Targeted Adversarial Examples for Black Box Audio Systems

The course teaches learners how to generate targeted adversarial examples for black box audio systems using a combination of genetic algorithms and gradient estimation. The learning outcomes include achieving a high targeted attack similarity and success rate while maintaining audio file similarity. The course focuses on the application of these techniques to fool deep neural networks in automatic speech recognition systems. The intended audience for this course is individuals interested in deep learning, security, and audio transcription. The teaching method involves a presentation delivered at a workshop, providing insights and findings from the research conducted in this area.