ARO PECASE: Credal Information Theory

Information Exchange Under Epistemic Uncertainty

Information theory provides the mathematical foundation for optimizing communication systems across bandwidth, latency, power consumption and reliability. Since Shannon’s seminal 1948 work, its foundational quantities have been defined relative to a single, precisely known probability distribution. While this framework rigorously characterizes aleatoric uncertainty, which is the irreducible randomness intrinsic to communication, it does not characterize epistemic uncertainty, which is our ignorance about the true source and channel distributions. In practice, epistemic uncertainty arises from partial sensor calibration, incompletely characterized fading and interference, adversarial perturbations and online learning algorithms that couple encoding and decoding to adaptive behavior. Existing approaches reduce this model uncertainty to a single worst-case scalar, discarding the structure of the uncertainty itself and providing no principled way to distinguish irreducible noise from model ignorance.

This project addresses the limitations described above by developing a new mathematical framework for information fusion and exchange under epistemic uncertainty, where entropy, mutual information and channel capacity become intervals that expand and contract with the degree of epistemic uncertainty. Thrust 1 develops a rigorous theory of credal information fusion for heterogeneous, partially dependent, and dynamically evolving sources, including certified conditions distinguishing when fusion genuinely reduces task-relevant epistemic uncertainty and regret guarantees under adversarial sensor contamination and drift. Thrust 2 establishes the information-theoretic limits of communication and compression under credal uncertainty, introducing a robust rate–distortion theory, probing-assisted channel coding with explicit rate-recovery guarantees, and a task-aware coding framework that makes downstream decision risk an explicit design objective. Thrust 3 validates these theories end-to-end on multi-view image benchmarks, real multimodal perception datasets, and experimental wireless testbeds at Northeastern, under explicit bandwidth, latency, and energy constraints.

ARO ECP: Neural-Approximate Beamforming

This project will establish the scientific foundations of a brand-new approach to MIMO communications named Neural-Approximate Beamforming (NAB). A critical issue of current approaches is that they leverages fixed digital signal processing paradigms to implement beamforming. Ultimately, this aspect (i) makes the wireless system prone to a series of critical security vulnerabilities; (ii) does not scale well as a function of bandwidth and number of antennas; and (iii) cannot dynamically achieve the complex trade-off between concealment, performance and resilience needed by tactical wireless networks. To address these issues, our key intuition is that beams do not need to be perfectly formed to deliver sufficient performance. Therefore, from a theoretical perspective, the communication and computational burden of beamforming can be approximated in the time, antenna and frequency dimensions by using dynamic neural networks adapted in real time based on mission requirements. There are key challenges that need to be solved to concretely realize NAB. First, new theoretical bounds must be computed to find how much beamforming approximation can be tolerated in a given propagation scenario and under different size, weight and power constraints without losing in performance and level of concealment. Next, such theoretical findings need to be incorporated in the design of novel dynamic neural networks architectures that will approximate not only the amount of beamforming feedback sent back to the beamformer, but also the digital signal processing itself, to always deliver the best performance at the right time with minimum computational and communication burden. To acquire the necessary real-world data needed to validate our theoretical findings and train/test the developed dynamic neural networks, we will perform extensive experimental evaluation with state-of-the-art wireless facilities at Northeastern University.

CICI: IPAAI: REPAIRT: Securing xApps in Open RANs with Reliable and Principled AI Red-Teaming

The Open Radio Access Network framework (O-RAN) is enabling open marketplaces where extensible radio-control applications (“xApps”) powered by artificial intelligence can be shared and deployed across multi-vendor cellular networks. Although this openness accelerates innovation, it also opens the door to security vulnerability. Indeed, malicious developers could hide software “backdoors” that silently degrade or manipulate the performance of competing networks. Currently, no testbed or certification pipeline verifies the security of these AI-driven applications before they enter the marketplace. The REPAIRT project addresses this pressing need by establishing the scientific and engineering foundations required to detect, neutralize, and ultimately prevent adversarial AI in O-RAN networks.

