Adversarial Attacks against Neural Networks in Audio Domain: Exploiting Principal Components

원문 초록

Adversarial attacks are inputs that are similar to original inputs but altered on purpose. Speech-to-text neural networks that are widely used today are prone to misclassify adversarial attacks. In this study, first, we investigate the presence of targeted adversarial attacks by altering wave forms from Common Voice data set. We craft adversarial wave forms via Connectionist Temporal Classification Loss Function, and attack DeepSpeech, a speech-to-text neural network implemented by Mozilla. We achieve 100% adversarial success rate (zero successful classification by DeepSpeech) on all 25 adversarial wave forms that we crafted. Second, we investigate the use of PCA as a defense mechanism against adversarial attacks. We reduce dimensionality by applying PCA to these 25 attacks that we created and test them against DeepSpeech. We observe zero successful classification by DeepSpeech, which suggests PCA is not a good defense mechanism in audio domain. Finally, instead of using PCA as a defense mechanism, we use PCA this time to craft adversarial inputs under a black-box setting with minimal adversarial knowledge. With no knowledge regarding the model, parameters, or weights, we craft adversarial attacks by applying PCA to samples from Common Voice data set and achieve 100% adversarial success under black-box setting again when tested against DeepSpeech. We also experiment with different percentage of components necessary to result in a classification during attacking process. In all cases, adversary becomes successful.