An Optimization Framework for Generalized Relevance Learning Vector Quantization with Application to Z-Wave Device Fingerprinting

Multivariate Stochastic Approximation to Tune Neural Network Hyperparameters for Criticial Infrastructure Communication Device Identification

“…The work herein considered the Z-Wave dataset used previously in [13] [34]. This dataset considered three devices, ND = 3, which were Aeon Labs' Aeotec Z-Stick S2 transmitters.…”

“…MDA is a linear ML algorithm whereas GRLVQI is a nonlinear Artificial Neural Network (ANN) algorithm. Generally, for Z-Wave, GRLVQI outperforms MDA in classification accuracy [13]. However, GRLVQI has multiple algorithmic settings, called hyperparameters, which influence the way in which the classifier is trained.…”

“…In doing so, this extends the initial work of [18] by allowing GRLVQI to find optimal settings in an iterative fashion for its four continuous hyperparameters: gradient descent learning rate ( ), relevance learning rate ( ), conscience rate 1 ( ), and conscience rate 2 ( ). Additionally, the work of [13] is extended by creating a more efficient approach to optimizing GRLVQI settings and by creating a more robust approach to GRLVQI optimization by averaging replication results.…”

“…[13], examined full factorial GRLVQI algorithm optimization relative to Z-Wave device identification for its five hyperparameters. However, the results and approach of [13] were not ideal since they: 1) were highly tailored to the example Z-Wave dataset and thus possibly not applicable to other WPAN device signals, 2) involved exploring only a small region of the design space and utilized subject matter expertise to find appropriate high (+) and low (-) settings, and 3) involved only a single replication and thus ignored the randomness existing in any ANN algorithm, including GRLVQI, result.…”

“…More efficient alternatives to the approach of [13] include 1) considering an alternative experimental designs, e.g. [16] [17], and 2) stochastic estimation, e.g.…”