REPAIRT couples the programmable wireless infrastructure of Saint Louis University and Northeastern University with new algorithms for systematic “red teaming” of AI algorithms in O-RAN networks. The project develops algorithms to flag adversarial AI behavior both before and after deployment, and introduces techniques that remove poisoned data and logic without costly retraining. A heterogeneous testbed supports reproducible experiments at scale and delivers an open framework that lets researchers and industry stress-test their own xApps under realistic traffic and threat conditions. These tools, data sets, and educational activities are enabling trustworthy xApp marketplaces, seeding advances in AI security applied to other dynamic cyber-physical systems, and prepare the future workforce in securing next-generation AI-driven wireless communication systems.

This award reflects NSF’s statutory mission and has been deemed worthy of support through evaluation using the Foundation’s intellectual merit and broader impacts review criteria.

Please see NSF abstract here.

DARPA YFA: RADAR: Resource-Aware Dynamic Adaptation for Resilient DNNs in Low-SWaP Tactical Systems

Autonomous mobile systems with strict size, weight, and power constraints increasingly need on-device AI based on neural networks for real-time perception and situational awareness. On the other hand, existing neural networks are not resilient to intentional and unintentional perturbations. The RADAR project proposes to address this long-standing issue with new techniques that will allow the deployed neural networks to be protected against perturbations throughreal-time fine-grained reconfiguration.

https://www.darpa.mil/work-with-us/communities/academia/young-faculty-award

FuSe: Deep Learning and Signal Processing using Silicon Photonics and Digital CMOS Circuits for Ultra-Wideband Spectrum Perception

The radio frequency (RF) spectrum weaves the very fabric of wireless communications. And it is among the most precious and scarcest of natural resources. Tomorrow’s tech applications such as digital twins, smart vehicles, and augmented reality demand Gigabit-per-second wireless connectivity everywhere all the time. Such demands call for effective mechanisms to guarantee efficient and secure RF spectrum access. Existing methods use simple techniques that can detect users’ presence in the spectrum but cannot sense the ?who, when, and how? of the spectrum being utilized. Nonetheless, emerging artificial intelligence (AI) methods including but not limited to machine learning (ML) techniques are promising for achieving ?RF perception.? A thorny problem in using AI algorithms for RF perception is the inability to process the massive sensed bandwidth of the spectrum. To solve this problem, this project will leverage a hybrid integration approach, where photonic and electronic small chips, or chiplets, will be synergistically combined to facilitate AI/ML-enabled RF perception over the entire RF spectrum. The education component of the project will address the dearth in the US-based semiconductor workforce through a combination of training on photonic and electronic chip design, AI/ML, and wireless technology skills. The FuSe team will mentor women and minorities who are underrepresented, in topics such as semiconductors, chip design, and wireless communication. Outreach to high-school students using AI-based projects will help build a pipeline of students to pursue engineering degrees focusing on semiconductors and computing. A critical educational emphasis is to fast-track training of students on newer FinFET nodes through a complete revamp of analog and digital IC design courses. The PIs will share the developed education and training material amongst the collaborators and make them available online.

To achieve AI-enabled spectrum sensing, this convergent FuSe project will co-integrate a photonic integrated circuit (PIC) with mixed-signal and energy-efficient asynchronous digital chiplets to realize real-time wideband RF perception. The PIC front-end will allow RF spectrum processing and channelization of over 24 GHz of bandwidth. The mixed-signal IC will interface the PIC?s output with digital AI accelerator chiplets. The team will create AI/ML algorithms for modulation recognition, spectrum sensing, and detection of wireless internet-of-things (IoT) devices or specific RF hardware front-ends using fast convolutional neural networks. PIs will employ high-level synthesis (HLS) of speed/power-efficient RF processing cores for real-time AI/ML algorithm implementation. These HLS prototypes will be custom optimized for minimum chip area and power consumption and will achieve low complexity and fast throughput using weight quantization, compressive processing, quantization-aware retraining, signal flow graph pruning, and power/area-optimized digital computing circuits. The team will synthesize the digital cores as asynchronous digital chiplets. Finally, the photonics and electronic chiplets will be taped-out and fabricated using state-of-the-art commercial foundries including the FinFET-based CMOS process, and then packaged for testing and evaluation.

Please see NSF abstract here.

