Li Lu

STEP introduces a novel, black-box, retraining-free detector that profiles audio samples using dual perturbation branches to detect backdoor attacks by exploiting the characteristic instability of hidden triggers.

The paper introduces AudioHijack, a framework that successfully demonstrates context-agnostic and imperceptible auditory prompt injection attacks, showing that commercial Large Audio-Language Models can be hijacked with high success rates.

The paper introduces ToxiAlert-Bench, a large-scale audio dataset that uniquely annotates both textual and paralinguistic sources of toxicity, and proposes a dual-head neural network that significantly improves speech toxicity detection by leveraging these paralinguistic cues.