Research

Our research focuses on vulnerabilities of machine learning-enabled systems, on methodologies for detecting attacks and on methodologies for making the systems robust to attacks. We have developed detection algorithms for physically realizable adversarial attacks (patch attacks) against computer vision algorithms, as well as algorithms for robustness against such attacks. We have developed detectors for attacks against cooperative multi-agent reinforcement learning as well as methodology for training robust cooperative multi-agent reinforcement policies. Our work combines theory with algorithm development and experimental validation.

Edge computing brings computing and storage resources closed to the network edge, potentially enabling a wide range of novel applications and services on energy-constrained Internet of things devices, including cooperative autonomous systems, computer vision-based services, automated surveillance for public safety, etc. Our research has focused on the joint management of communication and computing resources at the network edge, joint pricing and resource management for edge services and on performance isolation for machine learning workloads at the edge. Our work provides algorithmic solutions to these problems through methodological contributions in the area of distributed optimization, game theory, and machine learning, combined with trace-based and experimental evaluations.

Cyber-physical systems are pervasive in modern society, from electric power systems to unmanned aerial vehicles. They are physical systems that are monitored, controlled and managed by IT systems, and thus they heavily rely on communication and computing infrastrures. Our research has focused on exploring the vulnerabilities of cyber-physical systems, including electric power systems as well as unmanned aerieal vehicles, and on developing methodologies for mitigating the impact of potential attacks on these systems. We have contributed to the characterization of vulnerabilities in power system state estimation, time synchronization in power systems, as well as to the detection of attacks combining methodologies including time series analysis, machine learning, optimization and game theory.

Intrusion detection and response are essential for protecting modern networked systems from sophisticated adversaries. Our work studies problems including the detection of anomalous behavior in network traffic, identifying and localizing attackers in complex networked systems, and designing mechanisms that support effective semi-autonomous threat response, taking into account that threat response involves human interaction. Our work combines principled modeling and learning approaches, ranging from sequential hypothesis testing and active learning to transformer based approaches to anomaly detection, developing mechanisms that can significantly improve detection speed, precision, and robustness against sophisticated attacks.

Content delivery systems are a key component of the Internet, enabling scalable distribution of large volumes of content such as video streams to millions of users. Our work focuses on the design and analysis of peer-to-peer (P2P) streaming and content distribution systems, addressing problems such as overlay topology design, incentive mechanisms, robustness to churn, and efficient resource allocation. Our results provide algorithms and analytical models that improve scalability, stability, and resilience of P2P streaming and content delivery systems, including Bittorrent like systems, demonstrating that decentralized architectures can deliver high performance while reducing infrastructure costs.

As the demands on our communication infrastructures continue to rise, it becomes increasingly important that they operate with maximum efficiency. In wireless networks, it becomes critical to manage radio resources, such as transmit powers and shared spectrum. In data networks, it becomes essential to understand how the data traffic flows and fluctuates to direct the data from source to destination in the most efficient way. We have worked with both wired and wireless data networks, with a particular focus on the design of distributed resource management mechanisms which allow these networks to operate close to their practical capacity.