•  
  •  
 

Abstract

Autonomous Collaborative Combat Aircraft (CCA) operating in contested electromagnetic environments must classify Radio Frequency (RF) signals on edge silicon that degrades over the mission lifetime due to thermal stress, radiation, and manufacturing variation. Deep neural networks dominate RF classification on pristine hardware, but their weights are precise and interdependent, causing catastrophic accuracy collapse as the underlying chip ages. We investigate whether Hyperdimensional Computing (HDC), a brain-inspired paradigm that distributes information across thousands of dimensions, can provide a reliability floor where Deep Learning fails. Using the RadioML 2016.10A dataset filtered to five digital modulations relevant to drone command-and-control links, we trained a Random Fourier Features HDC model with dimension D = 10,000 and an adversarially trained "Steel Man" Multilayer Perceptron (MLP) baseline. After a kernel-bandwidth sweep identified gamma = 0.4 as the optimal HDC configuration, the MLP still achieved 83.9% high Signal-to-Noise Ratio (SNR) accuracy on clean CPU silicon while HDC reached 68.4%, a 15.5 percentage point deficit. However, when both models were deployed onto a Digital Twin of a Phase-Change Memory In-Memory Computing accelerator under post-training weight corruption, HDC lost only 1.4 percentage points at 20% hardware defects while the MLP lost 28.7 - a degradation rate roughly 20 times faster. HDC overtakes the MLP in absolute accuracy near the 10% defect rate, defining a mission-critical crossover. We argue that energy efficiency must be evaluated in terms of operational longevity rather than raw inference cost, and we propose a hybrid neuro-symbolic cascade for deployed CCA systems.

DOI

https://doi.org/10.5038/2378-0789.9.1.1162

Download

Share

COinS