GBAA-GA: General Black-Box Adversarial Attack Method for Audio Recognition Based on Genetic Algorithm

Automatic Speech Recognition (ASR) has brought great convenience, whereas security issues are not supposed to be neglected. Universal adversarial attack techniques, which are known for their generalizability, have received wide-spread attention in academia. Current research primarily focuses on white-box attacks, where the attacker has full knowledge of the target model. However, obtaining a wide range of information is difficult at present. Therefore, a general black-box adversarial example attack algorithm, GBAA-GA, is designed based on a genetic algorithm. Specifically, this study first trained a speech command classification model based on the Speech Commands dataset as the attack target. Under the assumption that only the model's input and output information is accessible, the study incorporated the concept of genetic algorithm and introduced an adaptive mutation probability strategy to address the issue of local optima in the algorithm. This approach successfully generated universal audio adversarial examples in a black-box setting. Finally, comparative experiments with the white-box universal perturbation attack algorithm PGD-UAP demonstrate that GBAA-GA achieves performance comparable to PGD-UAP in key metrics such as attack success rate and stealthiness, thereby fully validating the effectiveness and practicality of the proposed algorithm.