Targeted Adversarial Examples for Black Box Audio Systems

Explore a conference talk on targeted adversarial examples for black box audio systems presented at the 2nd Deep Learning and Security Workshop during the 2019 IEEE Symposium on Security & Privacy. Delve into the application of deep recurrent networks in automatic speech recognition (ASR) systems and the vulnerabilities they face from adversarial perturbations. Learn about a novel black-box approach to adversarial generation that combines genetic algorithms and gradient estimation techniques. Discover how this method achieves an 89.25% targeted attack similarity and a 35% targeted attack success rate after 3000 generations, while maintaining 94.6% audio file similarity. Gain insights into the challenges and potential solutions for securing ASR systems against sophisticated attacks in scenarios where model architecture and parameters are unknown.