Complex Convolutional Bidirectional Mamba With Feature-Wise Modulation for Multi-Device UAV Recognition
Abstract
The rapid proliferation of Uncrewed Aerial Vehicles (UAVs) necessitates reliable monitoring systems for airspace security. While Channel State Information (CSI) offers a precise and flexible modality for UAV sensing, leveraging CSI from multiple distributed devices can significantly enhance recognition accuracy through spatial diversity. However, existing deep learning-based methods often struggle to effectively integrate multi-source data, either neglecting explicit device interactions or suffering from prohibitive computational costs when scaling to larger networks. To address these challenges, we propose the Complex Convolutional FiLM-enhanced Multi-device Bidirectional Mamba Network (CFM-BiMamba). Specifically, we introduce a shared complex-valued feature extractor to reduce fusion redundancy, incorporating device-specific characteristics via Feature-wise Linear Modulation (FiLM). Furthermore, a bidirectional Mamba backbone is employed to capture long-range temporal dependencies efficiently, coupled with a lightweight attention module to robustly fuse multi-view features. Experimental results on both simulated and real-world measured datasets demonstrate that CFM-BiMamba consistently outperforms state-of-the-art baselines, striking a superior balance between recognition accuracy and computational efficiency in multi-device sensing scenarios.