The field of AI has experienced a renaissance fueled by modern data-driven Machine Learning (ML) techniques. In contrast to the expert-engineered processing and decision making of the previous generation of AI, ML-based techniques are able to learn to perform a task from raw sensory input. ML techniques have advanced the fields of image and speech recognition, and self-driving cars, enabled by sufficient training on digitized writing, spoken words, images, video streams, and other digital content.
AI’s first and ongoing wave consists of expert systems that rigidly codify human expertise and decision-making in predictable, rule-driven domains, such as simple game playing, tax preparation, and industrial process control. Such expert systems also have been deployed in RF contexts where, for example, engineers have been able to specify in computer code the rigid rules used by radios to switch to unused frequencies when they encounter interference. While effective, these systems have little understanding of what’s actually happening in the spectrum. RF applications of the second and emerging machine-learning wave of AI should yield far more agile and versatile capabilities: an RFML system, with a sufficiently rich training set of RF data, should be able to identify an enormous range of both known and previously unseen RF waveforms.
In contrast, the use of AI in the RF domain however has not kept pace with these advancements in ML, and little work has been conducted to explore the intersection of traditional RF signal processing and machine learning.(DARPA) is seeking ways to exploit autonomy and artificial intelligence (AI) to improve spectrum management and address the growing appetite for radio frequency (RF) communication in the military and civil domains.
At a time when adversaries have built capabilities to disrupt the RF spectrum, it has become critical to explore how machine learning could be applied to traditional RF signal processing. Through the explosive growth of RF devices and the Internet of Things, the number of connected devices such as phones, sensors, and drones makes it even more important to be able to identify signals intended to hack, spoof, or disrupt RF spectrum usage.
The goal of the Radio Frequency Machine Learning Systems (RFMLS) program is to develop the foundations for applying modern ML to the RF Spectrum domain and develop practical applications in emerging spectrum problems, which demand vastly improved discrimination performance over today’s hand-engineered first generation Cognitive RF systems. Ultimately, these innovations will result in a new generation of RF systems that are goal-driven and can learn from data, allowing experts to focus on the overall system capability and not individual subsystem performance.
One of the applications envisaged by DARPA is IOT security, ML algorithms shall be able to identify the real waveforms with spurious waveforms. As billions of phones, appliances, drones, traffic lights, security systems, environmental sensors, and other radio-connected devices sum into a rapidly growing Internet of Things (IoT), there now is a need to apply ML to the invisible realm of radio frequency (RF) signals, according to program manager Paul Tilghman of DARPA’s Microsystems Technology Office.
“What I am imagining is the ability of an RF Machine Learning system to see and understand the composition of the radio frequency spectrum – the kinds of signals occupying it, differentiating those that are ‘important’ from the background, and identifying those that don’t follow the rules,” said Tilghman. He would want that same system to be able to discern subtle but inevitable differences in the RF signals from what otherwise are identical, mass-manufactured IoT devices and to distinguish these from signals intended to spoof or hack into these devices. “We want to be able to understand and trust what is happening in the Internet of Things and to stand up an RF forensics capability to identify unique and peculiar signals amongst the proverbial cocktail party of signals out there,” said Tilghman.
The same situational awareness regarding the ever-changing composition of RF signals in any given space should also support a wireless communications management paradigm known as spectrum sharing. That’s a paradigm of shared spectrum use rather than the current practice of exclusive allocations governed by license agreements for specific frequencies. Tilghman is hoping to develop technologies to understand the current state of the spectrum for improved and extensive spectrum sharing—which can greatly expand the wireless communications capacity of the electromagnetic spectrum—both in the RFMLS program as well as in another major DARPA effort known as the Spectrum Collaboration Challenge. One of the requirement of Militaries is Cognitive Radio (CR), an adaptive, intelligent radio and network technology that can automatically detect available channels in a wireless spectrum and then change transmission parameters to communicate using them.
