Multivariate stochastic approximation to tune neural network hyperparameters for criticial infrastructure communication device identification

Abstract

The e-government includes Wireless Personal Area Network (WPAN) enabled internet-to-government pathways. Of interest herein is Z-Wave, an insecure, low-power/cost WPAN technology increasingly used in critical infrastructure. Radio Frequency (RF) Fingerprinting can augment WPAN security by a biometric-like process that computes statistical features from signal responses to 1) develop an authorized device library, 2) develop classifier models and 3) vet claimed identities. For classification, the neural network-based Generalized Relevance Learning Vector Quantization-Improved (GRLVQI) classifier is employed. GRLVQI has shown high fidelity in classifying Z-Wave RF Fingerprints; however, GRLVQI has multiple hyperparameters. Prior work optimized GRLVQI via a full factorial experimental design. Herein, optimizing GRLVQI via stochastic approximation, which operates by iterative searching for optimality, is of interest to provide an unconstrained optimization approach to avoid limitations found in full factorial experimental designs. The results provide an improvement in GRLVQI operation and accuracy. The methodology is further generalizable to other problems and algorithms.