III: Small: Enhancing Adversarial Robustness of Geospatio-Temporal Models

Recent advancements in artificial intelligence (AI) have significantly enhanced forecasting capabilities across various geo-scientific domains. In weather prediction, for example, sophisticated geospatio-temporal AI models have demonstrated higher accuracy and lower computational costs compared to traditional physics-based approaches. However, these models are vulnerable to adversarial attacks, where subtle modifications to input data may result in incorrect predictions. Since critical sectors such as agriculture, energy, transportation, and disaster management rely on accurate weather forecasts for planning and resource allocation, such manipulations could lead to poor management decisions, improper resource allocation, and inadequate disaster preparedness. This project will develop robust AI-based techniques for geo-scientific applications. By strengthening the reliability of the AI models, the project will contribute to more effective decision-making, benefiting both the nation and society. The investigators will also conduct various training and outreach activities to expand its impact and ensuring the methods and insights from the project are widely shared. These activities include mentoring students to conduct interdisciplinary research in AI for geo-scientific domains and participating in the annual Science Festival at Michigan State to showcase the research findings through presentations and hands-on demonstrations. The primary objectives of this project are to examine the vulnerabilities of existing geospatio-temporal AI models to adversarial attacks and to develop effective solutions that enhance their resilience. This will be accomplished by proactively identifying and mitigating these vulnerabilities to ensure stronger protection against potential threats. The innovations and technical contributions of this research will offer a deeper understanding of adversarial attacks on geospatio-temporal AI models and their implications for downstream applications. To achieve these goals, the project focuses on three key areas: (1) Assessing the adversarial robustness of current geospatio-temporal AI models, (2) Designing localized, targeted attacks that exploit the spatio-temporal dependencies in these models, and (3) Making Robust the geospatio-temporal AI models to withstand such attacks. This research will also contribute to scientific advancement by exploring the cascading impact of adversarial samples on downstream applications and adversarial attack techniques for ensemble models, two topics that have not been well-studied. To evaluate their effectiveness, the proposed techniques will be validated using deep learning-based weather forecasting models as a case study. 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.