NeTS: Medium: Resilient-by-Design Data-Driven NextG Open Radio Access Networks

Synopsis

Our society increasingly depends on cellular networks, making it critical to assure that the networks are secure against cyber attacks. Next-generation cellular networks are expected to rely on machine learning (ML) algorithms to achieve real-time resource optimization across space, time, frequency and devices. This project studies security threats to those ML algorithms and develops solutions to protect them, focusing on the Open Radio Access Networks (Open RAN) architecture which is rapidly becoming widespread. All project outputs (algorithms, hardware/software designs, and datasets) will be made publicly available through the NSF RFDataFactory website, helping to address the current lack of large-scale datasets for data-driven wireless research. As part of the project, several graduate students will develop unique expertise at the crossroads of ML, security, embedded systems and wireless networks. The project’s key findings will be integrated into new graduate courses in wireless ML security, and will enrich ongoing initiatives at Northeastern University for undergraduate and K-12 students coming from underrepresented minority groups.

Novel optimization frameworks are investigated to model adversarial ML attacks in Open RANs. These findings are used to design ML architecture search algorithms to find ML models for Open RANs that are resilient to attack while still satisfying constraints such as end-to-end latency and energy consumption. The project designs anomaly detection techniques to enhance resilience in dynamic settings, and dynamic defense strategies against real-time dataset poisoning attacks. The proposed techniques are evaluated using one or more of the following testbeds: the Colosseum network emulator, the OpenRANGym framework, and the NSF PAWR POWDER platform.

Personnel

Principal Investigator: Francesco Restuccia
Co-Principal Investigators: Alina Oprea, Tommaso Melodia
Senior Personnel: Salvatore D’Oro
Graduate Research Assistants: Tanzil Hassan, Andrea Lacava, Stefano Maxenti, Harsh Chaudhari, Giorgio Severi, Sazzad Sayyed, Shahriar Rifat

Publications

Meneghello, Francesca, Francesco Restuccia, and Michele Rossi. “WHACK: Adversarial Beamforming in MU-MIMO Through Compressed Feedback Poisoning.” IEEE Transactions on Wireless Communications, 2024.

Puligheddu, Corrado, Varshney, Nancy, Hassan, Tanzil, Ashdown, Jonathan, Restuccia, Francesco and Chiasserini, Carla Fabiana, “OffloaDNN: Shaping DNNs for Scalable Offloading of Computer Vision Tasks at the Edge”, IEEE 44th International Conference on Distributed Computing Systems (ICDCS), pp. 624-634, 2024.

Rifat, Shahriar and Ashdown, Jonathan and Restuccia, Francesco, “DARDA: Domain-Aware Real-Time Dynamic Neural Network Adaptation”, WACV 2025.

Rifat, Shahriar and De Lucia, Michael and Swami, Ananthram and Ashdown, Jonathan and Turck, Kurt and Restuccia, Francesco, “ADA: Adversarial Dynamic Test Time Adaptation in Radio Frequency Machine Learning Systems,” IEEE MILCOM 2024.

Zhang, Milin and Abdi, Mohammad and Rifat, Shahriar and Restuccia, Francesco, “Resilience of Entropy Model in Distributed Neural Networks,” ECCV 2024.

Zhang, Milin and De_Lucia, Michael and Swami, Ananthram and Ashdown, Jonathan and Turck, Kurt and Restuccia, Francesco, “HyperAdv: Dynamic Defense Against Adversarial Radio Frequency Machine Learning Systems,” IEEE MILCOM 2024.

Educational Activities and Outreach

This project is currently training several PhD students. These students are receiving rigorous training in adversarial machine learning and modern wireless technologies.  Moreover, Dr. Restuccia is actively involved in the UPLIFT and Young Scholars Program at Northeastern University for outreach to undergraduate researchers and K-12 students, respectively. In addition to preparing for research careers, these skills are increasingly in demand in major technology companies.

Dr. D’Oro has been invited to give a course on AI algorithms for O-RAN systems to members of the Taipei Computer Association (TCA). The course has a heavy focus on vulnerabilities of AI-based solutions in O-RAN, which includes adversarial learning and attacks.

Dr. Melodia and D’Oro have submitted a tutorial proposal to IEEE MILCOM 2024 to give a tutorial on these topics and to educate government agencies and contractors on security aspects related to O-RAN technologies based on AI. The tutorial has been accepted and will be given in late October 2024. Moreover, Dr. Melodia and D’Oro have given lectures at the 2024 PhD Summer School in Lipary (Italy), a summer school on networking aspects that attracts PhD students worldwide and is focused on developing the next-generation work force and train them on cutting-edge research topics.  

