trustworthiness of algorithmic design in the face of intentional attacks and manipulations. Unfortunately, many standard big data algorithms are susceptible to exploitation by adversarial input, resulting in major roadblocks to the widespread deployment of AI in safety-critical domains like healthcare and finance. The goal of this project is to identify and address new emerging directions in adversarial robustness and to develop the fundamental principles underlying vulnerabilities to adversarial input. The project will not only realize new elements of algorithmic design and mathematical tools, but also have immediate impact on the wide-ranging applications of trustworthy artificial intelligence. The research team will involve graduate students in this project and plans two workshops on new directions in adversarial robustness and adaptive data analysis. In addition, the team will initiate outreach with a local high school to mentor students in grades 5-12.The goal of this project is to develop new big data algorithms that are robust to adversarial input. At a high level, the primary focal points of this project are: 1) adversarial robustness in the black-box streaming setting, where an adversary has access to the previous outputs but not the internal states of the algorithm, 2) adversarial robustness in the white-box streaming setting, where an adversary additionally has access to the internal state and previous random bits used by the algorithm, and 3) adaptive data analysis with bounded space. Additionally, the research team will explore and integrate new attack models as they emerge, allowing for continuous adaptation to evolving challenges.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.