CAREER: Human-Machine Supervision Cycle for Trustworthy Biometrics

生物识别攻击检测的主流方法对偏离真实信息的类型做出了强有力的假设。这造成了实验室环境中观察到的可靠性与现实世界中预期的性能之间的严重差距，在现实世界中，未来的攻击是未知的。该项目填补了这一空白，并在人工智能（AI）和人类之间建立了有效的共生关系。该项目的新颖之处是(a)允许人工智能有效地向人类学习如何提高对生物识别系统的未知攻击的可检测性，以及(b)支持人类检查虚假生物识别输入的新方法。该项目更广泛的意义和重要性是：(a)可靠的生物识别系统，可以更好地识别从未见过的演示攻击，从而更好地保护消费者设备，银行账户并加强美国边境控制程序；(b)一个强大的教育计划，使K-12，本科生和研究生接触到生物识别技术的安全和伦理相关方面，并拓宽他们在国家关注的相关主题方面的知识；(c)由调查人员编写的公开讲座，这将扩大对负责任地使用生物识别技术的认识。在本项目中，将建立一个人机监督周期的整体框架，以实现(a)人类指导的计算机视觉方法设计，使生物特征呈现攻击检测机制更好地推广到未知的攻击工具，以及(b)创建计算机辅助方法，协助人类审查员检测虚假输入。本项目的基本技术贡献包括(1)拓宽了关于控制人类对虚假视觉信号感知的机制的知识，(2)发现了最有效的人类可解释的信息表示，以支持他们的决策并加快他们对新型攻击的学习。(3)开发信任评估的定量指标，并将人类和机器的决策连接成一个可信赖的串联，从而更好地判断生物特征输入的真实性；(4)将该框架应用于新生儿的虹膜识别，有助于与母亲和/或监护人更好地联系，从而提高从医疗保健系统中受益的机会，特别是在发展中国家。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Dominant approaches of biometric attack detection make strong assumptions about the type of deviations from authentic information. This creates a critical gap between reliability observed in laboratory settings and the performance expected in the real world, where future attacks are unknown. This project fills this gap and builds an effective symbiosis between Artificial Intelligence (AI) and humans. The project novelties are new methods that (a) allow the AI to effectively learn from humans how to increase detectability of unknown attacks on biometric systems, and (b) support humans in their examination of fake biometric inputs. The project's broader significance and importance are: (a) trustworthy biometric systems that better recognize never-seen presentation attacks, and thus better protect consumer devices, bank accounts and strengthen the US border control processes; (b) a strong educational program that exposes K-12, undergraduate and graduate students to both the security- and ethics-related aspects of biometrics, and broadens their knowledge in a relevant topic of national concern; (c) publicly available lectures prepared by the investigator, which will broaden the awareness of responsible use of biometrics.In this project, a holistic framework for human-machine supervision cycle will be established to enable (a) human-guided design of computer vision methods to make the biometric presentation attack detection mechanisms generalize better to unknown attack instruments and (b) creation of computer-aided methods of assisting human examiners in detecting of fake inputs. Fundamental technical contributions of this project include (1) broadening knowledge about mechanisms that govern human perception of fake visual signals, (2) discovering the most effective human-interpretable representations of information to support their decisions and speed up their learning of new types of attacks, (3) developing quantitative metrics of trust assessment and linking human and machine decisions into a trustworthy tandem that makes better judgements on the authenticity of biometric inputs, and (4) application of the framework to iris recognition of newborns, contributing to a better linkage with mothers and/or guardians, resulting in improved chances to benefit from healthcare systems, especially in developing countries.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.