Acoustic Interference Attack (AIA) introduced

Researchers demonstrated that specifically crafted innocuous audio tracks containing certain Acoustic Latent Semantics (ALS) patterns can disrupt the safety alignment mechanisms of Large Audio-Language Models (LALM) and lead to the inference path drift. In this context, audio acts not merely as a payload carrier but as an active tool for breaking the model's safety alignment. Unlike many prior attacks on LALMs, AIA leverages universal audio fragments and does not require tuning or text optimization for each individual prompt.

💬 Discuss