About

Marco Levorato is an Associate Professor in the Computer Science department at UC Irvine. He completed his Ph.D. in Electrical Engineering at the University of Padova, Italy, in 2009. His research interests are focused on distributed computing over unreliable wireless systems, especially for autonomous vehicles and healthcare systems. His work has received recognition including the best paper award at IEEE GLOBECOM in 2012, the UC Hellman Foundation Award in 2016, and the Dean mid-career research award in 2019. His research is funded by prominent organizations such as the National Science Foundation, the Department of Defense, Intel, and Cisco. He has served as vice chair of the IEEE Technical Committee on Smart Grid Communications and is actively involved in the IEEE and ACM conference communities as a TPC member and associate editor. His educational background includes a master's and bachelor's degree in Electrical Engineering from the University of Ferrara, Italy.

Research topics

• Computer Science
• Artificial Intelligence
• Computer network
• Machine Learning
• Distributed computing
• Operating system
• Telecommunications

Selected publications

• Efficient Tensor Compression and Reconstruction in Split DNNs for Edge-Based Object Detection

  IEEE Internet of Things Journal · 2026-01-28

  article

  Computer Vision (CV) tasks are among the most pivotal, yet challenging, operations for Uncrewed Aerial Vehicles (UAVs), especially in mission-critical applications. They require processing complex image data through Deep Neural Networks (DNNs), which demand computational resources far beyond UAVs’ capacity. To address this limitation, Split DNNs offer a promising solution by partitioning the model into: (i) a lightweight <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Head</i>, deployed on the UAV for rapid, albeit less precise, initial image representations, and (ii) a more complex <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Tail</i>, executed at the network edge for refined, higher-accuracy results. However, this solution necessitates transmitting large tensor data from the UAV to the edge server, leading to significant bandwidth consumption. We tackle this challenge by introducing a goal-oriented framework named Compressed Tensor-based DNN Split (CoTeD). Our framework integrates an application- and system-aware optimization model that orchestrates computing and transmission resources in real time. At the UAV, CoTeD dynamically selects relevant tensor information and optimally trades-off between DNN detection quality and bandwidth consumption, guided by application requirements and system operational conditions. At the edge server, CoTeD reconstructs the tensor, enabling efficient inference by the Tail model. This approach effectively balances bandwidth usage with quality of the CV task output. Experimental results, obtained through our hardware-software testbed and using datasets with different sizes and characteristics, show that CoTeD can reduce data transmission over the radio link by up to 90% without noticeable loss in object detection quality and inference latency by up to 70% compared to local DNN deployment onboard the UAV. Also, CoTeD yields an inference request success rate of at least 90%, with an increase of 20%-80% compared to direct DNN splitting, static JPEG compression, and DNN model quantization.

  PublisherDOI

• ADEx: Adaptive Early Exit DNNs for Inference Robustness to Popularity Drift

  Zenodo (CERN European Organization for Nuclear Research) · 2026-04-19

  articleOpen access
  PublisherDOI

• DynamicSound simulator for simulating moving sources and microphone arrays

  arXiv (Cornell University) · 2026-01-21

  preprintOpen access

  Developing algorithms for sound classification, detection, and localization requires large amounts of flexible and realistic audio data, especially when leveraging modern machine learning and beamforming techniques. However, most existing acoustic simulators are tailored for indoor environments and are limited to static sound sources, making them unsuitable for scenarios involving moving sources, moving microphones, or long-distance propagation. This paper presents DynamicSound an open-source acoustic simulation framework for generating multichannel audio from one or more sound sources with the possibility to move them continuously in three-dimensional space and recorded by arbitrarily configured microphone arrays. The proposed model explicitly accounts for finite sound propagation delays, Doppler effects, distance-dependent attenuation, air absorption, and first-order reflections from planar surfaces, yielding temporally consistent spatial audio signals. Unlike conventional mono or stereo simulators, the proposed system synthesizes audio for an arbitrary number of virtual microphones, accurately reproducing inter-microphone time delays, level differences, and spectral coloration induced by the environment. Comparative evaluations with existing open-source tools demonstrate that the generated signals preserve high spatial fidelity across varying source positions and acoustic conditions. By enabling the generation of realistic multichannel audio under controlled and repeatable conditions, the proposed open framework provides a flexible and reproducible tool for the development, training, and evaluation of modern spatial audio and sound-source localization algorithms.

