AI-Native FutureG
Networks.

Investigate whether RAN analytics exposed through the O-RAN Y1 interface can be exploited by a compromised consumer to enable targeted and resource-efficient wireless interference.

Developed an end-to-end over-the-air O-RAN security testbed integrating a near-RT RIC, KPM xApp, Y1 Producer/Consumer, srsRAN-based RAN and core, GNU Radio jammer, and USRP B210 radios, with threshold- and DBSCAN-based traffic-aware jamming strategies.

Y1-guided jamming achieved disruption close to always-on jamming while reducing jammer activity by 27%. Under constrained budgets, traffic-aware clustering caused up to an 18.1% bitrate drop while transmitting only 25% of the time.

Protect sensitive radio astronomy observations from harmful LEO satellite interference while preserving broadband coverage for users near radio astronomy sites.

Developed SkyShare, a predictive control-plane system integrating satellite orbital prediction, real-time radio astronomy observation data, ITU-compliant EPFD interference modeling, and SkySched for constellation-wide spot-beam scheduling.

Evaluated using real Starlink Gen2-mini constellation geometries and 25 Ku-band radio astronomy sites worldwide. SkyShare reduces unserved cells by up to 90.68% compared with existing boresight avoidance while maintaining EPFD compliance.

Address the limitations of static neighbor-based LEO topologies by dynamically optimizing inter-satellite links as satellite positions and network conditions evolve over time.

Developed DoTD, a Dynamic Time-Expanded Graph-based topology optimization framework that jointly considers network capacity, latency, and link churn to select high-performance ISLs over time.

Evaluated using Starlink TLE data and the SpaceNet LEO network emulator. Compared with +Grid, DoTD improves network capacity by up to 70.47%, reduces hop count by up to 81.82%, and reduces latency by up to 96.61%.

Integrate IRS, SWIPT, and AI within an open and programmable O-RAN architecture to improve wireless energy efficiency while satisfying network throughput requirements.

Developed O-NIIS, an O-RAN-based framework integrating IRS and SWIPT with a near-RT RIC and an unsupervised DNN (uDNN) xApp. The uDNN jointly optimizes active beamforming at the gNB and passive beamforming at the IRS using network CSI.

The proposed uDNN outperforms DQN, alternating optimization, and fixed beamforming, while achieving energy efficiency close to the global optimum obtained by exhaustive search with substantially lower computational complexity.

Address the scalability limitations of today's isolated ML xApps, where duplicated processing, independent model execution, and xApp queuing increase closed-loop latency and memory consumption as the number of RAN intelligence tasks grows.

Developed xFusion, a unified near-RT RIC architecture based on one xApp and one shared multi-task model for multiple RAN intelligence tasks. The shared execution pipeline eliminates redundant KPM preprocessing, duplicated feature extraction, and repeated model execution across isolated xApps.

At 115 cells, isolated ML xApps reach 2.33 s closed-loop latency and consume 4.99 GB of memory. xFusion reduces closed-loop latency to 205 ms and resident memory to approximately 0.62 GB, while remaining within the near-RT RIC E2 latency budget.

Enable scalable and explainable 5G network intelligence by forecasting multiple KPMs across large-scale cellular networks without site-specific model training, while providing domain-grounded explanations for predicted network degradations.

Developed TelcoAgent, a foundation model-based framework that combines an automatically constructed 3GPP knowledge graph, zero-shot time-series foundation models for multi-KPM forecasting, and a ReAct-based LLM agent for causal reasoning, root-cause analysis, and actionable network recommendations.

Evaluated TelcoAgent on three months of real-world 5G measurements spanning 200 cells and seven KPMs. The framework provides accurate zero-shot forecasting without site-specific retraining and achieves 99.8% numerical fidelity for its generated explanation reports.