Experimental & Comparative Study of Adversarial Attacks on Automatic Speech Recognition Systems

Abstract

Automatic Speech Recognition Systems are an integral part of modern-day artificial intelligence-based machines due to advancement in smart technology. Although ASRSs have provided users with the ability to perform important tasks through voice commands, their ability to withstand adversarial attacks is uncertain. The primary objective of adversarial artificial intelligence techniques is to disrupt the functioning of machine learning models with misleading data, by exploiting their vulnerabilities in decision-making process, causing them to make misclassifications or incorrect predictions. This poses a significant threat in the realm of AI and machine learning research, particularly in areas where machine performance is crucial. Therefore, it becomes imperative to study adversarial attacks, with a particular focus on assessing and comparing their severity. This would allow distinguishment between highly detrimental attacks and those with lesser impact, enabling the development of tailored defense strategies to effectively safeguard against them. This paper encompasses an experimental and comparative study of intensity of various attack methods on wav2vec2 model from the Torchaudio hub, particularly: Fast Gradient Sign Method, Basic Iterative Method, Projected Gradient Descent, Carlini and Wagner, and Imperceptible CW attack.