VOC-Certifire: Certifiably Robust One-Shot Spectroscopic Classification via Randomized Smoothing.

Area of Science:

• Spectroscopy
• Machine Learning
• Chemical Sensing

Background:

• Spectroscopic methods offer advantages for gas detection in safety and operational efficiency.
• Challenges like noise, interference, and unseen conditions limit conventional machine learning (ML) models in laser-based sensors.
• Existing data augmentation strategies improve ML robustness but do not fully address these limitations.

Purpose of the Study:

• To detect pressure-induced spectral broadening using effective augmentations.
• To develop a one-shot learning approach for identifying up to 12 volatile organic compounds (VOCs) with minimal data.
• To provide provable certification for model predictions using randomized smoothing.

Main Methods:

• Implementation of simple yet effective data augmentation techniques.
• Development of a one-shot learning model (VOC-certifire) for rapid VOC identification.
• Application of randomized smoothing for certified prediction robustness.

Main Results:

• The VOC-certifire model achieved performance comparable to the baseline VOC-net model.
• The one-shot learning approach successfully identified up to 12 VOCs.
• Predictions from VOC-certifire were robust, reliable, and certified within a predefined norm radius.

Conclusions:

• The proposed VOC-certifire model offers a robust and certified solution for gas detection.
• One-shot learning with randomized smoothing effectively addresses data limitations and enhances model reliability.
• This approach is crucial for applications demanding high precision and consistency in gas sensing.