Exploiting Spatio-Temporal Correlation in RF Data Using Deep Learning

Abstract

The pervasive presence of wireless services and applications have become an integral part of our lives. We depend on wireless technologies not only for our smartphones but also for other applications like surveillance, navigation, jamming, anti-jamming, radar to name a few areas of applications. These recent advances of wireless technologies in radio frequency (RF) environments have warranted more autonomous deployments of wireless systems. With such large scale dependence on use of the RF spectrum, it becomes imperative to understand the ambient signal characteristics for optimal deployment of wireless infrastructure and efficient resource provisioning. In order to make the best use of such radio resources in both the spatial and time domains, past and current knowledge of the RF signals are important. Although sensing mechanisms can be leveraged to assess the current environment, learning techniques are the typically used for analyzing past observations and to predict the future occurrences of events in a given RF environment. Machine learning (ML) techniques, having already proven useful in various domains, are also being sought for characterizing and understanding the RF environment. Some of the goals of the learning techniques in the RF domain are transmitter or emitter fingerprinting, emitter localization, modulation recognition, feature learning, attention and saliency, autonomous RF sensor configuration and waveform synthesis. Moreover, in large-scale autonomous deployments of wireless communication networks, the signals received from one component play a crucial role in the decision-making process of other components. In order to efficiently implement such systems, each component of the network should be uniquely identifiable. ML techniques, that include recurrent structures, have shown promise in creating such autonomous deployments using the idea of radio frequency machine learning (RFML). Deep learning (DL) techniques with the ability to automatically learn features, can be used for characterization and recognition of different RF properties by automatically exploiting the inherent features in the signal data. In this chapter, we present an application of such deep learning techniques to the task of RF transmitter fingerprinting. The first section concentrates on the application areas in the field of RF where deep learning can be leveraged for futuristic autonomous deployments. Section 2 presents discussion of different deep learning approaches for the task of transmitter fingerprinting as well as the significance of leveraging recurrent structures through the use of recurrent neural network (RNN) models. Once we have established the basic knowledge and motivation, we dive deep into the application of deep learning for transmitter fingerprinting. Hence, a transmitter fingerprinting technique for radio device identification using recurrent structures, by exploiting the spatio-temporal properties of the received radio signal, is discussed in Sects. 3 and 4. We present three types of recurrent neural networks (RNNs) using different types of cell models: (i) long short-term memory (LSTM), (ii) gated recurrent unit (GRU), and (iii) convolutional long short-term memory (ConvLSTM) for that task. The proposed models are also validated with real data and evaluated using a framework implemented using Python. Section 5 describes the testbed setup and experimental design. The experimental results, computational complexity analysis, and comparison with state of the art are discussed in Sect. 6. The last section summarizes the chapter.