Secure Industrial IoT Systems via RF Fingerprinting Under Impaired Channels With Interference and Noise

Industrial IoT-enabled critical infrastructures are susceptible to cyber attacks due to their mission-critical deployment. To ensure security by design, radio frequency (RF)-based security is considered an effective way for wirelessly monitored or actuated critical infrastructures. For this purpose, this paper presents a novel augmentation-driven deep learning approach to analyze unique transmitter fingerprints and determine the legitimacy of a user device or transmitter. An RF fingerprinting model is susceptible to various channel and environmental conditions that impact the learning performance of a machine/deep learning model. As data gathering cannot always be considered a feasible alternative, efficient solutions that can tackle the impact of varying propagation channels on learning performance are emergent. This work aims to shed light on the RF fingerprinting problem from a different angle when 4G, 5G, and WiFi data samples are collected from different transmitters by proposing a fine-grained augmentation approach to improve the learning performance of a deep learning model. This work also proposes an enhanced classifier structure following the fine-grained augmentation approach. Results of experiments, conducted on the POWDER dataset, demonstrate promising RF fingerprinting performance when training data are augmented in a waveform-specific fine-grained manner. Thus, the RF identification accuracy can be boosted to 97.84% on unseen RF data instances from our previously published work where we had achieved an accuracy of 87.94% using tapped delay line (TDL)/clustered delay line (CDL)-based augmentation approach. The paper also presents a sensitivity analysis of the fine-grained approach concerning different signal-to-noise-ratio (SNR), signal-to-interference-ratio (SIR) levels (20 dB and 30 dB), and signal-to-interference-plus-noise-ratio (SINR) levels (15 dB, 25 dB). The sensitivity analysis exhibits that it achieves 85.78% accuracy at 20 dB SIR on both Day 1 (train) and Day 2 (test) data. In addition, it achieves 92.37% accuracy even at 20 dB SNR on Day 2 data from POWDER dataset. Furthermore, it achieves 84.95% accuracy at 15 dB SINR on Day 2 data. Hence, these results exhibit the resiliency of the fine-grained augmentation approach against interference and noise.