From Inaudible Inputs to Model Failures: Low-Frequency Safety Risks in LALMs
Inaudible low-frequency audio can materially disrupt large audio-language models in black-box tests.
The paper introduces Intermittent Low-Frequency Lockout, a red-teaming method using a universal waveform template. Across six LALMs and multiple audio tasks, it cut accuracy by as much as 67 percentage points while remaining close to clean audio in human audibility ratings. The authors also propose Distributional Requery Guard, which detects low-frequency shifts and asks for a second recording; in their tests, it raised attacked accuracy from 28.5% to 46.1% after clean reacquisition. HF Daily Papers' note
The paper introduces Intermittent Low-Frequency Lockout, a red-teaming method using a universal waveform template. Across six LALMs and multiple audio tasks, it cut accuracy by as much as 67 percentage points while remaining close to clean audio in human audibility ratings. The authors also propose Distributional Requery Guard, which detects low-frequency shifts and asks for a second recording; in their tests, it raised attacked accuracy from 28.5% to 46.1% after clean reacquisition. HF Daily Papers' note
score 4