A Theoretical Foundation and Practical Platform for Adversarial Machine Learning

最近机器学习模型取得了令人印象深刻的成功，这使它们成为几乎所有需要大量数据分析的应用程序的有希望的候选解决方案。有了如此光明的前景，随之而来的是严格的审查：这些模型对对抗性攻击的鲁棒性令人惊讶。已经进行了许多实证研究，有时得出看似矛盾的结论。在这个项目中，我们的目标是为对抗性机器学习的研究建立一个共同的理论基础和实践平台。我们提供了一种公理方法来严格定义、计算和优化鲁棒性，以及其他更传统的指标，我们开发了一个平台，允许用户尝试不同的鲁棒性概念，并在准确性、鲁棒性和效率之间进行不同的权衡。我们的工作将在一个非常需要的统一和严格的框架内大大澄清和精简现有的经验工作。

The recent impressive success of machine learning models has made them a promising candidate solution for virtually every application that requires extensive data analysis. With such bright promise comes the great scrutiny: these models are surprisingly non-robust against adversarial attacks. Many empirical studies have been performed, sometimes with seemingly contradicting conclusions. In this project, we aim to build a common theoretical foundation and a practical platform for performing research in adversarial machine learning. We provide an axiomatic approach to rigorously define, compute, and optimize robustness, along with other more conventional metrics, and we develop a platform to allow users to experiment with different notions of robustness and with different trade-offs among accuracy, robustness and efficiency. Our work will greatly clarify and streamline existing empirical work in a much needed unified and rigorous framework.