Margin for Error: Large-Scale Simulation of In-Air Acoustic Attacks

The researchers propose a high-throughput realistic simulation framework and conduct over 8 million adversarial evaluations on speech control systems; the results show that on Whisper and wav2vec, accounting for acoustic factors can increase relative word error rate by up to 94.5%. The paper points out that migrating purely digital adversarial pipelines to the physical world faces scaling obstacles, and existing work has therefore abstracted away detectability and the effect of geometric structure on acoustics. The authors also formalize and propose Dual-Form Signal to Noise Ratio, used to decouple the stealthiness of the sound source from the attack effectiveness against the victim system, and state that this addresses a key limitation of current research and lays the foundation for reproducible, verifiable research.