About

Matthew Caesar is a professor at the Siebel School of Computing and Data Science at the University of Illinois Urbana-Champaign. He holds a Ph.D. in Computer Science from the University of California, Berkeley, obtained in 2007. His research focuses on simplifying management and improving the reliability of distributed systems and networks through principles of self-organization and self-diagnosis, with particular emphasis on wide-area networks and networked systems. Caesar's work concentrates on the design, analysis, and implementation of large-scale distributed systems and networks, especially in the context of network operations, measurement, and availability. His efforts aim to enhance the availability and performance of Internet infrastructure, including routing, DNS, and data centers.

Research topics

• Physics
• Astrophysics
• Computer Science
• Astronomy
• Mathematics
• Theoretical physics
• Demography
• Optics
• Quantum mechanics
• Computational physics
• Mathematical analysis

Selected publications

• Verifying Multi-vendor IoT Deployments Using Conditional Tables

  Lecture notes of the Institute for Computer Sciences, Social Informatics and Telecommunications Engineering · 2026-01-01

  book-chapter
  PublisherDOI

• Communication-Aware Multi-Agent Reinforcement Learning for Decentralized Cooperative UAV Deployment

  ArXiv.org · 2026-03-17

  articleOpen access

  Autonomous Unmanned Aerial Vehicle (UAV) swarms are increasingly used as rapidly deployable aerial relays and sensing platforms, yet practical deployments must operate under partial observability and intermittent peer-to-peer links. We present a graph-based multi-agent reinforcement learning framework trained under centralized training with decentralized execution (CTDE): a centralized critic and global state are available only during training, while each UAV executes a shared policy using local observations and messages from nearby neighbors. Our architecture encodes local agent state and nearby entities with an agent-entity attention module, and aggregates inter-UAV messages with neighbor self-attention over a distance-limited communication graph. We evaluate primarily on a cooperative relay deployment task (DroneConnect) and secondarily on an adversarial engagement task (DroneCombat). In DroneConnect, the proposed method achieves high coverage under restricted communication and partial observation (e.g. 74% coverage with M = 5 UAVs and N = 10 nodes) while remaining competitive with a mixed-integer linear programming (MILP) optimization-based offline upper bound, and it generalizes to unseen team sizes without fine-tuning. In the adversarial setting, the same framework transfers without architectural changes and improves win rate over non-communicating baselines.

  PublisherOA PDF

• Declarative Debugging for Modern Networks

  Lecture notes in computer science · 2026-01-01

  book-chapterSenior author
  PublisherDOI

• Communication-Aware Multi-Agent Reinforcement Learning for Decentralized Cooperative UAV Deployment

  arXiv (Cornell University) · 2026-03-17

  preprintOpen access

  Autonomous Unmanned Aerial Vehicle (UAV) swarms are increasingly used as rapidly deployable aerial relays and sensing platforms, yet practical deployments must operate under partial observability and intermittent peer-to-peer links. We present a graph-based multi-agent reinforcement learning framework trained under centralized training with decentralized execution (CTDE): a centralized critic and global state are available only during training, while each UAV executes a shared policy using local observations and messages from nearby neighbors. Our architecture encodes local agent state and nearby entities with an agent-entity attention module, and aggregates inter-UAV messages with neighbor self-attention over a distance-limited communication graph. We evaluate primarily on a cooperative relay deployment task (DroneConnect) and secondarily on an adversarial engagement task (DroneCombat). In DroneConnect, the proposed method achieves high coverage under restricted communication and partial observation (e.g. 74% coverage with M = 5 UAVs and N = 10 nodes) while remaining competitive with a mixed-integer linear programming (MILP) optimization-based offline upper bound, and it generalizes to unseen team sizes without fine-tuning. In the adversarial setting, the same framework transfers without architectural changes and improves win rate over non-communicating baselines.

