Large-Scale Simulation Reveals Acoustic Attack Vulnerabilities

A new study highlights the significant vulnerabilities of voice control systems to over-the-air acoustic attacks, a risk often underestimated due to the difficulty of scaling digital adversarial workflows to physical environments. Researchers developed a novel, high-throughput reality simulation framework to overcome these scaling barriers, enabling them to test over 8 million adversarial evaluations. This extensive testing revealed that incorporating acoustic awareness into attack strategies can dramatically increase the Word Error Rate (WER) in leading voice recognition models such as Whisper and wav2vec, with increases reaching up to 94.5%. The research also addresses methodological shortcomings in understanding these attacks by introducing a Dual-Form Signal to Noise Ratio. This new metric allows for the decoupling of an attack's stealth from its effectiveness against the target system, providing a more nuanced understanding of risk. By embracing the complexities of the acoustic environment rather than abstracting them away, this framework establishes a foundation for more repeatable and verifiable research into the security of human-AI voice communication.