  PublisherDOI

• ADEx: Adaptive Early Exit DNNs for Inference Robustness to Popularity Drift

  Zenodo (CERN European Organization for Nuclear Research) · 2026-04-19

  articleOpen access
  PublisherDOI

• DynamicSound simulator for simulating moving sources and microphone arrays

  ArXiv.org · 2026-01-21

  articleOpen access

  Developing algorithms for sound classification, detection, and localization requires large amounts of flexible and realistic audio data, especially when leveraging modern machine learning and beamforming techniques. However, most existing acoustic simulators are tailored for indoor environments and are limited to static sound sources, making them unsuitable for scenarios involving moving sources, moving microphones, or long-distance propagation. This paper presents DynamicSound an open-source acoustic simulation framework for generating multichannel audio from one or more sound sources with the possibility to move them continuously in three-dimensional space and recorded by arbitrarily configured microphone arrays. The proposed model explicitly accounts for finite sound propagation delays, Doppler effects, distance-dependent attenuation, air absorption, and first-order reflections from planar surfaces, yielding temporally consistent spatial audio signals. Unlike conventional mono or stereo simulators, the proposed system synthesizes audio for an arbitrary number of virtual microphones, accurately reproducing inter-microphone time delays, level differences, and spectral coloration induced by the environment. Comparative evaluations with existing open-source tools demonstrate that the generated signals preserve high spatial fidelity across varying source positions and acoustic conditions. By enabling the generation of realistic multichannel audio under controlled and repeatable conditions, the proposed open framework provides a flexible and reproducible tool for the development, training, and evaluation of modern spatial audio and sound-source localization algorithms.

  PublisherOA PDF

• Resource-Efficient Sensor Fusion at the Edge via System-Wide Dynamic Gated Neural Networks

  IEEE Transactions on Mobile Computing · 2025-07-08

  articleOpen access

  Next-generation mobile systems will support multiple AI-based applications, each leveraging heterogeneous sensors and data sources through deep neural network (DNN) architectures collaboratively executed within the network. In this context, to minimize the cost of the AI inference task subject to requirements on latency, quality, and – crucially – <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">reliability</i> of the inference process, it is vital to optimize (i) the set of sensors/data sources and (ii) the DNN architecture, (iii) the network nodes executing sections of the DNN, and (iv) the resources to use. To achieve these goals, we leverage dynamic gated neural networks with branches, and propose a novel algorithmic strategy called Quantile-constrained Inference (QIC), based upon quantile-Constrained policy optimization. QIC makes joint, high-quality, swift decisions on all the above aspects of the system, with the aim to minimize inference energy cost. We remark that this is the first contribution connecting gated dynamic DNNs with infrastructure-level decision making. We evaluate QIC using a dynamic gated DNN with stems and branches for optimal sensor fusion and inference, trained on the RADIATE dataset offering Radar, LiDAR, and Camera data, and real-world wireless measurements. Our results confirm that QIC closely matches the optimum and outperforms existing approaches in reducing energy consumption (compute, communication, and total) and application requirements failure by over 70%.

  PublisherOA PDFDOI

• SmartDepth: Motion-Aware Depth Prediction with Intelligent Computing for Navigation

  2025-06-09

  articleOpen accessSenior author

  Advancements in robotic platforms emphasize the need for efficient algorithm designs that minimize resource usage for sensing and computing. However, this objective clashes with the complexity of advanced perception tasks that are essential to robotics operations, such as depth perception and navigation. In this paper, we present SmartDepth, a new perception and navigation framework that significantly reduces resource usage compared to traditional approaches. SmartDepth considers a navigation model whose logic requires a chain of depth maps for a sequential number of steps. At its core, SmartDepth leverages highly accurate DNN-predicted depth maps to initialize the chain. To optimize resource efficiency, it then employs Motion-Aware Depth Prediction (MADP)—our novel, low-complexity, geometry-based algorithm—to extrapolate high-quality depth maps for the rest of the chain. The length of a chain is determined by a predefined efficiency level and should be set to ensure that the quality of MADP-predicted depth maps remains useful. In fact, extending the chain beyond this point would result in degradation, rendering the predicted depth maps ineffective. Experimental results show that, compared to a state-of-the-art navigation framework that uses relatively expensive depth extraction as each step, SmartDepth effectively reduces computational costs, in the form of execution time by 19%, average power consumption by 26%, and energy expenditure by 20%, and at a nominal degradation in navigation accuracy of 0.98% and a small increase in path length of 12%. While larger computational savings can potentially be obtained, this comes with a larger degradation in navigation accuracy and path length. Thus, we posit that SmartDepth is an intuitive evolution in perception-based robotics that improves computing efficiency, with a controlled accuracy loss.

