​Radio Frequency Machine Learning Systems (RFMLS) was a project early in the deep learning revolution (2017-2019) with the goal of developing the methods for applying the lessons we've learning from convolutional methods for vision to the RF domain. 

We designed, aggregated, and processed the largest machine learning quality RF domain dataset available. Defined a set of complex and challenging RF domain problems that could the dataset would be appropriate to experiment with and evaluated participant methods. This is effectively the government RF domain version of imagenet. 

This dataset consisted of complex domain signals which allowed us to explore both processing methods of data pre-training, complex valued neural network architectures, and fft matrix layer initialization techniques. 

---

Hardware Fingerprinting

Leveraging the perturbations in the received signals caused by the imperfections in the chips used to transmit, we were able to differentiate up to 10k unique wifi transmitters 

Broadband SOI Identification with confusors

Injecting signals of interest into broad band background recordings. We were able to differentiate between background and SOI signals and accurately identify classification of SOI. Injections were made in frequencies and with bandwidths that are not characteristic of the SOI injected forcing the model to learn non-trivial features

RL Resource Management of Receiver

A bespoke broadband receiver was build for this project that allowed us to develop an renforcement learning policy controller to exploit the spectrum to search for, identify, track and maintain SOI signals while continuing to explore the search space.