Mohammad Aaron Chegini

Topic

Efficient, Secure, and Intelligent Wireless Systems: From Edge AI to Network Management

Department of Computer Science

Date & location

• Wednesday, April 29, 2026

• 2:00 P.M.

• Virtual Defence

Reviewers

Supervisory Committee

• Dr. Amirali Baniasadi, Department of Electrical and Computer Engineering, University of Victoria (Supervisor)

• Mihai Sima, Department of Electrical and Computer Engineering, UVic (Member) 

External Examiner

• Dr. Sean Chester, Department of Computer Science, University of Washington 

Chair of Oral Examination

• Dr. Dennis Hore, Department of Chemistry, UVic

   

Abstract

The vision of AI-native 6G networks is blocked by three barriers. First, state of-the-art deep learning models are too large and power-hungry for battery-powered edge receivers, where modulation schemes can change every few milliseconds and must be classified in real time. Second, the edge hardware that hosts these models is vulnerable to physical denial-of-service attacks that degrade performance without any software breach. Third, when network faults occur, diagnosis remains a slow, manual process because existing AI systems cannot explain their reasoning to engineers. This thesis presents a unified framework that addresses all three barriers.

To solve the efficiency problem, we introduce RFNet, a lightweight neural network for Automatic Modulation Classification that reduces model size by over 90% compared to standard baselines. We validate this on real hardware with Tiny-RFNet, deployed on an NVIDIA Jetson Orin Nano, achieving 8,520 classifications per second at under 3 Watts, fast enough to track adaptive modulation changes in real time.

To address the security problem, we present NoCSNet, a deep learning framework that detects thermal attacks on chip interconnects with 93.8% accuracy, catching malicious patterns that evade conventional threshold-based monitors.

To enable interpretable diagnostics, we introduce TRACE, a cascaded system for 5G root cause analysis. TRACE achieves 99.65% diagnostic accuracy on the TeleLogs benchmark, outperforming even 32-billion-parameter Large Language Models (95.86%), while running in milliseconds on a single CPU and providing transparent reasoning traces that engineers can verify.

Together, these contributions demonstrate that trustworthy AI for next-generation wireless requires co-design across efficiency, security, and interpretability.