Perceptual-based deep-learning denoiser as a defense against adversarial\n attacks on ASR systems

2021/07/12 by Anirudh Sreeram, Nicholas Mehlman, Sreeram, Anirudh +7
Computer Science · Engineering · #Speech Recognition and Synthesis #Adversarial Robustness in Machine Learning #Geophysical Methods and Applications

paper · pdf · doi:10.48550/arxiv.2107.05222

Abstract

In this paper we investigate speech denoising as a defense against\nadversarial attacks on automatic speech recognition (ASR) systems. Adversarial\nattacks attempt to force misclassification by adding small perturbations to the\noriginal speech signal. We propose to counteract this by employing a\nneural-network based denoiser as a pre-processor in the ASR pipeline. The\ndenoiser is independent of the downstream ASR model, and thus can be rapidly\ndeployed in existing systems. We found that training the denoisier using a\nperceptually motivated loss function resulted in increased adversarial\nrobustness without compromising ASR performance on benign samples. Our defense\nwas evaluated (as a part of the DARPA GARD program) on the 'Kenansville' attack\nstrategy across a range of attack strengths and speech samples. An average\nimprovement in Word Error Rate (WER) of about 7.7% was observed over the\nundefended model at 20 dB signal-to-noise-ratio (SNR) attack strength.\n

Related

• Adversarial Attacks and Defenses for Speech Recognition Systems
• Defense against Adversarial Attacks on Hybrid Speech Recognition using Joint Adversarial Fine-tuning with Denoiser
• Catch Me If You Can: Blackbox Adversarial Attacks on Automatic Speech Recognition using Frequency Masking
• Recent improvements of ASR models in the face of adversarial attacks
• Adversarial Example Devastation and Detection on Speech Recognition System by Adding Random Noise