Contextual and Temporal Generative Time-Series Modeling
Abstract:
Generative time-series modeling delineates the problem of producing an accurate representation of sequential data or target variables based on observations. By learning the distributions of observed sequences with a probabilistic model, we aim to generate new sequential data points that effectively approximate the distribution of a given dataset. It is a challenging problem. Artificial intelligence, machine learning, and statistical modeling are essential to these endeavors. Through the advancement of contextualization in these disciplines, state-of-the-art models can generate realistic synthetic time-series data and even fabricate data of previously unobserved phenomena. In this dissertation, we present select problems in security and blockchain networks, presenting how time-series generative modeling is applied to advance these domains. We offer three works in two areas, demonstrating that auxiliary contextual information enhances synthetic data generation. In the first work, we propose the adaptive and online learning of network traffic to filter DDoS attacks. It demonstrates that our intrusion detection system performs well on publicly available datasets. The second work leverages contextual information as a control vector to generate poisoning attack traffic against online DDoS filtering. Lastly, we identify another application area, blockchain network transactions. The third work is on the contextual generation of Non-Fungible Token (NFT) transactions that project the value of NFT tokens over time. We then present a suite of new approaches and analyses. Collectively, these results provide a partisan path toward the discovery and use of contextual generative modeling that maximizes the synthetic generation of data.