Signal level detection stops spoofing in vehicular sensing systems
Green Learning-Based Signal-Level Detection of RIS Spoofing Attacks in Vehicular ISAC Systems
Information Theory
Summary
Vehicles and roadside units use wireless signals to sense the environment and communicate, but tricksters can send fake signals to confuse them. The authors present a new green learning method that detects fake reflected signals directly from the raw wireless echoes, rather than needing detailed vehicle paths first. This approach is both very accurate and efficient, working even when the fake signal devices change size. Their method helps keep vehicular sensing safe and trustworthy.
What this means in practice
- •For vehicular communication engineers: Detect spoofing attacks on roadside sensing units using efficient signal-level analysis to secure vehicle environment sensing.
- •For wireless security teams: Improve physical-layer security in integrated communication and sensing systems by identifying malicious signal reflections without trajectory reconstruction.
Authors
Po-Heng Chou, Tingyu Shui, Pin-Qi Fu, Lien-Sheng Chang, Walid Saad, Li-Chun Wang
Abstract
Integrated sensing and communication (ISAC) enables wireless infrastructure to jointly support communication and environmental sensing, but also introduces new physical-layer security threats. A malicious reconfigurable intelligent surface (RIS) can manipulate its time-varying reflection phases to generate spoofing echoes and distort the sensing results of a roadside unit (RSU). Existing trajectory-level detectors rely on reconstructed sensing parameters and vehicle trajectories, making their performance dependent on upstream estimation accuracy. In this paper, we propose a signal-level Green Learning (GL) framework that directly detects RIS spoofing from matched-filtered RSU echoes. The proposed method combines two-stage Subspace Approximation with Adjusted Bias (SAAB), Discriminant Feature Test (DFT), and eXtreme Gradient Boosting (XGBoost) for spatial-temporal representation learning, feature selection, and classification, respectively. Simulation results show 99.335\% detection accuracy for a 32-element malicious RIS at $-80$ dBm receiver noise power. GL remains competitive with Long Short-Term Memory (LSTM) and bidirectional LSTM (BiLSTM) baselines while requiring only 0.1119 million floating-point operations (FLOPs), and also generalizes to previously unseen RIS sizes.