MACHINE LEARNING-BASED RADIO FREQUENCY FINGERPRINTING FOR CYBERATTACK DETECTION: RECEIVER-AGNOSTIC AND COLLABORATIVE APPROACHES

Abstract

Radio frequency fingerprint identification (RFFI) exploits device-specific hardware im- pairments to provide a physical-layer authentication mechanism that complements con- ventional cryptographic credentials. In Internet of Things (IoT) deployments based on Long Range (LoRa) modulation, multiple distributed gateways naturally observe the same packet, creating an opportunity for collaborative identification. However, gate- ways with low signal quality or inferior hardware can dilute the collaborative decision when fused unconditionally with stronger receivers. This thesis proposes and evaluates a selective collaborative RFFI scheme in which only a well-chosen Top-K subset of receivers participates in each identification decision. Per-receiver convolutional neural network (CNN) classifiers are trained independently on short-time Fourier transform (STFT) log-magnitude spectrograms, and their softmax posterior vectors are fused by mean aggregation over the selected subset. Receivers are ranked by their single-receiver identification accuracy under each operating condition, defined by recording session and signal-to-noise ratio (SNR) regime. The scheme is evaluated on a public LoRa RFFI benchmark comprising 10 transmitters, 13 heterogeneous software-defined radio (SDR) receivers, and 4 recording sessions, across three SNR bands and three fixed SNR points. The proposed selective scheme consistently outperforms a baseline that fuses all available receivers, with accuracy gains ranging from 3 to 32 percentage points across all evaluated conditions. The optimal subset size is condition-dependent: Top-7 is best when receiver quality is gradually distributed, while Top-3 is best when quality is con- centrated in a small number of high-performing receivers. The highest identification ac- curacy reaches 99.32% at high SNR on Day 2 with Top-3 fusion. These results establish that receiver selection is a necessary design component in heterogeneous collaborative RFFI systems, and that collaborative RFFI provides useful physical-layer evidence for detecting device impersonation attacks, positioning it as a complementary mechanism within a defence-in-depth IoT security architecture.

Date of Award	2026
Original language	American English
Awarding Institution	HBKU College of Science and Engineering