ATD: Trustworthy Artificial Intelligence for Threat Detection

The primary objective of threat detection is to identify incidents that may pose a risk to computer systems, networks, data, social activities, or living communities. Advanced artificial intelligence (AI) algorithms, including machine learning and statistical learning methods, can analyze large amounts of data to identify features indicative of threats. Despite the success and promise of AI algorithms, they present challenges and concerns that warrant careful consideration. Many AI algorithms operate as "black boxes" and lack transparency in their decision-making processes, which can be problematic in critical applications. Additionally, inadequate training data can lead to biases, lack of robustness, and overfitting, resulting in inaccurate predictions, especially with new or unseen data. This project aims to mathematically address some challenges of trustworthy threat detection using AI algorithms, focusing specifically on effective learning from a very small number of samples, known as few-shot learning. The PI will investigate various problems related to few-shot learning. One such problem is few-shot domain adaptation, which uses knowledge from a related domain to build models for a target domain with limited unlabeled data. This is particularly relevant to automated threat detection, where events of interest are rare and have few examples. Effective metrics for learning from minimal target data will be explored. Another focus is few-shot graph generation, which can advance the theoretical understanding of machine learning on graphs with very limited training data. This is crucial for building trustworthy systems, as many relevant threat detection scenarios involve social and transportation networks. Since the most threatening events are rare, a few-shot generation is necessary. A primary emphasis of this proposal is on the mathematical understanding of learning effectively with very limited information. 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.