Feature Selection and Classifier Development for Radio Frequency Device Identification
Abstract
The proliferation of simple and low-cost devices, such as IEEE 802.15.4 ZigBee and Z-Wave, in Critical Infrastructure (CI) increases security concerns. Radio Frequency Distinct Native Attribute (RF-DNA) Fingerprinting facilitates biometric-like identification of electronic devices emissions from variances in device hardware. Developing reliable classifier models using RF-DNA fingerprints is thus important for device discrimination to enable reliable Device Classification (a one-to-many looks most like assessment) and Device ID Verification (a one-to-one looks how much like assessment). AFITs prior RF-DNA work focused on Multiple Discriminant Analysis/Maximum Likelihood (MDA/ML) and Generalized Relevance Learning Vector Quantized Improved (GRLVQI) classifiers. This work 1) introduces a new GRLVQI-Distance (GRLVQI-D) classifier that extends prior GRLVQI work by supporting alternative distance measures, 2) formalizes a framework for selecting competing distance measures for GRLVQI-D, 3) introducing response surface methods for optimizing GRLVQI and GRLVQI-D algorithm settings, 4) develops an MDA-based Loadings Fusion (MLF) Dimensional Reduction Analysis (DRA) method for improved classifier-based feature selection, 5) introduces the F-test as a DRA method for RF-DNA fingerprints, 6) provides a phenomenological understanding of test statistics and p-values, with KS-test and F-test statistic values being superior to p-values for DRA, and 7) introduces quantitative dimensionality assessment methods for DRA subset selection.
Document Details
- Document Type
- Technical Report
- Publication Date
- Dec 01, 2015
- Accession Number
- AD1003078
Entities
People
- Trevor J Bihl
Organizations
- Air Force Institute of Technology