Spectroscopic methods are well-established and widely used tools in analytical chemistry. They leverage the interaction between light and matter to extract information about chemical species and their abundances. Application of spectroscopic methods is hindered by the need for large datasets and the presence of unknown interference. These problems present significant challenges in developing reliable machine learning models for spectroscopic gas sensing. In many real-world applications, data is scarce, and absorbance signals are often corrupted by noise or overlapping spectral features from interfering species, making accurate detection and classification difficult. To address these challenges, we apply a set of targeted augmentation strategies aimed at improving model robustness and selectivity in gas sensing tasks. Specifically, we propose a one-shot learning approach with Voigt profile augmentation to handle pressure-induced spectral variations. Additionally, we use fictitious augmentations to mitigate the impact of unknown interfering species.Furthermore, we apply randomized smoothing to enhance resilience to unseen perturbations and domain shifts, promoting consistent performance in noisy, real-world conditions. Our models significantly outperform undefended baselines, offering a reliable, data-efficient solution for gas detection, which is critical for well-informed decision-making and risk mitigation. Work in this area has the potential for great societal impact as it can be applied for occupational safety (detecting hazardous/toxic gas exposure), health applications (identifying bio-markers in exhaled breath), and environmental applications (monitoring air pollutants and green house gasses).