Scholarlink

Signal classification in wireless communications confronts critical security vulnerabilities from adversarial samples and noise interference, posing significant risks to intelligent driving systems where reliable signal recognition underpins collision avoidance and autonomous navigation. In automotive applications, adversarial attacks can manipulate sensor inputs (e.g., LiDAR spoofing causing false obstacle detection), while environmental noise (e.g., heavy rain reducing camera accuracy by 40%) and feature-ambiguous signals (e.g., low-power jamming mimicking V2X communications) compromise perception reliability. Conventional deep learning solutions face dual limitations: excessive energy consumption (8–15 \(\mu \) J/inference) incompatible with automotive edge devices, and vulnerability to gradient-based attacks endangering safety-critical functions like emergency braking. To address this, we develop a novel spiking neural network (SNN) framework with Piecewise Adaptive Surrogate (PAS) gradient methodology. Our biologically inspired dual-branch architecture simultaneously discriminates between adversarial signals (FGSM/PGD attacks), noise-corrupted samples (Gaussian, impulse, narrowband), and feature-ambiguous hybrid signals. Leveraging statistical feature learning and dynamic \(\alpha \) -blending training ( \(\alpha \sim \mathcal {U}(0.35,0.65)\) ), the framework achieves Robust Accuracy (RA) of 89.5–93.1% under adversarial/noise conditions and Hybrid Recall Rate (HRR) of 82.1–88.3% for “gray zone” signals, validated on RadioML2016.10a/2018.01a and automotive datasets. The SNN implementation achieves 5.8 \(\times \) energy efficiency compared to CNNs, enabling real-time processing for safety-critical automotive applications.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Spiking Neural Network-Based Signal Classification on Adversarial Example and Signal with Common Noise

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Ping Ye,
  • Yunzhe Tian,
  • Xiangyu Shi,
  • Xiaoshu Cui,
  • Yaunwan Chen,
  • Yanfeng Gu,
  • Qiong Li,
  • Jiqiang Liu,
  • Wenjia Niu

摘要

Signal classification in wireless communications confronts critical security vulnerabilities from adversarial samples and noise interference, posing significant risks to intelligent driving systems where reliable signal recognition underpins collision avoidance and autonomous navigation. In automotive applications, adversarial attacks can manipulate sensor inputs (e.g., LiDAR spoofing causing false obstacle detection), while environmental noise (e.g., heavy rain reducing camera accuracy by 40%) and feature-ambiguous signals (e.g., low-power jamming mimicking V2X communications) compromise perception reliability. Conventional deep learning solutions face dual limitations: excessive energy consumption (8–15 \(\mu \) J/inference) incompatible with automotive edge devices, and vulnerability to gradient-based attacks endangering safety-critical functions like emergency braking. To address this, we develop a novel spiking neural network (SNN) framework with Piecewise Adaptive Surrogate (PAS) gradient methodology. Our biologically inspired dual-branch architecture simultaneously discriminates between adversarial signals (FGSM/PGD attacks), noise-corrupted samples (Gaussian, impulse, narrowband), and feature-ambiguous hybrid signals. Leveraging statistical feature learning and dynamic \(\alpha \) -blending training ( \(\alpha \sim \mathcal {U}(0.35,0.65)\) ), the framework achieves Robust Accuracy (RA) of 89.5–93.1% under adversarial/noise conditions and Hybrid Recall Rate (HRR) of 82.1–88.3% for “gray zone” signals, validated on RadioML2016.10a/2018.01a and automotive datasets. The SNN implementation achieves 5.8 \(\times \) energy efficiency compared to CNNs, enabling real-time processing for safety-critical automotive applications.