A challenging problem in Long Range (LoRa) communications enabled Industrial Internet of Things (IIoT) is the detection of rogue devices, which attempt to impersonate real devices by spoofing their authentic identifications in order to steal information and gain access to the system. Although machine learning (ML) offers a promising approach to detecting rogue devices, existing ML models rely on domain knowledge yet exhibit low detection accuracy and vulnerability against adversarial attacks. This paper proposes SmartLens, a novel real-time frequency domain feature based rogue device detection system, using a lightweight statistical ML algorithm and Mahalanobis distance to achieve high accuracy and low latency. We develop a method for extracting fine-grained sequential information from encrypted network traffic using frequency domain analysis that helps limit information loss and achieve high detection accuracy. Additionally, we formulate a constrained optimization problem to decrease the scale of temporal features. The effectiveness of SmartLens is evaluated on a real-world dataset collected using 60 LoRa devices. Our results demonstrate that SmartLens outperforms state-of-the-art systems with improved performance in terms of accuracy, latency and robustness. Specifically, SmartLens achieves over 86.82% detection accuracy under various evasion attacks, and 39.56% less detection latency than that of the baselines.