GWO-GEA: A Black-box Method for Generating Audio Adversarial Examples

In the context of launching adversarial assaults on Automatic Speech Recognition (ASR) systems, the intricacies of the target model, including its architecture and parameters, remain obscured to the attacker. Predominantly, assault techniques have been tailored for white-box models, leading to a notable struggle in crafting adversarial examples for black-box models, often marked by inefficiencies and low success probabilities. In response to this predicament, we introduce an innovative approach for generating voice adversarial examples specifically designed for black-box models. This method, dubbed GWO-GEA, integrates the Grey Wolf Optimizer with Gradient Estimation algorithms, leveraging these optimization techniques to enhance the generation process. This paper presents a method that fully leverages the fast convergence of the Grey Wolf algorithm and the fine optimization of objectives by the Gradient Estimation algorithm, allowing for the rapid and efficient generation of high-quality adversarial voice samples. This paper’s adversarial sample generation methods for two different scenarios are tested and analyzed on three public speech datasets: The Speech Commands Dataset, Mozilla Common Voice dataset, and LibriSpeech. The results show significant improvements in attack success rate, adversarial sample generation quality, and generation time compared to common methods. Furthermore, we have demonstrated on multiple commercial APIs that the adversarial examples generated by our method have good transferability and can adapt well to real-world attack scenarios.