earticle

영어

Our research focuses on the possibility of adversarial attacks that manipulate subtle audio signals to influence model predictions, considering the current situation where audio alterations expose security vulnerabilities. Against this backdrop, it is necessary to evaluate the impact of existing audio attacks on their detectability even under realistic physical attack conditions. In particular, our research centers on low-perceptibility attacks—such as Fade in/out, Sine mod, Bit reduction, and Hybrid attacks—that induce errors in speech recognition learning without affecting the human auditory system, aiming to identify their attack effectiveness. In the research methodology, we conducted experiments on six types of attack methods and analyzed the detection evasion capabilities and attack efficiency of the models through quantitative results, including attack success rates, prediction confidence, and visual similarity. The results revealed that recognition rates decreased in Fade in/out-based attacks, and Sine mod attacks induced emotion recognition error rates without degrading audio quality. Additionally, the Fade+ Sine modulation + Bit reduction based Hybrid attack model demonstrated the most balanced distortion and was confirmed to be the most successful in evading detection.