Radio Frequency Machine Learning Systems (RFMLS)

The goal of the Radio Frequency Machine Learning Systems (RFMLS) Program is to develop the foundations for applying modern data-driven Machine Learning (ML) to the RF Spectrum domain. These innovations form the basis of a new wave of Signal Processing technologies to address performance limitations of conventionally designed radio frequency (RF) systems such as radar, signals intelligence, electronic warfare, and communications.

Over the last decade ML has been applied successfully to numerous sensor modalities, and is now common place in many commercial applications, including object and facial recognition in images, speech recognition in acoustic signals, and text parsing and reasoning from documents. Key to each of these innovations was the evolution from hand-engineered approaches tailored to each problem, to solutions that learned from large datasets.

RF systems conversely are still designed using models and equations based on idealized assumptions and approximations regarding hardware, environment, and the problem being solved. The inaccuracy of these assumptions challenge our ability to perform tasks such as identification of signals among the ever-increasing myriad which populate the wireless landscape.

Under the program, RFMLS systems will seek to learn to perform four specific tasks. Each task emphasizes a core constituent capability of RF ML. The four solutions can be combined and applied to address DoD operational needs in the RF Spectrum.

1. RF Fingerprinting. Traditional wireless security relies on a software “identity” for each wireless device, which can often be hacked or otherwise cloned. The RFMLS system will aim to learn to recognize a specific transmitter based on the unique RF fingerprint naturally imparted by hardware imperfections within that transmitter. This task focuses on learning RF features.
2. RF Fingerprint Enhancement. To further enhance wireless security, a communication system learns to modify it’s transmit waveforms to enhance its natural fingerprint. This task focuses on learning to synthesize waveforms.
3. Spectrum Awareness. Traditional systems which monitor the RF spectrum use narrow bandwidths and relatively simple strategies (such as the frequency of transmission) to identify the signals occupying the wireless spectrum. Availability of commodity analog-to-digital converters with wide bandwidths combined with proliferation of software defined radios, spectrum sharing, and general wireless technology growth, challenge these approaches. RFMLS systems will learn to understand the difference between important and unimportant signals present in large bandwidths in order to build more useful and accurate spectrum awareness. This task emphasizes goal-driven attention.
4. Autonomous RF System Configuration. To further enhance spectrum awareness performance, a RFMLS system will seek to learn how best to tune and configure its hardware resources in order to maximize the number of important signals discovered in harsh RF environments. This task stresses hardware configuration and control.