Recent advances in wireless technologies have led to several autonomous deployments of such networks. As nodes across distributed networks must co-exist, it is important that all transmitters and receivers are aware of their radio frequency (RF) surroundings so that they can adapt their transmission and reception parameters to best suit their needs. To this end, machine learning techniques have become popular as they can learn, analyze and predict the RF signals and associated parameters that characterize the RF environment. However, in the presence of adversaries, malicious activities such as jamming and spoofing are inevitable, making most machine learning techniques ineffective in such environments. In this paper we propose the Radio Frequency Adversarial Learning (RFAL) framework for building a robust system to identify rogue RF transmitters by designing and implementing a generative adversarial net (GAN). We hope to exploit transmitter specific “signatures” like the in-phase (I) and quadrature (Q) imbalance (i.e., the I/Q imbalance) present in all transmitters for this task, by learning feature representations using a deep neural network that uses the I/Q data from received signals as input. After detection and elimination of the adversarial transmitters RFAL further uses this learned feature embedding as “fingerprints” for categorizing the trusted transmitters. More specifically, we implement a generative model that learns the sample space of the I/Q values of known transmitters and uses the learned representation to generate signals that imitate the transmissions of these transmitters. We program 8 universal software radio peripheral (USRP) software defined radios (SDRs) as trusted transmitters and collect “over-the-air” raw I/Q data from them using a Realtek Software Defined Radio (RTL-SDR), in a laboratory setting. We also implement a discriminator model that discriminates between the trusted transmitters and the counterfeit ones with 99.9% accuracy and is trained in the GAN framework using data from the generator. Finally, after elimination of the adversarial transmitters, the trusted transmitters are classified using a convolutional neural network (CNN), a fully connected deep neural network (DNN) and a recurrent neural network (RNN) to demonstrate building of an end-to-end robust transmitter identification system with RFAL. Experimental results reveal that the CNN, DNN, and RNN are able to correctly distinguish between the 8 trusted transmitters with 81.6%, 94.6% and 97% accuracy respectively. We also show that better “trusted transmission” classification accuracy is achieved for all three types of neural networks when data from two different types of transmitters (different manufacturers) are used rather than when using the same type of transmitter (same manufacturer).

中文翻译：

---

RFAL：用于射频发射器识别和分类的对抗性学习

无线技术的最新进展已导致此类网络的若干自主部署。由于分布式网络中的节点必须共存，因此所有发射器和接收器都必须了解其射频 (RF) 环境，以便它们能够调整其传输和接收参数以最适合其需求。为此，机器学习技术变得流行，因为它们可以学习、分析和预测表征 RF 环境的 RF 信号和相关参数。然而，在对手面前，干扰和欺骗等恶意活​​动是不可避免的，使得大多数机器学习技术在这种环境中无效。在本文中，我们提出了射频对抗学习 (RFAL) 框架，通过设计和实施生成对抗网络 (GAN) 来构建一个强大的系统来识别恶意 RF 发射机。我们希望通过使用深度神经网络学习特征表示，来利用发送器特定的“特征”，例如所有发送器中存在的同相 (I) 和正交 (Q) 不平衡（即 I/Q 不平衡）来完成此任务使用来自接收信号的 I/Q 数据作为输入。在检测和消除对抗性发射器之后，RFAL 进一步使用这个学习到的特征嵌入作为“指纹”来对受信任的发射器进行分类。进一步来说，我们实现了一个生成模型，该模型学习已知发射机的 I/Q 值的样本空间，并使用学习到的表示来生成模拟这些发射机传输的信号。我们将 8 个通用软件无线电外设 (USRP) 软件定义无线电 (SDR) 编程为受信任的发射器，并在实验室中使用 Realtek 软件定义无线电 (RTL-SDR) 从它们那里收集“空中”原始 I/Q 数据环境。我们还实现了一个鉴别器模型，该模型以 99.9% 的准确率区分可信发射器和伪造发射器，并使用来自生成器的数据在 GAN 框架中进行训练。最后，在消除对抗性发射器后，使用卷积神经网络 (CNN) 对可信发射器进行分类，一个完全连接的深度神经网络 (DNN) 和一个循环神经网络 (RNN)，以演示如何使用 RFAL 构建端到端的稳健发射机识别系统。实验结果表明，CNN、DNN 和 RNN 能够分别以 81.6%、94.6% 和 97% 的准确率正确区分 8 个可信发射机。我们还表明，当使用来自两种不同类型的发射器（不同制造商）的数据而不是使用相同类型的发射器（同一制造商）时，所有三种类型的神经网络都能实现更好的“可信传输”分类精度。和 RNN 能够分别以 81.6%、94.6% 和 97% 的准确率正确区分 8 个可信发射机。我们还表明，当使用来自两种不同类型的发射器（不同制造商）的数据而不是使用相同类型的发射器（同一制造商）时，所有三种类型的神经网络都能实现更好的“可信传输”分类精度。和 RNN 能够分别以 81.6%、94.6% 和 97% 的准确率正确区分 8 个可信发射机。我们还表明，当使用来自两种不同类型的发射器（不同制造商）的数据而不是使用相同类型的发射器（同一制造商）时，所有三种类型的神经网络都能实现更好的“可信传输”分类精度。