  PublisherDOI

• mmWave-SAR dataset: large high-resolution heatmap and point cloud dataset for static object detection and other machine-learning applications

  2025-04-11

  articleSenior author

  This research paper introduces a novel dataset collection methodology as well as a sample dataset for mmWave radars that applies synthetic aperture radar (SAR) techniques to enhance object detection and machine learning applications. The dataset features high-resolution heatmaps and point clouds, capturing static objects in various environments. By utilizing SAR, the spatial accuracy and resolution of mmWave data are significantly improved, resulting in more detailed object representations. This approach overcomes the typical resolution limitations of mmWave technology, delivering a high-quality dataset that performs well in noisy, complex environments where traditional sensors often fail.

  PublisherDOI

• AutoPGT: LLM-Driven Automated Policy Generation for Securing Industrial Control Systems

  2025-09-29

  article

  Industrial Control Systems (ICSs) face growing exposure to cyberattacks due to increasing interconnectivity and continued digitization. Effective protection from these threats requires well-designed security policies that are adapted to the target system’s unique operating environment. However, security policy development remains manual and error-prone, constrained by system complexity and the need for specialized technical expertise. To address these limitations, this paper presents AutoPGT, an Automated Policy Generation Tool that produces security policy code based on natural language descriptions and system data. The tool uses Large Language Models (LLMs) to first generate reusable system-agnostic policy templates from user descriptions, and then combines them with system data to produce system-specific policy code. We evaluate AutoPGT by gauging the alignment between its drafted policies and the original user intent under different generation configurations. Our results show that including representative examples and requiring the model to justify its choices yields stronger alignment between generated policies and user intent. Given these findings, AutoPGT enables the development of intuitive and robust security policies that strengthen the resilience of critical cyber-physical systems.

  PublisherDOI

• SIFT-Graph: Benchmarking Multimodal Defense Against Image Adversarial Attacks With Robust Feature Graph

  ArXiv.org · 2025-11-11

  preprintOpen accessSenior author

  Adversarial attacks expose a fundamental vulnerability in modern deep vision models by exploiting their dependence on dense, pixel-level representations that are highly sensitive to imperceptible perturbations. Traditional defense strategies typically operate within this fragile pixel domain, lacking mechanisms to incorporate inherently robust visual features. In this work, we introduce SIFT-Graph, a multimodal defense framework that enhances the robustness of traditional vision models by aggregating structurally meaningful features extracted from raw images using both handcrafted and learned modalities. Specifically, we integrate Scale-Invariant Feature Transform keypoints with a Graph Attention Network to capture scale and rotation invariant local structures that are resilient to perturbations. These robust feature embeddings are then fused with traditional vision model, such as Vision Transformer and Convolutional Neural Network, to form a unified, structure-aware and perturbation defensive model. Preliminary results demonstrate that our method effectively improves the visual model robustness against gradient-based white box adversarial attacks, while incurring only a marginal drop in clean accuracy.

  PublisherOA PDFDOI

• mmPrism: High-Throughput Perception-Aware mmWave Radar Object Recognition in Contested Environments

  2025-10-06

  article

  Millimeter-wave (mmWave) radars have demonstrated strong performance under challenging visibility conditions and support high-resolution imaging through techniques such as synthetic aperture radar (SAR). The increasing availability of compact, low-cost mmWave systems has made them particularly appealing for low-power platforms in mission-critical tactical scenarios. However, traditional SAR imaging demands extensive sampling, leading to significant delays and excessive data collection—particularly in environments where only a small subset of the scene contains mission-relevant information.To overcome this inefficiency, we introduce mmPrism, a perception-aware attention scheduling framework for mmWave radar that intelligently concentrates sensing and computation on Target Areas of Interest (TAIs). By combining rapid coarse scanning with spotlight-mode SAR under the control of a perception-aware TAI scheduler, mmPrism reduces unnecessary overhead while preserving task-relevant detail. Experimental results demonstrate substantial improvements in recognition throughput and latency, with minimal compromise in imaging resolution.

