ORACLE: Optimized Radio clAssification through Convolutional neuraL nEtworks

Abstract

This paper describes the architecture and performance of ORACLE, an approach for detecting a unique radio from a large pool of bit-similar devices (same hardware, protocol, physical address, MAC ID) using only IQ samples at the physical layer. ORACLE trains a convolutional neural network (CNN) that balances computational time and accuracy, showing 99% classification accuracy for a 16-node USRP X310 SDR testbed and an external database of >100 COTS WiFi devices. Our work makes the following contributions: (i) it studies the hardwarecentric features within the transmitter chain that causes IQ sample variations; (ii) for an idealized static channel environment, it proposes a CNN architecture requiring only raw IQ samples accessible at the front-end, without channel estimation or prior knowledge of the communication protocol; (iii) for dynamic channels, it demonstrates a principled method of feedback-driven transmitter-side modifications that uses channel estimation at the receiver to increase differentiability for the CNN classifier.The key innovation here is to intentionally introduce controlled imperfections on the transmitter side through software directives, while minimizing the change in bit error rate. Unlike previous work that imposes constant environmental conditions, ORACLE adopts the 'train once deploy anywhere' paradigm with nearperfect device classification accuracy.

Data Augmentation for Deep Learning-Based Radio Modulation Classification

“…For the image, after rotation or flip augmentation, the same cat is displayed but from different viewpoints. In the constellation diagram of the QPSK modulated radio signal, the black circles indicate four ideal reference points, and the red crosses are the received symbols which are shifted due to the imperfection of transmitter/receiver hardware and wireless channel [23]. In Fig.…”

Transmitter Classification with Supervised Deep Learning

“…The first step is to establish a benchmark with the simplest scenario. This actually shows a comparison with the network in [8] without artificial impairments. In [8], they use a very similar setup of 16 static USRPs placed in a room with no moving object.…”

“…This actually shows a comparison with the network in [8] without artificial impairments. In [8], they use a very similar setup of 16 static USRPs placed in a room with no moving object. Their network achieves a classification accuracy of 98.6 % whereas the one studied here achieves 99.9 % with more classes (21 instead of 16) Fig.…”

“…This ensures that the change in signal propagation is exactly the same for the 3 scenarios. The Plain scenario suffers from a big loss of accuracy, as was noticed in [8], but the Robot one is very resilient to this kind of change. Finally, Fig.8 focuses on the ability to generalise to other scenario types.…”

“…However, in those works, either simulations or simple datasets were used which may not provide a biased free classifier that actually focuses on RF signatures. Furthermore, in those works, no attention is given on the quality of the dataset itself for the identification task at hand (absence of bias, reproducibility, interference with outside sources), and in [8], artificial impairments remove possible security claims: if identification is done on software added elements, any malicious software can do the same and impersonate the user.…”