Compressed CNN for Inferring Rapid RF Fingerprints using Memristor Crossbar Array

Josh Li1, Jianbin Huang2, Michael B. Jiang3, Kang Jun Bai4
1University of Maryland Baltimore County, 2San Francisco State University, 3University of Illinois Urbana-Champaign, 4Air Force Research Laboratory


Abstract

In response to the growing demand of authenticating wireless devices in a large pool of internet of things (IoT) network, a convolutional neural network (CNN) has proposed to identify Wi-Fi transmitters based on their radio frequency (RF) fingerprints – raw in-phase and quadrature (I/Q) components – with decent accuracy. Meanwhile, the in-memory operator using memristor crossbar array for analog vector-matrix multiplication (VMM) has demonstrated its high energy efficiency and small form factor when implemented deep learning models liked CNN. In this work, we showcase a compressed CNN using average pooling methodology to reduce model parameters and required hardware resources for the inference operation. Specifically, by linearizing pooling layers with averaging approach, trained convolutional and pooling layers can be integrated into one single matrix – via matrix-matrix multiplication – which can be realized by a memristor crossbar array for efficient and rapid inference operation. Prototype and proof-of-concept demo were made for authentication of Wi-Fi transmitters based upon their raw RF fingerprints. With merely 2 crossbar arrays made of 3-bit memristors alongside the temporal encoding mode, our model demonstrates 91% accuracy, making such a neural network a good candidate when implemented in power-limited mobile edge devices.