PI Restuccia is currently teaching a course named “Deep Learning and Edge Computing in Wireless Networks”. During the course, students present “AI in wireless” and “wireless for AI” papers, in addition to working on research assignments. As such, Restuccia is integrating the research conducted in this project with his educational efforts by (i) offering research-oriented  and development-oriented projects from the project; (ii) let the students present research papers published by the PI team on the topic covered by the project.

PI Oprea taught a seminar course on machine learning security and privacy in Fall 2023. The project has supported the development of adversarial machine learning modules for the class, which have been made publicly available through the course website. In Fall 2024, PI Oprea is teaching a seminar course on trustworthy generative AI for undergraduate and MS students that will also include modules on poisoning attacks and mitigations.

NSF Abstract: https://www.nsf.gov/awardsearch/showAward?AWD_ID=2312875&HistoricalAwards=false

SHF: Self-Adaptive Interference-Avoiding Wireless Receiver Hardware through Real-Time Learning-Based Automatic Optimization of Power-Efficient Integrated Circuits

The sheer number of devices in the Internet of Things (IoT) is creating an extremely crowded and dynamic spectrum environment. As such, continuous and seamless adaptation of wireless-communication parameters will become essential in the next few years. On the other hand, typical radio-frequency (RF) integrated circuits are statically optimized for a fixed set of parameters and communication standards, which leaves limited room for real-time optimization at the intersection of hardware and software. Indeed, many of today’s devices are tailored for worst-case scenarios associated with a particular communication standard, which leads to limited performance and excessive power consumption. Conversely, real-time optimization allows to quickly reconfigure circuits for energy-efficient operation while achieving system-level performance goals. This challenge will be addressed by devising Radio Real-Time Machine Learning (RadioRTML), a platform that will demonstrate the feasibility of automatic RF integrated-circuit optimization through machine learning (ML) techniques directly implemented with reconfigurable hardware. This research will transform how the optimization of radio frequency systems is done today by demonstrating that real-time ML-based adaptation of RF parameters is able to achieve significant performance improvements. The outcomes will have long-lasting benefits for the design and optimization of low-power RF circuits for adaptive energy-efficient communication. Furthermore, the project will provide unique training for graduate and undergraduate students at the crossroads of machine learning and integrated circuit design.

Automatic machine-learning-based optimization of RF integrated circuits will be investigated through digital control of analog RF front-end circuits. Novel deep reinforcement-learning (DRL) algorithms will be developed to deliver unprecedented flexibility while improving energy efficiency and minimizing interference impacts. A key challenge in the application of DRL is to design a policy network expressive enough to achieve the required performance, yet implementable in a resource-constrained embedded IoT platform. For this reason, new techniques for effective and efficient policy network design will be created. Since DRL is known to exhibit slow convergence times and high energy consumption, this research will include the design of novel transfer-learning techniques to speed up DRL convergence, and it will leverage edge-computing techniques to significantly reduce the energy consumption of the platform. At the RF circuit level, customized topologies and design techniques will be created to construct a flexible receiver front-end. RadioRTML will be prototyped on a System-on-Chip (SoC)-based software-defined radio (SDR) connected to a custom-designed printed circuit board for the RF front-end chip. To thoroughly train and test the RadioRTML algorithms, large-scale data collection will be performed utilizing Arena (a 64-antenna 24-SDR system located at Northeastern University), Colosseum (the world’s largest network emulator), and the NSF POWDER testbed.

This award reflects NSF’s statutory mission and has been deemed worthy of support through evaluation using the Foundation’s intellectual merit and broader impacts review criteria.

NSF Abstract: https://www.nsf.gov/awardsearch/showAward?AWD_ID=2218845

SWIFT: AI-based Sensing for Improved Resiliency via Spectral Adaptation with Lifelong Learning

Resilience to interference via improved spectrum access requires fast sensing, cognition, and actionable intelligence to algorithmically enforce compliance in real-time. The ability to measure spectrum usage, quantify legitimate uses, detect violations and enforce compliance directly leads to improved spectrum utilization, coexistence of multiple competing users, and enhanced security. To this end, this SWIFT project will demonstrate a system for spectral situational awareness through radio frequency (RF) machine learning (ML). The key objective is to obtain actionable spectrum intelligence in the sub-6 GHz legacy bands through a real-time understanding of waveform shapes, spectral content, and modulation schemes. The research will create lifelong incremental learning approaches to spectrum management and dynamic spectrum access, enabled by advanced hardware innovations.