  PublisherOA PDFDOI

• AI-Based Classification of Adversarial Attacks vs. Hardware Fault Corruptions in the Split Computing Context

  2025-07-07

  article

  Split Computing has emerged as a promising paradigm for deploying Deep Neural Networks in Edge and Inter-net of Things systems, enabling inference tasks to be distributed between resource-constrained edge devices and cloud servers. This approach is particularly attractive for autonomous systems, where security and reliability may be critical. However, interme-diate feature maps transmitted between devices are vulnerable to corruption, which may result from intentional adversarial attacks or unintentional hardware faults. Distinguishing whether corruption originates from an external adversary or an inherent system fault is crucial for implementing appropriate counter-measures-reinforcing security mechanisms against attacks or improving system reliability to mitigate the effects of hardware-related faults. To the best of our knowledge, this work is the first to propose a machine learning-based classification mechanism capable of differentiating adversarial attacks from hardware defects in Split Computing systems. The proposed approach analyzes the intermediate feature maps transmitted from the edge device to the server, classifying the source of corruption to guide appropriate responses. Experimental results demonstrate that one of the proposed classifiers can distinguish between intentional and unintentional feature map corruptions with an accuracy of 93.91 %.

  PublisherDOI

• A novel middleware for adaptive and efficient split computing for real-time object detection

  Pervasive and Mobile Computing · 2025-02-22 · 3 citations

  articleOpen access

  Real-world applications requiring real-time responsiveness frequently rely on energy-intensive and compute-heavy neural network algorithms. Strategies include deploying distributed and optimized Deep Neural Networks on mobile devices, which can lead to considerable energy consumption and degraded performance, or offloading larger models to edge servers, which requires low-latency wireless channels. Here we present Furcifer, a novel middleware that autonomously adjusts the computing strategy (i.e., local computing, edge computing, or split computing) based on context conditions. Utilizing container-based services and low-complexity predictors that generalize across environments, Furcifer supports supervised compression as a viable alternative to pure local or remote processing in real-time environments. An extensive set of experiments coversdiverse scenarios, including both stable and highly dynamic channel environments with unpredictable changes in connection quality and load. In moderate-varying scenarios, Furcifer demonstrates significant benefits: achieving a 2x reduction in energy consumption, a 30% higher mean Average Precision score compared to local computing, and a three-fold FPS increase over static offloading. In highly dynamic environments with unreliable connectivity and rapid increases in concurrent clients, Furcifer’s predictive capabilities preserves up to 30% energy, achieving a 16% higher accuracy rate, and completing 80% more frame inferences compared to pure local computing and approaches without trend forecasting, respectively. • Adaptive Split Computing: an efficient strategy for real-time vision applications. • A Middleware for automated, dynamic, and resilient flexible computing. • A Low-Complexity Manager for efficient power, higher FPS rate, and enhanced accuracy.

  PublisherDOI

• Distributional RL for Task Offloading, Resource Allocation and Early Exit Selection at the Edge

  2025-11-25

  articleOpen access

  Early Exiting (EE) is an emerging paradigm in deep learning that equips Deep Neural Networks (DNNs) with intermediate classifiers, enabling a trade-off between inference accuracy and latency. In this work, we investigate the integration of EE mechanisms into edge computing architectures, focusing on a representative use case involving task execution in resourceconstrained computing and communications environments for connected and automated vehicles (CAVs). We develop a detailed system model that captures the complex interplay among timevarying system components, including wireless channel coherence and the dynamic availability of computational and communication resources. Building on this model, we formulate a joint optimization problem encompassing task offloading, resource allocation, and early exit selection. We demonstrate how EE enhances system adaptability under stringent constraints, such as limited bandwidth, computing capacity, or delay requirements. To tackle the complexity of the proposed optimization, we adopt a novel solution approach based on the distributional Soft Actor-Critic (SAC) Deep Reinforcement Learning (DRL) algorithm, which quantifies the uncertainty of the learned policy. Simulation results confirm that integrating EE with edge computing significantly improves the trade-off between inference accuracy and latency compared to conventional architectures, with an increase in performance of up to 212%, in terms of completed tasks.

  PublisherOA PDFDOI

Recent grants

• Multi-Scale Analysis and Control of Smart Energy Systems

  NSF · $260k · 2016–2019

• MLWiNS: Ultra-Reliable Collaborative Computing for Autonomous Unmanned Aerial Vehicles

  NSF · $300k · 2020–2023

• S&AS: FND: Cognitive and Reflective Monitoring Systems for Urban Environments

  NSF · $500k · 2018–2021

Frequent coauthors

• Michele Zorzi

  University of Padua

  69 shared

• Urbashi Mitra

  35 shared

• Sabur Baidya

  University of Louisville

  23 shared

• Nikil Dutt

  20 shared

• Davide Callegaro

  University of California, Irvine

  19 shared

• Leonardo Badia

  University of Padua

  18 shared

• Roberto Valentini

  University of L'Aquila

  15 shared

• Francesco Restuccia

  14 shared

Awards & honors

• Best Paper Award at IEEE GLOBECOM (2012)
• UC Hellman Foundation Award (2016)
• Dean mid-career research award (2019)
• Emerging Innovator of Year Award (2024)
• Keynote at IEEE HealthCom (2022)