Transformer-Based Incomplete Time Series Classification Framework for Fast UAV Detection Over Encrypted Wi-Fi Traffic

Abstract

As the production of consumer unmanned aerial vehicles (UAV) with video streaming capabilities increases every year, public security and personal privacy are increasingly put at risk.Thus, efficient and early detection of UAVs remains a pressing issue. To detect UAVs as early as possible, we formulate the task of UAV detection as a univariate time series classification problem, eliminating the time-consuming and computationally expensive process of calculating statistical features frequently used in neural networks for UAV detection. Initially used in natural language processing and recently applied to time series tasks, Transformer neural network models are able to determine relationships within sequential data in order to make classification decisions or predictions about future data points. Since many consumer UAVs exhibit a pattern in packet length during wireless WiFi communication between the drone and controller, we propose using Transformer models to analyze packet length over time. In addition, our research addresses the problem of packet loss over long distances. Since Transformers can generate or predict information, we use a Transformer to impute missing packet lengths, greatly increasing one UAV detection radius when there is packet loss. Overall, our Transformer model uses existing univariate time series data to impute missing packet lengths before classifying the traffic flow, allowing the model to detect UAVs quickly, efficiently, and within a larger radius.