Project Information

• NSF award number: 2312448

• Project period: 10/1/2023 – 9/30/2027

• Principal investigator: Dr. Huacheng Zeng

• Graduate student:

  • Geo Zhou (Ph.D. student)

  • Kunzhe Song (Ph.D. student)

  • Qijun Wang (Ph.D. student)

  • Jie Lu (Ph.D. student)

  • Peihao Yan (Ph.D. student)

  • Shichen Zhang (Ph.D. student)

Project Overview

Recent movement to open up radio access network (RAN) interfaces, led by the O-RAN Alliance, has introduced a new paradigm for future wireless networks. With its key features of openness and intelligence, O-RAN enables a “mix-and-match” approach to RAN development and deployment, allowing telecom carriers to select the best hardware and software from different vendors. Such openness also catalyzes the integration of machine learning (ML) based intelligence into the RAN and promises further performance improvement. This project aims to address several major challenges in O-RAN, with the objective of enhancing its performance, intelligence, and trustworthiness. Through innovation in wireless algorithm and protocol design, ML, and network security, this project expedites the evolution of O-RAN ecosystem. Moreover, the project promotes the participation of women and students with diverse backgrounds in wireless communications and computer science research while enhancing pedagogical activities through new course materials.

This project aims to enhance the performance, intelligence, and trustworthiness of O-RAN by tackling several fundamental challenges across its control loops of three different time scales. The project consists of three interconnected research thrusts. The first thrust focuses on real-time multi-user multi-input and multi-output (MU-MIMO) beamforming in O-RAN's distributed unit (O-DU). It develops a data-driven approach for beamforming that accounts for channel uncertainty. The second thrust focuses on the design of ML algorithms for MU-MIMO control within the near-RT RAN Intelligent Controller (RIC). It establishes an optimization-based framework to generate high-quality labeled datasets for training ML models. The third thrust aims to advance knowledge of the vulnerabilities of ML models in the non-RT RIC of O-RAN and develop safeguard solutions against data manipulation attacks.

The team at Michigan State University will focus on the second research thrust: designing data-driven approaches to enhance the performance and security of O-RANs. This includes, for example, the development of signal processing pipelines for DU and AI-based xAPPs in Near-RT RIC.

Publications

• Efficient User Association and Wireless Scheduling with Shorter Time-Scale Rate Adaptation
  Xiaoyi Wu, Huacheng Zeng, and Bin Li,
  IEEE Transactions on Networking, 2026.

• Integrating Health Sensing into Cellular Networks: Human Sleep Monitoring Using 5G Signals
  Ruxin Lin*, Peihao Yan*, Jie Lu*, Qijun Wang* and Huacheng Zeng
  IEEE INFOCOM WKSHPS: ISAC-FutureG 2026.
  [PDF] [arXiv]

• EExApp: GNN-Based Reinforcement Learning for Radio Unit Energy Optimization in 5G O-RAN
  Jie Lu*, Peihao Yan*, and Huacheng Zeng
  IEEE INFOCOM 2026. [Acceptance rate: 18.9%]
  [PDF] [arXiv] [Source Code]

• MOGUL: A Model-Guided Learning Approach for Scheduling in 5G O-RAN
  Yubo Wu, Huacheng Zeng, Wenjing Lou, and Y. Thomas Hou
  IEEE Internet of Things Journal, 2025.

• xDiff: Online Diffusion Model for Collaborative Inter-Cell Interference Management in 5G O-RAN
  Peihao Yan*, Huacheng Zeng, and Y. Thomas Hou
  IEEE/ACM Transactions on Networking, 2025.
  [PDF] [Source Code]

• Near-Real-Time Resource Slicing for QoS Optimization in 5G O-RAN Using Deep Reinforcement Learning
  Peihao Yan*, Jie Lu*, Huacheng Zeng, and Y. Thomas Hou
  IEEE/ACM Transactions on Networking, 2025.
  [PDF] [Source Code]

• SPP: Achieving Low-Probability-of-Intercept Cellular and Wi-Fi Communications via MIMO-Based Spatial Pilot Perturbation
  Peihao Yan*, Malid Afshari*, and Huacheng Zeng
  IEEE Transaction on Wireless Communications, 2025.
  [PDF]

