Vibrational\nSpectroscopy Can Be Vulnerable to Adversarial\nAttacks

Nondestructive detection methods based on vibrational\nspectroscopy\nhave been widely used in many critical applications in a variety of\nfields such as the chemical industry, pharmacy, national defense,\nsecurity, and so on. As these methods/applications rely on machine\nlearning models for data analysis, studying the threats associated\nwith adversarial examples in vibrational spectroscopy and defenses\nagainst them is of great importance. In this paper, we propose a novel\nadversarial method to attack vibrational spectroscopy, named SynPat,\nwhere synthetic peaks produced by a physical model are placed at key\nlocations to form adversarial perturbations. Our new attack generates\nperturbations that successfully deceive machine learning models for\nRaman and infrared spectrum analysis while they blend much better\ninto the spectra and hence are unnoticeable to human operators, unlike\nthe existing state-of-the-art adversarial attacking methods, e.g.,\nimages and audio. We verified the superiority of the proposed SynPat\nby an imperceptibility test conducted by human experts and of defense\nexperiments by an AI detector. To the best of our knowledge, this\nis a first thorough study on the robustness of vibrational spectroscopic\ntechniques against adversarial samples and defense mechanisms. Our\nextensive experiments show that machine learning models for vibrational\nspectroscopy, including conventional and deep models for Raman or\ninfrared classification and regression, are all vulnerable to adversarial\nperturbations and thus may pose serious security threats to our society.