DARE: Deceiving Audio–Visual speech Recognition model

Abstract

Audio–Visual speech recognition (AVSR) is an effective way to predict text corresponding to the spoken words using both audio and face videos, even in a noisy environment. These models find extensive applications in various fields like assisting hearing-impaired, biometric verification and speaker verification. Adversarial examples are created by adding imperceptible perturbations to the original input resulting in an incorrect classification by the deep learning models. Attacking an AVSR model is quite challenging, as both audio and visual modalities complement each other. Moreover, the correlation between audio and video features decreases while crafting an adversarial example, which can be used for detecting the adversarial example. We propose an end-to-end targeted attack, Deceiving Audio–visual speech Recognition model (DARE), which successfully performs an imperceptible adversarial attack while remaining undetected by the existing synchronisation-based detection network, SyncNet. To this end, we are the first to perform an adversarial attack that fools the AVSR model and SyncNet simultaneously. Experimental results on the publicly available dataset using state-of-the-art AVSR model reveal that the proposed attack can successfully deceive the AVSR model while remaining undetected. Furthermore, our DARE attack circumvents the well-known defences while maintaining a 100% targeted attack success rate.