Reproduce “GhostPrefix: Synchronization-Free Universal Prefix Attacks for ASR Manipulation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Shengding Liu, Ashwin Innuganti, Qiben Yan
GhostPrefix learns model-specific audio prefixes prepended to speech input to steer automatic speech recognition to add a cue, produce an exact transcript, or degrade recognition. It trains prefixes using white-box access, then reuses them across unseen utterances for each model. On digital LibriSpeech tests, prompted success averaged 97.6% across 13 models; results do not establish cross-model transfer or real-world effectiveness.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.