Uncover the Power of Multipath : Detecting NLOS Drones Using Low-Cost WiFi Devices
In recent years, consumer UAV technology has seen considerable advances. Consumer UAVs have become an ideal vector for privacy invasions due to their affordability, size, maneuverability, and their ability to stream live high-quality video. There is considerable proliferation of drones in both civil and military domains. Hence it is critical to detect invading unmanned aerial vehicles (UAVs) or drones in a timely manner for both security and safeguarding privacy. Currently available solutions like active radar, video or acoustic sensors are very expensive (especially for individuals) and have considerable constraints (e.g., requiring visual line of sight).
Recent research on drone detection with passive RF signals provides an opportunity for low-cost deployment of drone detectors on commodity wireless devices. The state of the arts in this direction mainly focus on detecting drones using line-of-sight (LOS) RF signals which are less noisy as compared to their non-LOS (NLOS) counterparts. To the best of our knowledge, there is no existing cost-effective solution for the general public to enable non-LOS(NLOS) detection for drone privacy invasion, which is the most common condition and it still remains an open challenge.
This thesis research provides a low-cost UAV detection system for privacy invasion caused by customer drone. Our model supports NLOS detection with low-cost hardware under $50, and hence it is affordable for the general public to deploy in their house, apartments, and office. Our work utilizes inherent drone motions (i.e., body shifting and vibrations) as unique signatures for drone detection. Firstly, we validated the relationship between drone motions and RF signal under the NLOS condition using extensive experiments. This is motivated by the fact that under NLOS conditions slight changes to the position or motion of a drone could lead to dramatic change in multi-path components in received RF signals. The NLOS condition “amplifies" the RF signatures introduced by drone motions.
We designed a deep learning model to capture the complex features from NLOS RF signals. In particular, we designed and trained a long short-term memory (LSTM) neural network [15, 27], a generative model which can effectively extract features of inputs for NLOS drone detection. Moreover, without knowing the presence of drones, our system starts with classifying any detected RF signals into LOS signals and NLOS signals before the NLOS drone learner is used. Classification of LOS and NLOS signals is feasible because they exhibit different combined features such as strength, variance, and distribution due to their differences in multipath effects. We used the supervised support vector machine (S-SVM)  as the learning model, which is effective for binary classification.
This design is validated via extensive experiments using commodity drones in resident areas with other Wi-Fi enabled mobile devices.