CAREER: Learning in Adversarial and Nonstationary Environments

大多数机器学习算法依赖于数据从固定概率分布中采样的假设。这一假设在实践中经常被违背，导致分类和回归策略远不是最优的，甚至是可靠的。最近的研究表明，对手可以显著损害预处理技术和分类的结果。不幸的是，尽管需要这种技术的应用程序越来越多，但还没有解决一个统一的框架，用于在对手存在的情况下从流数据中学习。本职业将学习理解特征选择何时以及为什么在对手面前失败。本研究不仅关注特征选择失败的原因，还关注特征选择中的黑盒攻击和白盒攻击的可移植性。该项目还提出了开发新的方法来攻击信息理论算法和弹性信息理论特征选择方法。该职业还解决了在有对手存在的非固定环境中学习的问题。将为每个任务执行一组综合的和真实的基准测试。该研究侧重于这一未满足的需求，并解决了来自机器学习不同子领域的各种对抗性学习问题：具体而言，是非平稳环境中特征选择和学习的算法。该研究计划的成功实施将对机器学习和应用驱动领域产生更广泛的影响。该教育计划包括指导和培训目前需求量很大的数据科学家的未来劳动力，在大学的协作学习环境中通过多层次的教育引入机器学习。CAREER项目还包括综合研究，修订研究和学习，以社区为基础的教育研究整合，以吸引更多各级学生参与STEM和机器学习。这个职业将吸引图森的K-12学生，通过实践教学技术促进STEM教育和机器学习。还将根据CAREER的研究成果和机器学习的最新趋势，向数据科学界进行公开讲座。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。

The majority of machine learning algorithms rely on the assumption that data are sampled from a fixed probability distribution. This assumption is often violated in practice, which results in classification and regression strategies that are far from optimal or even reliable. Recent work has shown that an adversary can significantly compromise the outcome of preprocessing techniques and classification. Unfortunately, a unified framework for learning in the presence of an adversary from streaming data has not been addressed despite the growing number of applications that need such techniques. This CAREER will study to understand when and why feature selection fails with an adversary. Not only will this research focus on understanding why feature selection fails, but also the transferability of black and white box attacks on feature selection. This project also proposes to develop novel methods to attack information-theoretic algorithms and approaches for resilient information-theoretic feature selection. This CAREER also addresses the problem of learning in a nonstationary environment with the presence of an adversary. A comprehensive set of synthetic and real-world benchmarks will be performed for each of the tasks. The research focuses on this unmet need and tackles a variety of adversarial learning problems drawn from different subfields of machine learning: specifically, algorithms for feature selection and learning in nonstationary environments.A successful implementation of the proposed research plan will have broader impacts on machine learning and application-driven domains. The education plan includes mentoring and training the future workforce for data scientists, who are currently in high demand, by introducing machine learning through multiple levels of education in a collaborative learning environment at the university. The CAREER project also includes integrated then integration research, revise research and learning with a community-based integration of research in education to draw more students at all levels for STEM and machine learning. This CAREER will engage K-12 students in Tucson to promote STEM education and also machine learning through hands-on teaching techniques. There will also be public talks to the data science community based on the CAREER research outcomes and the most recent trends in machine learning.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.