The [US] Department of Defense (DoD) is the ideal proving ground to try to bring AI to the spectrum, he noted. “We are usually an expeditionary force which means we don’t have an infrastructure to rely on and we don’t have time to conduct site surveys and determine the best place to put a cell tower; we get what we get and have to figure out how to work with it,” Tilghman explained, noting that the diverse range of radios operated by the armed forces further increase the challenge.
DARPA announced its new Radio Frequency Machine Learning Systems (RFMLS) program aimed to develop spectrum access strategies and a new wireless paradigm in which radio networks, will autonomously collaborate and reason about how to share the RF spectrum, avoid interference, and jointly exploit opportunities to achieve the most efficient use of the available spectrum.
Though the RF environment presents familiar ground for ML researchers, there are some important differences that make RFML unique. Compared to other domains, the RFML data rate is much higher. A high definition video stream at 60 Hz could be on the order of 3 Gbps (ignoring coding compression savings). 1 GHz of RF spectrum data has approximately 10 times that data rate, and is not nearly as amenable to decimation in order to improve processing.
Additionally, RF waveforms are typically captured and represented as complex numbers, underscoring the importance of both amplitude and phase of the signal. Although there has been interest recently in complex-valued neural networks, the technology for learning naturally in the complex plane is not fully developed and relies on treating complex variables as two real numbers. The RF environment offers a more compelling reason for developing a native complex-valued approach.
The RFMLS program features four technical components that would integrate into future RFML systems:
Applications of first-generation AI to the spectrum currently depend on hand-engineered features which an expert has selected based on the belief that they best describe RF signals pertinent to a specific RF task. Conversely, Deep Learning applied in other domains has achieved excellent performance in vision and speech applications by learning features similar to those learned by the brain from sensory data. Recently it has been shown that machine learning of RF features has the potential to transform spectrum problems in a way similar to other domains. Specifically, it
is expected that RFML Systems will be capable of learning the appropriate features to describe RF signals and associated properties from training data
From data sets of RF signals, RFML systems will need to learn the characteristics used to identify and characterize signals in various civilian and military settings.
Attention and Saliency:
Next-generation RF systems must evolve from characterizing MHz of spectrum to characterizing GHz of spectrum. To effectively make this step, RFML Systems must have a sense of which signals are important, focusing processing resources on those signals, and conserve resources by disregarding other signals that are unimportant to a particular task.
Humans are exquisite at consuming, prioritizing, and processing visual and auditory information. A major contributor to this ability is a sophisticated sense of importance defined by attention and saliency. Top-downattention is a goal-driven mechanism, which causes us to focus our cognitive processing on visual information most pertinent to a task at hand. For example, while looking for faces, humans will tend to focus their attention at eye-level since the floor is unlikely to contain many faces. Adhering strictly to goal-driven attention, however, would ultimately be myopic, missing new information relevant to the task at hand. Hence attention is complemented by a bottom-up e.g. data-driven mechanism called saliency, which enables us to capture changes or contrast differences in the visual or auditory scene.
Top-down attention in vision and speech is important to select the relevant low-level features needed for a classification or decision while removing distractors. Bottom-up saliency is important to call for top-down and cognitive attention due to, for example, a bright moving object or a loud clap. Similarly, for understanding an RF scene, stored RF concepts such as signals and transmitter types can be used to identify RF objects and model behavior. ML methods that include recurrent structures have been shown to create attentional functions in other modalities creating a path for such functions in an RFML System.
Researchers who win contracts to work on the RFMLS program will need to devise an equivalent within the RF domain of our own so-called salience detection, that is, the ability to identify and recognize important visual and auditory stimuli. The presence of a communications signal in a frequency band usually devoted to radar signals would be an example of a signal-of-interest that an RFMLS’s salience-detection capability would have to notice.
Autonomous RF Sensor Configuration:
Our eyes automatically adjust to changing light levels and they move and focus to keep the most important aspects of a dynamic visual scene in the most sensitive portions of the retina. The RFML systems that DARPA envisions would have an equivalent ability to automatically tune their receptivity to signals and signal features the systems deem to be most effective at accomplishing the task at hand.
