Technical Perspective: Physical Layer Resilience through Deep Learning in Software Radios

Resilience is the new holy grail in wireless communication systems. Complex radio environments and malicious attacks using intelligent jamming contribute to unreliable communication systems. Early approaches to deal with such problems were based on frequency hopping, scrambling, chirping, and cognitive radio-based concepts, among others. Physical-layer security was increased using known codes and pseudorandom number sequences. However, these approaches are not up to modern standards; they do not improve resilience and are rather easy to attack by means of intelligent jamming.

Conceptually, dynamic changing waveforms and physical layer parameters would help overcoming many of these issues. However, almost all modern radio technologies are rather inflexible when it comes to changing physical layer parameters on the fly. For example, Bluetooth is limited to frequency hopping, and Wi-Fi to switching channels and modulation/encoding schemes based on active scanning. What is needed is a system that continuously changes physical layer configurations, that is, carrier frequency, FFT size, symbol modulation, and even the position of header and pilots. This way, the overall resilience of the wireless system would be improved significantly.

There are some works that address the sender side, that is, how to enable a sender to vary physical layer parameters, switching even between completely different waveform designs. Off-the-shelf chips are usually not able to do this—for a good reason, they are bound to operate within the limits of the standardized protocols. There is, however, a concept called cross-technology communication (CTC), which acts as an enabler for dynamic changes of waveforms, modulation, and more. For example, it has been shown that normal Wi-Fi chips can be used to emulate Bluetooth, LTE, ZigBee, and many other waveforms by sending carefully chosen Wi-Fi signals. This way, communication between entirely different technologies becomes possible—significantly enhancing the resilience of the communication system.

The following paper tackles the problem from a completely new perspective. The presented PolymoRF concept represents a polymorphic wireless receiver entirely based on physical-layer deep learning. PolymoRF focuses entirely on the receiver system: The receiver can determine on the fly the current parameterization used by the sender. The main idea is to apply a novel deep learning-based model, named RFNet, to infer physical layer parameters from I/Q samples collected by a software radio. The I/Q representation stands for in-phase and quadrature components using real and imaginary parts of complex numbers; and allows to accurately represent phase and amplitude of a signal. Modulated signals, thus, span an image in the plane, which is called a constellation diagram.

Conceptually, the authors go one step beyond classic software-defined radio approaches that use an analog radio frontend in combination with a software frontend (typically a mix of FPGA and CPU-based parts). As all signal processing up to demodulation and decoding is realized in software, new functionality, here the RFNet model, can be deeply integrated with the software radio in the FPGA implementation. This helps bring deep learning much closer to wireless signal processing; they achieve three goals eminent for resilient wireless communications:

A deep learning architecture is created to analyze radio signals without feature extraction and selection. This is achieved by conceptually arranging I/Q samples represented in the complex I/Q plane in the form of an image, which can be approached and analyzed more easily using convolutional neural networks. This is a rather novel approach to I/Q sample processing. A general-purpose hardware/software architecture is designed that is like existing software-defined radio solutions, but unique in the way that pipelining and unrolling techniques have been integrated helping to reduce the latency of the RFNet model in the signal processing chain. A system-level feasibility—most important from a systems perspective and for combining theory and practice—is performed by implementing the proposed concepts in a testbed. For this, the RFNet model has been implemented in software and realized as a system on a chip using FPGA technology. Results demonstrate this can be done without introducing much additional latency. This reality check clearly outlines the advantages of the polymorphic receiver concept.

What is missing and what can we expect in the future from this line of research? The polymorphic wireless receiver concept is without doubt a major step forward, offering a door to next-generation resilient wireless communication systems. This physical layer deep-learning approach on the receiver side now must be combined with adequate transmitter concepts. This can be classic wireless transmitter modules or novel multidimensional scrambler techniques that make jamming more difficult and helps overcoming natural disturbances of the wireless radio signals including interference from other communication protocols. In this research, cross-technology communication may help reusing existing chips, like Wi-Fi chips, to enable such features. We can expect to see implementations combining polymorphic wireless receivers with approaches using learning concepts on the sending side. Eventually, this line of research will improve resilience in wireless communications well beyond our current imagination.