• Anti-Jamming 5G Millimeter-Wave Communication via Joint Analog and Digital Beamforming: A Bayesian Optimization Approach
  Peihao Yan*, Bowei Zhang*, Shichen Zhang*, Kai Zeng, and Huacheng Zeng
  IEEE Transactions on Machine Learning in Communications and Networking, 2025.
  [PDF]

• Dissertation: Wireless Communication and Sensing System Design: A Learning-based Approach
  Shichen Zhang,
  Michigan State University, June 2025.

• Is Driver on Phone Call? Mobile Device Localization Using Cellular Signal
  Shichen Zhang*, Huacheng Zeng, and Y. Thomas Hou
  IEEE Journal on Selected Areas in Communications, 2024.
  [PDF]

Research Activities and Outcomes

Year 3: Energy-Efficient Radio Control in 5G O-RAN

• Problem description: Radio units (RUs) account for a large share of the energy consumed by a 5G radio access network. Their transmissions are often bursty, creating short opportunities for sleep between active periods. Putting an RU to sleep for longer, however, leaves fewer transmission opportunities for users and can reduce throughput or increase delay. This work addresses how an O-RAN system can jointly schedule RU sleep and allocate DU resources to save energy while meeting the quality-of-service (QoS) needs of multiple traffic slices.

• Technical Challenge: RU sleep decisions are discrete, whereas resource slicing decisions allocate continuous shares of radio resources; both interact under changing traffic, channel conditions, and user populations. An xApp must learn this tradeoff from network measurements while producing policies quickly enough to guide the DU and RU through the near-real-time RIC.

• Our approach and method: We developed EExAPP, a deep reinforcement learning xApp that jointly selects an RU sleep pattern and the DU's resource shares for network slices. Its dual-actor, dual-critic proximal policy optimization design assigns separate learning components to energy efficiency and QoS. A Transformer encoder converts measurements from a changing number of users into a fixed-size state representation, and a bipartite graph attention network coordinates feedback from the two critics. We implemented the xApp on a 5G O-RAN system using the FlexRIC near-real-time RIC, modified the OpenAirInterface DU to support sleep and slicing control, and connected it to a commercial RU and smartphones.

• Experimental results: Over-the-air experiments used one commercial RU and eight smartphones under light, medium, and heavy traffic with two, four, or eight slices. EExAPP's learning stabilized after about 500 steps, compared with about 750 steps for a single-actor, single-critic version. Across the evaluated scenarios, the full model achieved higher combined energy-efficiency and QoS rewards than that version and the Kairos and O-RAN DRL baselines, while keeping QoS violation rates low. Kairos recorded the lowest QoS violation rates overall, but EExAPP achieved the higher overall reward through longer RU sleep periods. Ablation studies showed that both the Transformer encoder and graph attention coordination contributed to performance, particularly under heavier traffic and more slices.

Year 3: Learning-Based Inter-Cell Interference Management in 5G O-RAN

• Problem description: Neighboring 5G small cells often reuse the same spectrum to increase capacity, but simultaneous transmissions can interfere with one another. This interference can cause unstable throughput, long queues, and even dropped connections. In O-RAN, each DU must schedule users at a fast time scale, while the near-real-time RIC sees measurements from multiple cells and can coordinate them. This work addresses how to use that shared view to improve user throughput and delay as interference, channel conditions, and traffic demands change.

• Technical Challenge: The near-real-time RIC cannot assign every resource block at the DU's millisecond scheduling rate, and DUs cannot coordinate every scheduling decision directly because of communication delay. A useful policy must therefore be compact, quick to update from streaming measurements, and flexible enough for each DU to respond to its local conditions.

• Our approach and method: We developed xDiff, an online reinforcement learning xApp that collects key performance indicators and MAC-layer measurements from multiple DUs through the E2 interface. A conditional diffusion model generates continuous preference values for user and resource-block combinations. These values guide each DU's local scheduler toward allocations that reduce inter-cell interference without fixing every real-time decision in the RIC. Double Q-learning networks evaluate candidate policies, and training combines denoising and reward objectives so the xApp can adapt to changing network conditions. We implemented xDiff in an OpenAirInterface-based 5G O-RAN testbed.

