earticle

영어

Our study analyzed and evaluated the attack performance of hybrid combinations of Gaussian, Salt & Pepper, and Sine wave-based noise attack models. The experimental results confirmed that the hybrid attack model had the greatest impact on classification accuracy. Although some attacks did not change the output scores, misclassification was observed in emotional label classification. Our study validated that Gaussian, Salt & Pepper, and Sine wave-based noise attacks exploit security vulnerabilities that affect the learning of speech recognition systems and demonstrated the threat posed by hybrid attack models. The Gaussian model showed a significant decrease in emotion scores, while other models maintained high emotion scores. In terms of classification accuracy, changes were observed in the Gaussian, salt and pepper, and hybrid Gaussian models. Emotion attacks were demonstrated in the hybrid Gaussian, Salt & Pepper, and Hybrid sine wave experiments.