CAREER: Taming Networks in the Wild: A Safety-Centric Network Learning Framework

尽管最近取得了成功，但人工智能（AI）在安全等领域也有所欠缺。因此，必须优先考虑安全性，将其作为人工智能研究各个领域的基本方面。值得注意的是，由于网络（图）数据的无处不在，网络学习技术在过去十年中对人工智能产生了巨大影响，广泛部署在社交网络、医疗保健和网络安全等应用中。然而，现实世界的网络通常包括数据问题和不可预见的环境危害等挑战，导致使用这些网络时存在风险的人工智能技术和不安全的结果。现有的安全网络学习研究缺乏通用性，效率和跨多个安全维度的综合集成。它们难以确保在实际场景中进行及时、安全的预测，无法全面解决数据、模型和使用方面的问题。这些挑战共同阻碍了他们的能力和有效性，目前缺乏一个全面的框架来充分解决网络学习中的安全问题。为了弥合这一研究差距，本项目的目标是设计，开发和评估一种新的安全为中心的网络学习框架（SNL）在网络上的安全决策。该项目的成果将对网络学习研究产生重大影响，并提供先进的解决方案，以应对公共卫生、网络安全和社交媒体等不同领域的挑战。此外，该项目将促进跨学科合作，并促进向工业界的技术转让。项目成果将向公众开放并广泛传播。此外，该项目将通过新颖的课程开发和学生辅导活动将研究与教育结合起来，重点关注代表性不足的群体，旨在培训和教育后代有效开发和利用人工智能，同时确保人工智能安全。该项目将全面努力开发SNL，优先考虑可靠性，稳定性和可解释性的关键安全维度，并进一步以通用，高效和集成的方式涵盖数据，模型和使用的关键方面。形式上，SNL将提供可靠的网络数据和网络学习模型（可靠性），同时生成稳定和一致的输出（稳定性），并伴有易于理解的使用解释（可解释性）。具体来说，在这个项目中，涉及创新理论，算法和模型的研究组成部分是四重的。首先，设计新颖的网络学习算法，以识别和生成受数据和环境问题影响最小的可靠网络数据。第二，创建新的网络表征学习模型和训练策略，以提高学习网络表征的效率和可靠性。第三，设计创新的数据到模型优化理论，以确保网络学习的稳定性。最后，开发新的生成学习方法，以提高网络学习的使用可解释性和输出可接受性。统一的框架允许不同研究组成部分之间的无缝协作和相互加强。通过融合研究计划，该项目不仅将在网络学习和人工智能安全研究方面取得重大进展，还将为应对各种社会挑战提供新的见解，最终造福整个社会。该奖项反映了NSF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。

Despite recent successes, Artificial Intelligence (AI) has also shown to be lacking in areas such as safety. Therefore, it is imperative to prioritize safety as a fundamental aspect across various domains of AI research. Notably, due to the ubiquity of network (graph) data, network learning techniques have strongly impacted AI in the past decade, being widely deployed across applications such as social networks, healthcare, and cybersecurity. However, real-world networks often include challenges like data issues and unforeseen environmental hazards, resulting in risky AI techniques and unsafe outcomes when these networks are employed. Existing studies for safe network learning lack versatility, efficiency, and comprehensive integration across multiple safety dimensions. They struggle to ensure timely, safe predictions in practical scenarios and fail to holistically address data, model, and usage aspects. These challenges collectively hinder their capability and effectiveness, and there is currently a lack of a holistic framework to adequately tackle safety issues in network learning. To bridge this research gap, the goal of this project is to design, develop, and evaluate a novel Safety-centric Network Learning framework (SNL) for safe decision-making on networks in the wild. The project outcomes will substantially impact network learning research and offer advanced solutions to address challenges across diverse domains such as public health, cybersecurity, and social media. Additionally, the project will foster interdisciplinary collaborations and facilitate technology transfer to industry. The project outcomes will be made publicly accessible and broadly disseminated. Moreover, the project will integrate research with education through novel curriculum development and student mentoring activities with an emphasis on underrepresented groups, aiming to train and educate future generations in effectively developing and utilizing AI while also ensuring AI safety. The project will involve comprehensive efforts to develop SNL that prioritizes the critical safety dimensions of reliability, stability, and explainability, and further encompasses the crucial aspects of data, model, and usage in a general, efficient, and integrated manner. Formally, SNL will provide reliable network data and network learning models (reliability) while generating stable and consistent outputs (stability) accompanied by easily understandable usage explanations (explainability). Specifically, the research components that engage innovative theories, algorithms, and models in this project are fourfold. First, design novel network learning algorithms to identify and generate reliable network data that are minimally impacted by data and environmental issues. Second, create new network representation learning models and training strategies to promote efficiency and reliability in learning network representations. Third, devise innovative data-to-model optimization theories to ensure the stability of network learning. Finally, develop novel generative learning methods to advance the usage explainability and output receptivity of network learning. The unified framework allows seamless collaboration and mutual reinforcement among different research components. Through the convergent research program, the project will not only make significant advancements in network learning and AI safety research but also shed novel insights to tackle various societal challenges, ultimately benefiting society at large.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.