• Experimental results: The testbed included three cells and ten smartphones. In a three-cell case study, independent scheduling led to severe throughput and delay fluctuations and two user disconnections; with xDiff, the users met their throughput and delay targets most of the time and no disconnections were observed. Five-hour comparisons under varying traffic showed that xDiff met throughput demands and achieved the lowest queueing delay among the evaluated interference-management methods in the lab setting. It also gave the lowest delay in a building-scale setting with lighter interference, although its throughput gain there was modest. Ablation experiments found faster adaptation and higher reward than policy generators without the diffusion model; the measured xDiff policy inference time was 21.8 ms.

Year 2 - Secure 5G Cellular Communications against Eavesdropping

• Problem Description This project tackles the challenge of enabling low-probability-of-intercept (LPI) communications in real-world 5G and WiFi systems. Existing physical-layer LPI techniques often rely on custom hardware or are incompatible with billions of deployed user devices, limiting their practicality. The key vulnerability arises because eavesdroppers exploit the same pilot signals as legitimate receivers to estimate the channel and decode data. When pilots and data symbols experience identical effective channels, both legitimate users and eavesdroppers can successfully demodulate transmissions. The central problem, therefore, is how to prevent eavesdroppers from decoding data—without modifying user devices or requiring knowledge of eavesdropper channels.

• Technical Challenges Several challenges complicate this objective. First, any new technique must be fully compatible with existing WiFi and 4G/5G devices, meaning no changes to user hardware or software can be assumed. Second, the system must ensure that intended receivers continue to observe aligned pilot and data channels, while forcing eavesdroppers to observe misaligned ones. This requires careful precoder design under practical constraints such as limited transmit power and a finite number of antennas. Finally, the solution must be effective against both out-of-network eavesdroppers (with no channel knowledge available) and in-network eavesdroppers (with partial channel knowledge), while balancing the tradeoff between security and communication throughput.

• Approach / Methodology We propose Spatial Pilot Perturbation (SPP), a novel MIMO-based precoding technique. The core idea is to apply different precoders to pilot and data symbols within the same transmission frame. For intended receivers, the precoders are designed so that pilots and data experience identical compound channels, enabling correct decoding. For eavesdroppers, however, the pilot and data experience mismatched compound channels, disrupting channel estimation and preventing demodulation. SPP introduces two key design elements: a transmission vector, optimized as if no eavesdroppers exist, and a perturbation vector, chosen from the nullspace of the legitimate user’s channel to distort eavesdropper reception. The approach generalizes to multi-user MIMO, single versus multiple eavesdroppers, and scenarios with or without eavesdropper CSI. Importantly, SPP requires no modifications to user devices and leverages the standard pilot structures already present in 5G and WiFi systems.

• Main Results We validated SPP through both numerical simulations and real-world experiments on 5G and WiFi software-defined radio testbeds. Results demonstrate that, even without any knowledge of eavesdroppers, SPP reduces eavesdropping rates to ≤0.2% in 5G and ≤0.9% in WiFi, with only a 10–18% reduction in throughput. Further experiments show that when limited eavesdropper CSI is available, the error vector magnitude (EVM) gap between users and eavesdroppers increases by several dB, providing additional security. Overall, SPP is shown to be highly effective, backward-compatible, and practical for deployment in existing wireless systems, marking an important step toward real-world adoption of physical-layer LPI techniques.