  PublisherDOI

• D-planner: An Efficient Surrounding-aware Multi-drone System for Urban Monitoring

  2024-10-28 · 1 citations

  article

  In monitoring urban areas with dense infrastructures, drone swarms emerge as an efficient means for locating and collecting data from urban targets. In this task, the drone swarms are required to visit a set of valuable sites. However, many challenges exist in the data collection process of drone swarms: complex streets and obstacles require drones to use computer vision and pathfinding algorithms to perceive the environment, avoid collisions, and manage power in real time with their limited onboard computation and battery. In extreme situations where battery limitations prevent the surveillance of all intended sites, drones are desired to collect information from as many valuable targets as possible. It calls for extra smart path planning. Meanwhile, dynamically changing zone priorities and environments can suddenly change target values or render planned sites dangerous, necessitating immediate path recomputation to ensure efficiencies of urban target surveillance.In this paper, we address the following question: Can we enhance the performance, including safety and efficiency, of drone swarms to conduct urban target monitoring, despite constraints like battery life, limited computational resources, and evolving environments? We introduce our solution, D-planner, a system that performs efficient drone swarm path planning with intelligent navigation algorithms, carries out collision-free navigation using a computer vision and pathfinding module, and integrates incremental dynamic path computation that leverages geometries and greedy strategies to figure out the priority of data sites and plan safe paths with a limited deviation from the original paths. To evaluate the system, we constructed a simulation environment of a large city with meter-level precision using Google Maps. Experiments on our system show that D-planner can improve the planning speed of target-rich paths by up to 6.5× and the total value by up to 25.60% compared with baseline solutions.

  PublisherDOI

• Fine-grained Distributed Data Plane Verification with Intent-based Slicing

  arXiv (Cornell University) · 2024-05-31

  preprintOpen accessSenior author

  Data plane verification has grown into a powerful tool to ensure network correctness. However, existing methods with monolithic models have memory requirements tied to network sizes, and the existing method of scaling out is too limited in expressiveness to capture practical network features. In this paper, we describe Scylla, a general data plane verifier that provides fine-grained scale-out without the need for a monolithic network model. Scylla creates models for what we call intent-based slices, each of which is constructed at the rule-level granularity with only enough to verify a given set of intents. The sliced models are retained and incrementally updated in memory across a distributed compute cluster in response to network updates. Our experiments show that Scylla makes the scaling problem more granular -- tied to the size of the intent-based slices rather than that of the overall network. This enables Scylla to verify large, complex networks in minimum units of work that are significantly smaller (in both memory and time) than past techniques, enabling fast scale-out verification with minimal resource requirement.

  PublisherOA PDFDOI

Recent grants

• FIA: Collaborative Research: NEBULA: A Future Internet That Supports Trustworthy Cloud Computing

  NSF · $581k · 2010–2014

• CAREER: Getting Rid of Bugs: Realizing Interactive Debugging of Networked Systems

  NSF · $472k · 2011–2017

• NeTS-NECO: Collaborative Research: Fixing the Reliability Problem in Network Software From its Root

  NSF · $650k · 2008–2013

Frequent coauthors

• J. van den Brand

  128 shared

• L. Sun

  Bundesministerium für Klimaschutz, Umwelt, Energie, Mobilität, Innovation und Technologie

  88 shared

• A. Heidmann

  84 shared

• N. Arnaud

  Université Paris-Saclay

  81 shared

• I. M. Pinto

  Enrico Fermi Center for Study and Research

  79 shared

• E. Polini

  Laboratoire d’Annecy de Physique des Particules

  76 shared

• M. Seglar-Arroyo

  76 shared

• M. Croquette

  Laboratoire Kastler Brossel

  72 shared

Education

• Ph.D., Computer Science

  University of Illinois at Urbana-Champaign

  2005

• M.S., Computer Science

  University of Illinois at Urbana-Champaign

  2001

• B.S., Computer Science

  University of Illinois at Urbana-Champaign

  1999

Awards & honors

• IEEE Fellow (2023)