RiftNet is a powerful neural network architecture tailored and designed for RF fingerprint applications and novel device detection. Several innovations contribute to its capabilities. At its foundation RiftNet leverages a stack of dilated causal convolutional (DCC) layers to efficiently extract a wide array of discriminating features from raw IQ data. The DCC layers allow RiftNet to exponentially increase the receptive field of its deeper latent representations without the large computational cost that comes from traditional convolutional layers.
RiftNet also has two processing chains for handling the presence of encoded soft-IDs, like a MAC address, which can frustrate learning RF fingerprints. The non-ID processing chain is optimal for targeting constant or non-ID containing portions of a signal like the synchronization sequences found in many communications protocols. The soft-ID sensitive processing chain is setup to handle portions of signals which may contain soft-IDs. The combination of these two processing chains give RiftNet the ability to extract maximal useful information from any kind of RF signal. To keep the implementation simple for new users we don't include the soft-ID processing chain in this release, but performance without it is still high.
RiftNet is also very sample efficient learning well from as few as 100 signals per device. This computational efficiency means most test cases can train on one NVIDIA Tesla V100 GPU in a matter of hours.