Artificial intelligence in software-defined SIGINT systems

Overview

The combination of wideband front ends and powerful processors makes software defined radios an
ideal platform for signals intelligence applications. Artificial intelligence and deep learning
techniques can train a system to detect signals faster than hand-coded algorithms. Learn how
DeepSig applies artificial intelligence and deep learning (DL) with COTS SDR.

Artificial intelligence and deep learning with COTS SDR

As wireless protocols grow more complex, spectrum environments become more contested and
electronic warfare (EW) grows more sophisticated. The degree of baseband processing required by
military radios becomes more complex and specialised.

The hostile environment represents an optimisation problem that quickly becomes unmanageable.
Fully optimising RF systems with this level of complexity has never been practical. Designers have
relied on simplified closed-form models that don’t accurately capture real-world effects and have
fallen back on piecemeal optimisation, wherein individual components are optimised, but full end-
to-end optimisation is limited.

In the last few years, there have been significant advances in AI, especially in the DL class of
machine-learning techniques. Where human designers have toiled at great effort to hand-engineer
solutions to difficult problems, DL directly targets large sets of complex, problem-specific data.

SDR hardware architectures

There are typically two types of COTS cognitive radio (CR) systems for defence:

1. Compact, deployed systems in the field that use AI to determine actionable intelligence in
real time. These systems utilise FPGAs and general-purpose processors (GPPs), sometimes
with the addition of a compact graphics-processing unit (GPU) module.

2. Modular, scalable, more compute-intensive systems typically consisting of CRs coupled to

high-end servers with powerful GPUs for offline processing. These systems frequently
require larger-form-factor RF instruments, and increased data volumes often require higher
throughput buses, such as PCIe.

For low size, weight, and power systems (low-SWaP), FPGA hardware-processing efficiency and low-
latency performance, coupled with GPP programmability, makes a lot of sense. While the FPGA may
be more complicated to program, it is key to achieving low SWaP in real-time systems. In this
situation, Universal Software Radio Peripherals (USRPs) from National Instruments and Ettus
Research provide an off-the-shelf platform in a compact form factor. User-programmable FPGAs are
an inherent aspect of USRP devices, and direct integration with either LabVIEW or open-source
software, such as RF Network on Chip (RFNoC), alleviates the challenge of programming the FPGA in
a hardware description language.
For larger compute-intensive systems, it is important to have a hardware architecture that scales
and can heterogeneously utilise best-in-class processors. These architectures typically comprise
FPGAs for baseband processing, GPPs for control, and GPUs for AI processing. GPUs offer a nice
blend of being able to process massive amounts of data while being relatively easy to program. A
downside of GPUs is their long data pipelines, which lead to higher transfer times, although this is
only an issue in systems that require ultra-low latency. Of course, there are ranges of devices within
each category that trade power at the expense of performance, which you should weigh up in design
analysis.

Processor
SWaP Real-time Ease of use
type

Non-optimal latency and

determinism; improved
Typically large and by using a real-time OS Easiest to program with
GPP
power hungry (though that adds LabVIEW, C/C++, or Python
development
complexity)

Typically difficult to program

with hardware description
Generally SWaP-
FPGA Low latency languages; simplified through
friendly
tools in C/C++, OpenCL,
LabVIEW, or RFNoC

Typically large and Easy to program with

GPU Typically higher latency
power hungry TensorFlow or OpenCL/CUDA

Table 1. Processor options for cognitive radio

An example of a larger compute-intensive system is the DARPA Colosseum testbed used in the
Spectrum Collaboration Challenge. This system includes 128 two-channel USRPs (Ettus X310) with
onboard FPGAs, ATCA-3671 blades with multiple FPGAs for data aggregation, and high-end servers
with powerful GPUs for AI processing.
Figure 1. The DARPA Colosseum testbed features 128 Ettus X310 USRPs and NI ATCA-3671 processing
units.

AI and radio

To understand how AI can address RF system design complexity, it’s helpful to have a high-level
knowledge of recent advances that have driven the AI-based system explosion. The term ‘AI’ has
been used for decades, and broadly encompasses problem solving where a machine is making
decisions to find a solution. Machine learning (ML) refers to a type of AI where a machine is trained
with data to solve a specific problem. DL is a class of ML capable of ‘feature learning’, a process
whereby the machine determines what aspects of data to use in decision making, as opposed to a
human designer specifying salient characteristics.
For example, designers historically hand-coded facial recognition algorithms based on years of
feature-recognition technique research. The DL approach combines a data set of images containing
human faces with operator training to point out where the faces are. The machine learns to
recognise what constitutes a face, without a designer defining the algorithm.