Year 1 - xAPP design and implementation for near-real-time resource slicing in 5G O-RAN

• Overview. Open-Radio Access Network (O-RAN) has become an important paradigm for 5G and beyond radio access networks. This paper presents an xApp called xSlice for the Near-Real-Time (Near-RT) RAN Intelligent Controller (RIC) of 5G O-RANs. xSlice is an online learning algorithm that adaptively adjusts MAC-layer resource allocation in response to dynamic network states, including time-varying wireless channel conditions, user mobility, traffic fluctuations, and changes in user demand. To address those network dynamics, we first formulate the Quality-of-Service (QoS) optimization problem as a regret minimization problem by quantifying the QoS demands of all traffic sessions through weighting their throughput, latency, and reliability. We then develop a deep reinforcement learning (DRL) framework that utilizes an actor-critic model to combine the advantages of both value-based and policy-based updating methods. A graph convolutional network (GCN) is incorporated as a component of the DRL framework for graph embedding of RAN data, enabling xSlice to handle a dynamic number of traffic sessions. We have implemented xSlice on an O-RAN testbed with 10 smartphones and conducted extensive experiments to evaluate its performance in realistic scenarios. Experimental results show that xSlice can reduce performance regret by 67% compared to state-of-the-art solutions. The below figure shows the high-level architecture of our proposed xSlice system in an O-RAN.

• Testbed Hardware. The below figure shows our O-RAN testbed, which includes 5G Core, O-CU, O-DU, O-RU and 10 COTS smart- phones. We implement the O-RAN system using 4 servers for O-DU, O-CU, Near-RT RIC, and 5G Core. Specific parameter settings are detailed in Table 3. The hardware configuration of this testbed is designed to support Internet access for COTS smartphones. Local network connection is provided by a Net- gear GS308v3 Ethernet switch, which interconnects the 5G core, O-CU, O-DU, and O-RU components. The O-RU is set up on the USRP N310 device, supporting 4x4 MIMO config- urations. USRP N310 uses external clock GPSDO CDA-2990 to improve its clock accuracy. The radio units are tuned to a center frequency of 3319.68 MHz within the TDD n78 band, with a subcarrier spacing of 30 kHz. The testbed also includes 10 smartphones from OnePlus, Google Pixel, Motorola, and Samsung Galaxy, enabling a diverse range of testing scenarios and performance evaluation. The n78 band was used for this project under an FCC Experimental License with Call Sign

• 5G Core Support. We have extended 5G core OAICN for this project. The AMF manages user access, authenti- cation, and mobility, while the UPF handles user data traf- fic routing and quality of service. We have added various Slice Service Types (SST), which identify the slice type (e.g., eMBB, URLLC), and Slice Differentiators (SD) to the UPF database. When the NSSAI of a slice matches the correspond- ing SST and SD code in oai-cn5gdatabaseoai_db.sql, the session is assigned to that slice. OAI Modifications. We use OpenAirInterface (OAI) 5G RAN for our experiments. However, OAI does not support sliceing. Therefore, we rewrote the downlink schduler functions in oaiopenair2LAYER2 NR_MAC_gNB and slicing control E2 interface in oaiopenair2 E2APRAN_FUNCTION. We define two attributes SD and SST, for different slices, which serve as unique identifiers for the slices. When a UE connects to the 5G core network, the core database assigns each UE with an SST and SD, as well as 5G QoS Indicator (5QI) and priority level. The DU then assigns a Radio Network Temporary Identifier (RNTI) to each new UE to distinguish it. Additionally, in the Single Slice model, all sessions utilize a single BWP, where the BWP size corre- sponds to the total number of PRBs. For xSlice, we need to define a BWP for each slice, including its starting position and bandwidth size. The size can be obtained through xSlice, while the starting position is determined by traversing through all of the PRBs. xSlice Implementation. We use Mosaic5g Flexric as our near-RT RIC. Flexric contains E2 Node agent, near-RT RIC and xApp. It provide a flatbuffers encodingdecoding scheme as alternative to ASN.1. We use SWIG as an interface generator to enable CC and Python for the xApps. We build our own xApp with E2AP v2.03 and KPM v2.03. We implement xSlice as an xApp within the Near-RT RIC. The Actor-Critic algorithm of the DRL agent is integrated into the PPO framework, supporting online training. The DRL environment is developed in Python for compatibility. xSlice includes both the DRL model and the E2 interfaces for data collection and processing. We have divided the task into two xApps: Monitor xApp and AI-model xApp. The Monitor xApp is responsible for interacting with the RAN to retrieve information (KPM and MAC data) and for sending updated slicing policies to the RAN. The AI-model xApp primarily focuses on model training and online learning. By employing this multiprocess paradigm, xSlice can promptly apply updated policies to the RAN while ensuring that the states acquired accurately reflect the implementation of the new policy actions.