The project is expected to improve at least 100x over software-based systems, through a combination of array processing, reliable AI with lifelong learning algorithms, low-complexity AI, and digital signal processing. Specifically, AI techniques will be used to achieve spectrum intelligence, and more specifically data driven techniques, such as deep learning, towards real-time processing of wideband multi-directional RF signals carrying a diverse set of waveforms, modulations, and protocols. Led by Florida International University (FIU) – South Florida’s largest public research R1 university with 67+% Hispanic students, this SWIFT team will include many under-represented students, who in summer research, will learn key concepts in spectrum sensing. PIs at Embry-Riddle Aeronautical University will spearhead efforts in mentoring women in science, technology, engineering, and mathematics. The PI at Northeastern University will focus on creating graduate teaching materials in wireless communications and RF-ML based on Colosseum (the world’s largest RF emulator) and the PAWR platforms. The team will develop and maintain public open datasets for training AI radios for usability and reproducibility of the scientific community.

This award reflects NSF’s statutory mission and has been deemed worthy of support through evaluation using the Foundation’s intellectual merit and broader impacts review criteria.

NSF Abstract: https://www.nsf.gov/awardsearch/showAward?AWD_ID=2229472

ONR YIP: Polymorphic Wireless Computing for Ultra-Wideband 6G Spectrum Dominance

Achieving technological dominance in 6G wireless networks is of fundamental strategic importance to the economic well-being of the United States. To concretely deploy terabit-per-second communication links, fundamental research breakthroughs are needed to rethink how computing will be conducted in 6G wireless systems. To this end, this project will lay the foundations of a new paradigm named polymorphic wireless computing. We will research novel techniques that will seamlessly adapt not only the underlying algorithmic structure, but also the hardware and software structure of the 6G wireless platform according to ongoing mission-driven objectives, existing network/spectrum operating conditions, and current performance metrics of interest, while being able to operate at several GHz of bandwidth. To this end, we will perform highly interdisciplinary research at the intersection of machine learning, embedded systems, wireless networking and wireless security. To reduce our research to practice, we will prototype our techniques on software-defined radio platforms equipped with reconfigurable hardware; and leverage unique state-of-the-art facilities at Northeastern University to perform extensive experimental evaluation in realistic use-case scenarios.

AFOSR YIP: Dynamic Data Driven Open Radio Access Systems

In this project, we will research the underlying algorithmic foundations and theoretical performance bounds of future dynamic data-driven wireless systems. The key target of this project is to achieve strategic mission-critical application-level objectives by guaranteeing assured wireless communications in congested, contested, concealed, and contaminated environments. First, we mathematically optimize network operations through dynamic allocation of network, computation, and memory resources, taking into account the semantics of the sensed data and the current spectrum conditions into the problem formulation. Importantly, our optimization dynamically compresses sensed data according to application-level semantics, so that network load can be minimized while still guaranteeing key application performance indicators. Next, we investigate novel techniques to guarantee the performance of the proposed data-driven optimization and classification techniques in terms of latency, accuracy, and resource occupation in challenged and constrained scenarios. Our key theoretical findings will be extensively validated through extensive data collection campaigns leveraging several wireless testbeds available at the Institute for the Wireless Internet of Things.

The Air Force Office of Scientific Research, or AFOSR, expands the horizon of scientific knowledge through its leadership and management of the Department of the Air Force’s basic research program. As a vital component of the Air Force Research Laboratory (AFRL), AFOSR’s mission is to discover, shape, champion and transitions high- risk basic research that profoundly impacts the future Air and Space Forces. AFOSR accomplishes its mission through global investment in advanced discovery research efforts in relevant scientific areas. Central to AFOSR’s strategy is the transfer of the fruits of basic research to industry, the supplier of Department of the Air Force acquisitions; to the academic community, which can lead the way to still more accomplishment; and to the other directorates of AFRL that carry the responsibility for applied research leading to acquisition.

AFOSR Young Investigator Award Press Release: https://www.afrl.af.mil/News/Article-Display/Article/3245790/afrlafosr-awards-25-million-via-young-investigator-research-program/