Isolated and Ensemble Audio Preprocessing Methods for Detecting Adversarial Examples against Automatic Speech Recognition

Understand

An adversarial attack is an exploitative process in which minute alterations are made to natural inputs, causing the inputs to be misclassified by neural models.

• In the field of speech recognition, this has become an issue of increasing significance.
• Although adversarial attacks were originally introduced in computer vision, they have since infiltrated the realm of speech recognition.
• In 2017, a genetic attack was shown to be quite potent against the Speech Commands Model.