VOC-Certifire:\nCertifiably Robust One-Shot Spectroscopic\nClassification via Randomized Smoothing

Spectroscopic methods are advantageous for gas detection\nwith applications\nranging from safety to operational efficiency. Despite the potential\nof laser-based sensors, real-world challenges, such as noise, interference\nand unseen conditions, hinder the accurate identification of species.\nThe use of conventional machine learning (ML) models is constrained\nby extensive data requirements and their limited adaptability to new\nconditions. Although augmentation-based strategies have proven to\nimprove the robustness of machine learning models, they still do not\noffer a complete defense. To address these challenges, this study\nfocuses on three primary goals: first, to detect pressure-induced\nspectral broadening using simple yet effective augmentations; second,\nto bypass the need for extensive data sets by deploying a one-shot\nlearning approach that can identify up to 12 volatile organic compounds\n(VOCs); and third, to provide a provable certification for the one-shot\nlearning model predictions via randomized smoothing. To assess the\neffectiveness of our proposed augmentations and randomized smoothing,\nwe perform a comparative study with four distinct models: VOC-net,\nVOC-lite, VOC-plus, and VOC-certifire. Remarkably, the one-shot learning\nmodel proposed herein, VOC-certifire, delivers predictions that match\nthe baseline model VOC-net. The VOC-certifire predictions not only\nexhibit robustness and reliability but are also certified within a\npredefined l2 norm radius. Such a certification is particularly\nuseful for gas detection, where the robustness, precision and consistency\nare key to well-informed decision-making.