RiftNet Reconstruction Model for Radio Frequency Domain Waveform Representation and Synthesis

Waveform representation, manipulation, and synthesis are challenging problems in the RF domain traditionally demanding expert knowledge to produce transparent and efficient solutions. In this work we present a low-complexity neural network architecture for waveform representation, manipulation, and synthesis. We demonstrate this architecture's performance by training it to represent Wi-Fi 802.11a/g waveforms and modify them with the objective of enhancing waveform distinguishability for RF fingerprint classification. We further present analysis of the network waveforms' latent representation to discover time and frequency properties of the learned transform. We discuss these properties in the context of traditional signals processing transforms to increase understanding and transparency of the algorithm and inspire future research into this domain. Although we target RF domain applications, we expect this architecture's performance and benefits to have high transferability to other domains.