Black-box Speech Adversarial Attack with Genetic Algorithm and Generic Attack Ideas

Deep learning technology has been widely used in many fields related to people's life. However, current research shows that speech recognition models designed based on deep neural networks are extremely susceptible to adversarial samples. Black-box attack methods are closer to the actual situation, but the black-box attack techniques now in use have drawbacks such as low attack success rate, easy perception of the adversarial samples by humans, and more iterations required to generate the adversarial samples. Therefore, constructing black-box adversarial examples for speech recognition systems continues to be an extremely difficult challenge. In this paper, we propose a blackbox speech adversarial attack method based on genetic algorithm and generalized attack idea to address the problem of low attack success rate of existing black-box speech attack methods. By effectively combining a speech attack method based on a genetic algorithm design with a general speech attack method, at the same time, we designe a switching condition to switch the genetic algorithm module to a general attack method module to continue the attack process under certain conditions. This effectively improves the success rate of targeted attacks on DeepSpeech, a speech recognition model with good current speech recognition, and provides a new idea for the design of black-box speech adversarial attack methods in the future.