CAREER: Towards Fairness in the Real World under Generalization, Privacy and Robustness Challenges

人工智能(AI)算法被广泛应用于各种现实世界的应用中，如社交媒体挖掘和健康信息学。在人工智能算法中确保公平变得越来越重要，以避免放大不平等和强化现有的偏见。虽然公平算法最近取得了很大的进步，但当它们部署到现实世界中时，它们仍然面临着实用的泛化、保密性和健壮性的挑战。首先，在域和时间移位等分布移位的情况下，公平性性能会显著降低。其次，以前的大多数公平算法都需要直接访问准确的人口统计属性，但由于人们的隐私意识和法律法规，这通常是不可行的。此外，研究表明，解决公平问题可能会增加隐私泄露的风险。第三，恶意行为者可以通过在训练阶段注入中毒样本或在推理阶段操纵数据来放大人工智能算法的人口统计学偏见。本项目的目标是研究上述问题对公平性的影响，并在泛化、隐私和健壮性挑战下制定有效的解决方案来确保公平性。为了实现研究目标，本项目系统地研究了域和时间转移下的公平性、隐私机制实施和隐私泄露风险下的公平性、偏差放大攻击和防御方法。该项目的成果有助于推进公平人工智能的最新研究，并介绍了：(1)从信息论的角度和元学习框架来确保时间不变的公平性；(2)在局部差异隐私机制下提高公平性能的算法，实现公平图学习的同时将隐私泄露降至最低；(3)针对公平属性的毒化和逃避攻击，以及相应的以模型为中心和以数据为中心的防御方法。更广泛地说，这个项目将对改进实践中的公平算法产生直接和强烈的影响，使负责任的数据分析能够在现实世界中使用先进的值得信赖的人工智能范例。该奖项反映了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。

Artificial Intelligence (AI) algorithms are widely adopted in various real-world applications such as social media mining and health informatics. It becomes increasingly essential to ensure fairness in AI algorithms to avoid amplifying inequalities and reinforcing existing prejudice. Although fairness algorithms have achieved great progress recently, when deployed in the real world, they still face practical generalization, privacy and robustness challenges. First, the fairness performance can be significantly degraded under distribution shifts such as domain and temporal shifts. Second, most previous fairness algorithms require direct access to the exact demographic attributes, which is usually infeasible due to people's awareness and legal regulations on privacy. Moreover, research indicates that addressing fairness may increase privacy leakage risks. Third, malicious actors can amplify the demographic bias of AI algorithms by injecting poisoning samples in the training stage or manipulating the data in the inference stage. The goal of this project is to investigate the impact of the aforementioned issues on fairness and develop effective solutions to ensure fairness under generalization, privacy and robustness challenges.To achieve the research goal, the project systematically investigates the key directions of fairness under domain and temporal shifts, fairness faced with privacy mechanism enforcement and privacy leakage risks, bias amplification attack and defense methods. The project outcomes help advance state-of-the-art research on fair AI and introduce: (1) fairness in domain adaptation from an information-theoretical perspective and a meta-learning framework to ensure temporal-invariant fairness; (2) algorithms improving fairness performance under local differential privacy mechanism and achieving fair graph learning while minimizing the privacy leakage; and (3) poisoning and evasion attacks on fairness properties, as well as model-centric and data-centric defense methods for such attacks accordingly. More broadly, this project will have an immediate and strong impact on improving fairness algorithms in practices, enabling the responsible data analysis with advanced trustworthy AI paradigms in the real world.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.