Collaborative Research: ATD: Fast Algorithms and Novel Continuous-depth Graph Neural Networks for Threat Detection

在算法威胁检测中，理解多变量时间序列之间的相互作用是至关重要的。具有注意力机制的图神经网络(GNN)已被证明在学习和预测这类时间序列方面是有效的。该项目旨在研究GNN以提高加速和精度。这项研究将在人工智能(AI)、交通分析、电力系统和健康分析等领域具有广泛的适用性。该项目将为人数不足的学生提供培训机会并促进STEM教育。该项目旨在解决在多变量时间序列中进行威胁检测的三个关键挑战：1)利用深层GNN保持准确性，2)使用有限的数据训练GNN，3)减少训练和部署具有关注层的深层GNN的计算成本。提出了基于偏微分方程(PDE)理论的连续深度神经网络和高效注意算法。通过利用GNN的连续观点，该项目旨在开发具有理论基础和计算效率的算法，用于在有限监督下进行准确的图形深度学习。该项目将集中于三个研究方向：推力A：桥接扩散方程理论和GNN体系结构设计，以开发一类新的基于图上扩散方程的GNN。这些GNN克服了过度平滑，在有限的监督下可靠地学习和预测。推力B：为GNN和注意力训练、测试和推理开发快速算法。推力C：新算法在异常检测和软件开发中的应用，特别是在基准图学习任务和交通流、电力分配和疫情数据中的异常检测中。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。

In algorithmic threat detection, understanding the interactions of multivariate time series is crucial. Graph neural networks (GNNs) with attention mechanisms have proven effective in learning and predicting such time series. This project aims to investigate GNNs for improved acceleration and accuracy. The research will have broad applicability in fields such as Artificial Intelligence (AI), traffic analysis, power systems, and health analytics. The project will provide training opportunities and promote STEM education for underrepresented students.The project aims to address three key challenges in threat detection within multivariate time series: 1) maintaining accuracy with deep GNNs, 2) training GNNs with limited data, and 3) reducing computational costs in training and deploying deep GNNs with attention layers. The research advances continuous-depth GNNs and efficient attention algorithms based on the partial differential equation (PDE) theory. By leveraging the continuous viewpoint of GNNs, the project aims to develop theoretically-grounded and computationally efficient algorithms for accurate graph deep learning with limited supervision. The project will focus on three research thrusts: Thrust A: Bridging diffusion equation theory and GNN architecture design to develop a new class of GNNs based on diffusion equations on graphs. These GNNs overcome over-smoothing and reliably learn and predict with limited supervision. Thrust B: Developing fast algorithms for GNN and attention training, testing, and inference. Thrust C: Application of the new algorithms to anomaly detection and software development, specifically in benchmark graph learning tasks and anomaly detection in traffic flow, power distribution, and epidemic data.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.