Exploiting Neural Audio Codec Latents for Adversarial Audio Attacks

The article presents a novel generative attack framework that operates in the continuous latent space of a neural audio codec, enabling targeted adversarial audio attacks with high efficiency. The method employs a conditional generator to produce class-specific perturbations in a single forward pass, achieving success rates of up to 99% with an inference latency reduced to under 7 ms, which is 24 times faster than existing generative baselines. This advancement is significant for practitioners as it offers a more effective and efficient approach to assessing the robustness of audio classification systems against adversarial threats.