| 摘要: |
| 在基于深度学习的辐射源个体识别研究中，对抗样本会导致性能良好的分类模型出现显著的分类错误，严重影响模型的可靠性。为了应对这一挑战，设计了一种基于梯度的攻击算法，旨在提高对抗样本在白盒和黑盒环境下的攻击成功率。首先，采用随机切片的方法对辐射源个体信号ADS-B（Automatic Dependent Surveillance-Broadcast）进行数据增强，以提高数据的多样性和鲁棒性；接着，针对ADS-B信号的IQ特性，设计了适用于IQ信号的特征提取方法。该方法通过随机抽取特征并将未抽取部分的特征进行了缩放处理，以降低对信号波形的形变；最后，将提取方法与动量迭代结合，提出了特征动量迭代法（Feature Momentum Iterative Fast Gradient Method，FMIM）。实验结果表明，与现有的攻击算法相比，FMIM在白盒环境下的攻击成功率提高了 3%~23.9%，在黑盒环境下提高了 3.6%～7.1%。 |
| 关键词: 辐射源个体识别 深度学习 对抗样本 特征梯度 |
| DOI：10.20079/j.issn.1001-893x.240328002 |
|
| 基金项目:国家自然科学基金资助项目（62371463） |
|
| Adversarial Attacks Based on Gradient Feature for Specific Emitter Identification |
| LIU Shaolong,ZHANG Tao,ZHAO Chen,LIU Fenghui |
| (1.School of Electronic and Information Engineering,Nanjing University of Information Science and Technology,Nanjing 210044,China;2.The 63rd Research Institute of National University of Defense Technology,Nanjing 210007,China) |
| Abstract: |
| In the study of deep learning-based specific emitter identification,adversarial examples can lead to significant classification errors in otherwise high-performing classification models,severely affecting the reliability of these models.To address this challenge,the authors design a gradient-based attack algorithm aimed at improving the attack success rate of adversarial examples in both white-box and black-box environments.First,a random slicing method is applied to enhance the data diversity and robustness of automatic dependent surveillance-broadcast(ADS-B)signals.Next,a feature extraction method suitable for IQ signals is designed,which involves randomly selecting features and scaling the unselected features to reduce signal waveform deformation.Finally,this feature extraction method is integrated with momentum iteration to propose the feature momentum iterative fast gradient method(FMIM).Experimental results show that,compared to existing attack algorithms,FMIM improves the attack success rate by 3%to 23.9%in white-box environments and by 3.6%to 7.1%in black-box environments. |
| Key words: specific emitter identification deep learning adversarial samples feature gradient |
