Reproduce “ScrambleMark: Zero-Query Watermark-Agnostic Black-Box Evasion of Modern Speech Watermarking” 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
Binhao Ma, Weikang Ding, Hanqing Guo, Wenwei Zhao, Xiaowen Li, Leah Ding, Zhuo Lu, Rui Duan
ScrambleMark attacks speech watermarks without querying the target detector during an attack. It scrambles speech harmonics with randomized frequency-band shifts, selecting candidates using a public surrogate model and requiring neither clean target-speaker speech nor paired clean and watermarked examples. Across three commercial platforms, it averaged 0.98 attack success rate and 0.91 STOI; these results are limited to tested samples.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.