Research papers · Posted 2020.11.24 · Publication / acceptance: author-reported
Provably-Robust Runtime Monitoring of Neuron Activation Patterns
Provably-Robust Runtime Monitoring of Neuron Activation Patterns
Overview
The monitor construction uses symbolic reasoning to bound neuron values under input perturbations before abstracting training activation patterns. Multi-bit intervals extend the representation, aiming to reduce false alarms while retaining provable robustness properties.
Publication and version information
Publication, acceptance or venue information is author-reported on arXiv. The exact source wording is preserved below.
Author-supplied source information
Author-supplied information on arXiv.
Paper accepted at IEEE DATE'21
Version: arXiv v2 — manuscript revision