TUM Logo

Language Dependencies in Generating Audio Adversarial Attacks

Language Dependencies in Generating Audio Adversarial Attacks

Supervisor(s): Karla Markert
Status: finished
Topic: Others
Author: Donika Mirdita
Submission: 2020-09-15
Type of Thesis: Masterthesis
Thesis topic in co-operation with the Fraunhofer Institute for Applied and Integrated Security AISEC, Garching

Description

Audio Speech Recognition (ASR) systems are ubiquitous presences in our 
online devices and their vulnerability to attacks has become a matter of 
research interest. While the core machine learning algorithms that 
enable these systems have already been analyzed in detail, there exists 
no comparative analysis of vulnerabilities between identical machine 
learning architectures trained on different language datasets. This 
project investigates how ASR models for English and German hold up under 
a set of attacks and whether one of the languages is more susceptible to 
manipulations than the other. The results of this experiment suggest 
statistically relevant differences between English and German in terms 
of computational effort necessary for the successful generation of 
adversarial examples.