earticle

영어

Audio recognition systems based on artificial intelligence are vulnerable to adversarial attacks. We designed audio adversarial attack models and experimented with clip, adaptive noise, high frequency emphasis, and invert phase attack models, as well as hybrid attack models based on these, to confirm through experimentation how these attack models affect adversarial attacks. The research method involved experimenting with nine attack models using the Interactive Emotional Dyadic Motion Capture Database dataset and evaluating how the attacks changed emotional states based on superb/wav2vec2-base-superb-er. Evaluation metrics included emotional scores and Peak Signal to Noise Ratio. Our research results confirmed attack success in the adaptive noise model, hybrid adaptive clip model, and hybrid invert phase model. Our research contributes to voice authentication and defense technique design.