• Performance Evaluation. The below figure presents the experimental results in terms of their average values. Numerically, xSlice demonstrates superior performance in terms of regret, throughput, and latency compared to existing policies. Notably, when compared to the state-of-the-art solution Zipper, xSlice reduces performance regret by 67%.

Broader Impacts

• Broadening Access to Intelligent 6G O-RAN Research through Community Engagement and Public Dissemination
  On August 6, 2026, Dr. Huacheng Zeng delivered a research talk to the ARA team at Iowa State University as part of the broader impacts activities of this project. The presentation shared research toward intelligent, application-aware, and perception-enabled 6G Open Radio Access Networks (O-RAN). It covered three complementary directions: deep reinforcement learning and graph-based network representations for automated network control; intelligent handover and mobility management to support reliable, low-latency edge AI applications; and integrated sensing and communication to enable environmental perception. These topics connected advances in wireless networking and AI to the capabilities and potential societal benefits of future wireless infrastructure.
  
  The activity supported knowledge exchange across institutions and exposed the ARA team to methods relevant to programmable wireless systems and application-driven network design. By connecting network optimization, distributed AI, and wireless sensing, the presentation provided an interdisciplinary perspective useful for research and education at the intersection of communications, computing, and machine learning. These directions also have potential relevance to ARA’s mission in rural wireless connectivity and applications, including precision agriculture and community services, offering opportunities to explore how emerging O-RAN capabilities could address practical needs.
  
  To extend access beyond the live audience, the talk recording was made publicly available on YouTube. This dissemination provides an educational resource for students, researchers, practitioners, and other interested viewers, allowing them to engage with the material regardless of location or availability during the event. By combining engagement with an external research team and public access to the presentation, the activity broadened the availability of project-related knowledge and supported continued learning about next-generation wireless technologies.

• International Research Exchange on Integrated Sensing and Communication in O-RAN
  On May 22, 2026, Dr. Huacheng Zeng presented “OFDM-Based Monostatic ISAC in O-RAN” at the NSF–NICT Joint Workshop in Japan, supporting the broader impacts of this project through international research dissemination and knowledge exchange. The presentation connected the project’s focus on intelligent, programmable O-RAN infrastructure with emerging integrated sensing and communication (ISAC) capabilities. Examples included three-dimensional scene reconstruction, human pose and facial expression estimation, and sensing using 5G phone signals. These examples illustrated how wireless infrastructure could support services beyond connectivity, with potential applications in vital sign monitoring, elderly fall detection, localization, and activity recognition. Sharing these research directions with the workshop audience broadened awareness of the potential societal benefits of combining wireless communications, sensing, and artificial intelligence.
  
  The presentation also supported U.S.–Japan research exchange by identifying open challenges and opportunities for future collaboration. These included developing shared datasets and benchmarks, improving AI model generalization in unfamiliar environments, combining radio sensing with camera observations, and assessing the reliability, cost, and power requirements of different sensing architectures. The discussion of security and privacy emphasized considerations relevant to the responsible development of wireless sensing, particularly in applications involving people and their daily activities. By presenting both emerging capabilities and unresolved challenges, the talk provided a basis for collaborative investigation and helped connect the project’s O-RAN research to a broader international community working toward useful, reliable, and socially beneficial next-generation wireless systems.