Similarly, RF signal classification and spectrum-sensing algorithms can benefit hugely from DL
methods. Whereas previous automatic modulation classification (AMC) and spectrum-monitoring
approaches required labor-intensive efforts to hand-engineer feature extraction (often taking teams
of engineers months to design and deploy), a DL-based system can train for new signal types in
hours.

DL also permits end-to-end learning, whereby a model jointly learns an encoder and decoder for a
complete transmit-and-receive system. Instead of needing to attempt to optimise a system in
piecemeal fashion by individually tuning each component (such as digital-to-analog converters
[DACs], analog-to-digital converters [ADCs], RF converters, wireless channel, and receiver network)
and stitching them together, the model treats the system as an end-to-end function and learns to
optimise the system holistically.

AI in deployed systems

Performing signal detection and classification using a trained deep neural network takes a few
milliseconds. Compared to iterative and algorithmic signal search, detection, and classification using
traditional methodologies, this can represent several orders of magnitude in performance
improvement. These gains also translate to reduced power consumption and computational
requirements, and the trained models typically provide at least twice the sensitivity of existing
approaches.

DeepSig, a US-based startup focused on signal processing and radio systems, has commercialised DL-
based RF sensing technology in its OmniSIG Sensor software product, which is compatible with NI
and Ettus Research USRPs. Using DL’s automated feature learning, the OmniSIG sensor recognises
new signal types after being trained on just a few seconds’ worth of signal capture.

Figure 2. The OmniSIG sensor performs detection and classification of signals within the cellular
band, using a general-purpose SDR.

For learned communications systems, including end-to-end learning that facilitates training directly
over the physical layer, you can use DeepSig’s OmniPHY software to learn communications systems
optimised for difficult channel conditions, hostile spectrum environments, and limited hardware
performance. These include non-line-of-sight communications; anti-jam capabilities; multi-user
systems in contested environments; and hardware-distortion-effect mitigation.
One of the advantages of learned communications systems is easy optimisation for different
missions. While many users care most about throughput and latency, some might prioritise
operational link distance, power consumption, or even signature and probability of detection or
interception. Moreover, with machine learning, the more you know about the operational
environment, the more effective your trained solution can be.

Combining DL-based sensing and active radio waveforms makes possible entirely new classes of
adaptive waveforms and EW capable of coping with today’s contested spectrum environments. For
DL-based system training, processor performance is key, but once trained, the model can readily be
deployed into low-SWaP embedded systems, such as edge sensors and tactical radios.

Why SDR for SIGINT?

SDR’s core elements—RF front ends and processing units—make it ideal for prototyping and
deploying AI-based signals intelligence systems. USRP’s low SWaP makes it well suited to
communications intelligence deployment, for detecting signals at sub-6 GHz frequency.

For higher frequencies and computationally intensive applications, PXI platform instrumentation
extends to mmWave frequencies to address up to Ka band, greater data throughput across x8 PCIe
links, and massive processing with ATCA modules employing multiple Xilinx Virtex-7 FPGAs.

Detecting adverse signals at unknown frequencies—which could include using spread-spectrum or

frequency-hopping techniques—requires wideband receivers. COTS SDRs integrate the latest
wideband ADCs and DACs to address this. Additionally, you can form multichannel systems to either
extend the effective bandwidth by tuning receiver channels to adjacent bands or share local
oscillators to affect phase coherence across channels. This provides not only signal detection and
identification but also direction finding and localisation.

National Instruments and Ettus Research USRPs utilise heterogeneous architectures to process
signals both onboard the SDR and on a host PC. Inline processing onboard the SDR is important in
cognitive systems, where the output signal is generated in response to spectrum sensing or a
received signal. Processing onboard FPGAs can offer multiple benefits, including lower latency
(compared to shuttling the data to and from a host computer) and data reduction across the data
link or bus, by channelising or storing only signals of interest.

Hostile electromagnetic environments require SIGINT systems that can detect unknown signals and
rapidly adapt to emerging threats. Algorithms with DL capability can be trained to recognise new
signals with reduced development time, and, with low SWaP, real-time processing, wideband front
ends, and flexible programming, SDR is the optimal architecture for deploying AI-based SIGINT
systems.