Unlike traditional RF systems, visual processing is an autonomous sensory-motor task. If the environment is bright our eyes adjust. Saccadic vision directs eye movement to capture the most relevant portions of the visual scene. Much of the early research in Adaptive RF Systems focused on the analog RF electronics (i.e., the RF front end), building systems that can be reconfigured, so that the operating characteristics are not set in stone at design time.
Similarly recent advances have resulted in large-scale, all-digital RF systems that are intrinsically flexible. These advances are largely untapped because there has been little focus on coupling this adaptability with the necessary intelligence to reconfigure the RF sensor for better performance of the overall system task.
Learning for autonomous control has revolutionized game playing and self-driving cars. Although the sensors and actuators are different in RF applications, an RFML system can learn to optimize the huge number of possible configurations (analog electronics, beam steering angle, bandwidth sensitivity, position, etc.). ML methods that combine sensory processing (e.g. Convolutional Neural Networks) and decision-making when also combined with appropriate training methods (e.g. Reinforcement Learning) have enabled machine-learning systems to take on tasks like the game of Go, previously thought to be the domain of humans for at least the next decade.
The space of RF configurations is combinatorically large but, as in the game of Go, susceptible to ML. RFML Systems will be capable of autonomously reconfiguring the sensor over a large control space to improve overall task performance.
A full RFML system also should be able to digitally synthesize virtually any possible waveform, much as human beings can pronounce any new word or add inflections or pauses to infuse gravitas or nuances of meaning into what they saying. This capability to create new waveforms tailored to the specific RF devices they emanate from should give other sophisticated radios the improved ability to identify friendly systems.
Tasks such as waveform generation in present-day RF systems amount to decisions on a small set of discrete choices – use waveform A, or use waveform B. This fails to achieve true waveform synthesis – the ability to generate wholly new waveforms. Discrete selection approaches are akin to a speech-generation system preprogrammed with a discrete set of words, which it can choose between. Such a system could not pronounce new words or apply more
nuanced concepts such as syllabic stress common in modern speech synthesis systems. Similar challenges exist in the RF domain in coping with very large decision spaces. ML approaches, which enable unsupervised or semi/self-supervised learning of efficient waveform encodings, will enable RFML Systems to learn fundamentally new waveforms able to achieve one or more simultaneous objectives.
“If we get this right, we will have RF systems with the ability to discern and characterize signals in the ever-more-crowded spectrum. And that will give emerging automated systems, and the military commanders that rely on them, much needed information to understand the landscape of the wireless domain,” said Tilghman. “I hope our new RFMLS program will forge the technical foundations for a new domain and community of AI research.”
BAE Systems wins DARPA contract to apply machine learning to RF spectrum
The U.S. Defense Advanced Research Projects Agency (DARPA) has awarded McLean, VA-based BAE Systems a contract valued at $9.2 million for its Radio Frequency Machine Learning System (RFMLS) program. As part of the program, the company aims to develop new, data-driven machine learning algorithms that will help to decipher the ever-growing number of RF signals, providing commercial or military users with greater situational understanding of an operating environment, the company announced November 2018.
The inability to uniquely identify signals in an environment creates operational risk due to the lack of situational awareness, inability to target threats, and vulnerability of communications to malicious attack,” said Dr. John Hogan, product line director of the Sensor Processing and Exploitation product line at BAE Systems. “Our goal for the RFMLS program is to create algorithms that will enable a whole new level of understanding of the RF spectrum so users can identify and react to any signals that could be putting them in harm’s way.”
Under this Phase 1 contract, BAE Systems’ scientists intend to create machine learning algorithms, using cognitive approaches, that will use feature learning techniques to differentiate signals. In addition, researchers aim to create algorithms that can learn to differentiate important versus unimportant signals in real-time scenarios through a deep learning approach.