PHYSICAL LAYER SECURITY IN POWER LINE COMMUNICATIONS VIA DEEP LEARNING

Abstract

PLC technology is an emerging technology which transmit data over power lines. This enables real-time monitoring and control but introduces security vulnerabilities to the CPS. Cryptographic security fails when keys are compromised or when devices lack computational resources. This dissertation addresses physical-layer security in \ac{PLC} systems through three contributions: device fingerprinting for authentication, analysis of radio jamming effects, and anomaly detection methods. The first contribution develops device fingerprinting for PLC authentication. The method converts IQ samples to images using bi-variate histograms and applies CNN for classification. Experiments with 8 USRP X310 SDR and 2 power line couplers achieve >90% identification accuracy. The key finding is dual fingerprinting: device signatures contain both radio hardware characteristics and coupler characteristics. Models trained on one coupler fail when tested on another (random-guess accuracy). ResNet-18 achieves 0.0039-second inference time. The second contribution examines radio jamming in PLC systems. Experiments with 9 SDR across 126 configurations show that jamming might have a dual purpose: privacy protection and security threat. At 10 dB attenuation and \acf{RJP} = 0.2, fingerprinting accuracy drops to 14.29\% while BER remains 99\%$ detection accuracy at 50-76 meters from the jammer where BER $\approx$ 0. Supervised methods detect lower jamming power (RJP = 0.1) but require jammed training samples. Unsupervised methods need higher jamming power (RJP = 0.4) but work without adversarial training data. All experiments use real hardware transmitting over power lines in office environments, taking into account channel noise, attenuation, and hardware variations. Device fingerprinting operates independently of cryptographic keys. Anomaly detection works at distances where traditional metrics fail, providing early warning before communication degrades.

Date of Award	2025
Original language	American English
Awarding Institution	HBKU College of Science and Engineering