Adversarial Attacks on ASR Systems
Course Project - Automatic Speech Recognition
Instructor - Prof. Preethi Jyothi, Dept of Computer Science, IIT Bombay
- Reviewed literature on targeted, imperceptible, white & black-box adversarial attacks on the ASR systems - Trained a Bi-RNN CTC-based network on the SpeechCommands dataset with a WER of 16%
- Implemented Gradient-descent based adversarial attack achieving a 0% classification accuracy along with a Signal-to-Noise Ratio (SNR) of 30dB in the perturbed audio examples