Arbitrarily Accurate Classification Applied to Specific Emitter Identification

TL;DR

This paper presents a method for achieving arbitrarily high classification accuracy by iteratively evaluating subsamples, demonstrated on specific emitter identification using deep learning on bispectra of I/Q signals, with minimal computation.

Contribution

It introduces a subsample evaluation technique that guarantees arbitrary accuracy, applied to emitter identification with a multi-channel CNN on bispectra, showing rapid error reduction.

Findings

Error rate decreases logarithmically with sample count

High accuracy achieved with minimal computation time

Each additional eight samples reduces error by tenfold

Abstract

This article introduces a method of evaluating subsamples until any prescribed level of classification accuracy is attained, thus obtaining arbitrary accuracy. A logarithmic reduction in error rate is obtained with a linear increase in sample count. The technique is applied to specific emitter identification on a published dataset of physically recorded over-the-air signals from 16 ostensibly identical high-performance radios. The technique uses a multi-channel deep learning convolutional neural network acting on the bispectra of I/Q signal subsamples each consisting of 56 parts per million (ppm) of the original signal duration. High levels of accuracy are obtained with minimal computation time: in this application, each addition of eight samples decreases error by one order of magnitude.