Text

theses

With the widespread use of machine learning techniques in many areas of our life (e.g., recognizing images, videos, and voice commands), their security vulnerabilities have become an increasing concern of the public. For instance, a deep neural network (DNN), as a popular class of machine learning techniques, has been proved defenseless against deliberately designed adversarial attacks. Existing studies mainly focus on feeding adversarial images to fool an image classification system. They can make the system misclassify the images as any target object by adding carefully-crafted perturbations.
In this thesis, we propose an adversarial attack against automatic speech recognition (ASR) systems (e.g., Siri and Google Assistant). We demonstrate that an adversary can embed malicious voice commands into regular songs, and these embedded commands can be recognized by the ASR system but go unnoticed by humans. Particularly, we use a genetic-based algorithm to craft the original song with the probability density function (PDF) identifier of the malicious comments, allowing the crafted song to be recognized as the embedded comments. Evaluation demonstrates that the commands in the crafted songs can be successfully recognized by the ASR system (e.g., Kaldi) with an average success rate of over 58%. By calculating the covariance between the crafted song and the original one, we show that the similarity between them is over 95%, making it hard to be noticed by humans.

ETD_9872

M.S.

Includes bibliographical references

doi:10.7282/t3-2zvt-vy33

ETD graduate

The author owns the copyright to this work.