• Zeng hosts undergraduate students to study integrated sensing and communication (ISAC) solutions
  Dr. Zeng’s INSS laboratory is hosting two undergraduate students from the Department of Computer Science and Engineering for an innovative research project focused on ISAC. This outreach activity aims to bridge the gap between theoretical knowledge and practical applications, fostering a collaborative environment where students can engage in cutting-edge research. The primary objective of this ISAC project is to enhance the capabilities of a single WiFi device by integrating advanced sensing functionalities. The students embarked on a comprehensive study to explore various solutions that enable concurrent transmission and reception of radio signals. This dual functionality is crucial for developing sophisticated applications that can leverage existing wireless infrastructure while adding significant value through new sensing capabilities. The students conducted extensive literature reviews and feasibility analyses. They evaluated current methodologies and technologies that allow for simultaneous signal processing, focusing on the challenges and opportunities within the realm of WiFi devices. Building on their findings, the students then transitioned to the development of deep learning models tailored for novel applications. Among the applications they explored are human body skeleton estimation and mask segmentation. These applications not only demonstrate the practical implications of their research but also address real-world problems, such as improving human-computer interaction and enhancing security measures in various settings. As part of this outreach activity, the students participated in regular presentations and discussions with their peers, faculty, and industry professionals. This interaction is designed to promote knowledge sharing and provide feedback that refine their research direction. Additionally, the students obtained the opportunity to showcase their work at departmental seminars, encouraging other students to engage in research and explore the possibilities within the ISAC field.

• Zeng gave talk on 5G signal implementation on SDR to George Mason University students
  Dr. Zeng was invited to deliver a guest lecture on 5G signal implementation to students in the Department of Electrical and Computer Engineering at George Mason University. The lecture aimed to bridge theoretical foundations with practical system design, providing students with both a high-level perspective of 5G networks and a detailed technical understanding of their operation. The talk began with an introduction to the 5G signal frame structure, where Dr. Zeng explained the principles of numerology, subcarrier spacing, and time-slot organization, emphasizing how these design choices enable high data rates, low latency, and flexible spectrum usage. He then discussed the MAC-layer protocols, illustrating how scheduling, HARQ mechanisms, and link adaptation contribute to efficient radio resource management. At the transport layer, he described the challenges of ensuring reliability, low latency, and scalability, particularly in high-mobility scenarios. Building on this foundation, Dr. Zeng integrated results from this project to highlight the evolution of network architectures. He introduced the Open RAN (O-RAN) paradigm, stressing its role in decoupling hardware and software, enabling multi-vendor interoperability, and fostering innovation through software-defined and virtualized components. He further explored how artificial intelligence and machine learning are becoming integral to modern network design, offering capabilities such as adaptive resource management, self-optimization, and anomaly detection. The lecture concluded with a discussion on future research challenges and opportunities. The talk provided students not only with a technical dive into 5G but also with a forward-looking perspective on how emerging technologies will shape the next generation of wireless systems.

• Zeng hosts undergraduate student to work on experimental research on O-RAN
  Zeng's INSS Laboratory provides undergraduate the student with unique opportunities to participate in experimental research on Open Radio Access Networks (O-RAN). The lab emphasizes hands-on learning, enabling the student to engage directly with both the software and hardware that form the backbone of modern wireless communication systems. By immersing themselves in practical experimentation, the student develop not only technical expertise but also the ability to tackle complex systems-level challenges. The student studies open-source RAN software platforms, including srsRAN and the OpenAirInterface (OAI) fronthaul interface (FHI) 7.2 split. These platforms provide a flexible and transparent environment for exploring 5G networks, making them ideal for academic research. The student work on implementing a complete O-RAN system. This involves integrating Benetel O-RUs (O-RAN Radio Units) with srsRAN-based gNB components. The integration process gives the student direct experience with the O-RAN Alliance's principles of disaggregation and openness, where hardware and software from different vendors can be combined to create a fully functioning system. This stage of the project highlights the real-world engineering challenges of interoperability across heterogeneous components. In addition to system integration and debugging, the student conduct performance evaluations of the deployed O-RAN setup. He measured and analyzed system behavior using commercial smartphones connected through the Benetel RU and srsRAN core components. Performance metrics such as throughput, latency, and stability are assessed under different configurations, allowing the student to understand how design choices in open-source RAN implementations affect end-user experience. These experiments not only test the limits of the platforms but also contribute to ongoing efforts to make open RAN technologies viable for real-world applications. By working with real hardware and open-source software, the student develop a holistic view of the opportunities and challenges in building O-RAN systems.