Over the past few years, machine learning has been widely utilized to enhance cybersecurity. In this talk, we will present two of our recent projects in this area: the use of autoencoders to defend against adversarial attacks and the detection of darknet traffic over Tor. In the first project, we observed that machine learning-based intrusion detection systems (IDSs) have become increasingly popular due to their strong detection capability and adaptability.

What does it feel like to cross the road in front of a car with no driver? Or to watch a delivery robot get stuck on a footpath and wonder “do I help”? This presentation takes these everyday encounters seriously. Drawing on research in autonomous vehicles–pedestrian interaction and human–robot collaboration, we explore how pedestrians and bystanders navigate a world increasingly shared with autonomous systems, and what design research can do to make those interactions feel more natural, legible, and human.

Autonomous Vehicle (AV) technology holds the promise of transforming future transportation systems, but its large-scale deployment requires rigorous safety evaluation. To address this need, we propose a behavioral safety assessment program to evaluate AV driving intelligence, consisting of two key components: (1) high-fidelity simulations that replicate real-world driving conditions, and (2) efficient testing strategies to accelerate validation. Central to this is modeling a Naturalistic Driving Environment (NDE) with distribution-level accuracy. We develop a deep learning-based framework to simulate statistically realistic human driving behavior, particularly for those long-tail safety-critical events. However, directly testing in the NDE is inefficient due to the rarity of such events. To address this, we introduce the Naturalistic and Adversarial Driving Environment (NADE), reducing test mileage without sacrificing accuracy. The proposed AV behavioral safety assessment program has been successfully deployed at physical testing facilities for real-world AV evaluation.

Modular autonomous vehicles (MAVs) are transforming transit design through en-route (de)coupling and in-motion transfer. To evaluate this disruptive technology, we conduct a comprehensive cost-effectiveness analysis of an MAV-based system featuring a main corridor service integrated with local pick-up-and-delivery routes. We develop a continuum approximation (CA) model to quantify passenger and agency costs, comparing it against traditional fixed-route and on-demand services. Our analysis uncovers the complex interactions between system parameters and performance, identifying favorable operating regimes and providing managerial guidance for modular transit implementation.

Ninety years ago, Greenshields’ 1935 study of vehicular flow introduced what we now call the fundamental diagram. It captures a basic congestion effect: as density increases, speed drops from its free-flow value and goes to zero as density approaches the jam level. This simple relation gives a constitutive law that closes macroscopic traffic equations, and it underpins classical continuum models such as LWR and later second-order variants. In parallel, optimal transport has developed into a general framework for how distributions move under constraints. It is now widely used in swarm control, inference, and diffusion-based generative modeling. The viewpoint is to evolve densities rather than track individuals, and to choose the evolution that minimizes an action, often a kinetic-energy-type cost.

Transportation systems are essential to economic activity and social mobility, yet they face growing challenges from rapid urbanization, increasing travel demand, and the expanding integration of emerging technologies. This talk will cover part of my research that aims at advancing efficient and robust transportation systems, from both the operational decision-making and infrastructure planning perspective.

Structural health monitoring (SHM) is crucial for ensuring the safety, durability, and operational efficiency of civil infrastructure such as bridges and buildings. Traditional monitoring methods are limited by low accuracy, high computational costs, and poor adaptability to complex working conditions, making it difficult to meet the demands of real-time and intelligent evaluation. Computational intelligence has emerged as a quite popular means for advancing SHM, addressing these limitations through applying innovative algorithmic structures or modifications.

This talk will address the critical challenge of ensuring the robustness of graph learning models by walking through a structured, three-step processing pipeline. We will begin by dissecting adversarial attacks, exploring how graph-structured data can be subtly manipulated to mislead model predictions. The discussion will then pivot to defense strategies, first covering empirical defense methods that have shown practical promise but lack formal guarantees. Subsequently, we will delve into the more rigorous field of certified...

Modern global supply chains exhibit complex, multi-tier dependencies that remain largely invisible to decision-makers, creating vulnerabilities to cascading disruptions. This presentation introduces a novel AI-driven methodology that combines data scraping, generative AI, and network science to illuminate deep-tier supply chain networks and quantify systemic risks. The approach leverages diverse data sources—including 10-K filings, earnings calls, and trade databases—to construct large-scale supply chain networks spanning over 55,000 firms, 230,000 relationships, 93 countries, and 278 industries, enabling...

Unlike previous generations of wireless networks, which were primarily designed to meet the requirements of human communications, 5G networks enable extensive data collection from machines. As we transition to 6G, the emphasis moves beyond connectivity toward leveraging this machine-generated data for a new spectrum of control applications, such as UAV swarms, collaborative robots, and cooperative autonomous vehicles. Designing communication systems for these advanced control applications introduces a distinct set of challenges. These include meeting stringent...