From Inaudible Inputs to Model Failures: Low-Frequency Safety Risks in LALMs

Intermittent Low-Frequency Lockout (ILL) identifies safety vulnerabilities in Large Audio-Language Models (LALMs) by using inaudible, low-frequency waveforms to trigger model failures. The method employs Sentence Attention Scale Estimation and Frequency Confusion Transfer to construct these signals. Distributional Requery Guard (DRG) is proposed to mitigate these risks.