Study: on average, collusion attacks can break the traceability of multiple audio watermarks

1. 14:39
   arXiv

   Averaging collusion attacks undermine the traceability of personalized neural audio watermarking

   The researchers found that averaging two personalized watermark copies of the same recording sample by sample can make the tracking failure rate of five neural audio watermarking systems reach 87.7% to 99.7%, while audio quality is largely preserved. Decoder outputs react differently to the composition of collusion payloads across systems; in TimbreWM most consistencies are higher, while in WMCodec they are markedly lower. In attacks with a known payload setting, mixing weights can steer the decoder toward a chosen non-member, reaching a 92.3% target success rate on TimbreWM. A preliminary confidence screening can reject many collusion-induced attributions while retaining most valid single-copy outputs. Based on this, the authors propose that multi-copy traceability should be an additional evaluation dimension beyond conventional single-copy robustness; code and results are publicly available on GitHub.

Current attention 36·Peak in the comparable period 36 (Nov 8 19:00)·Change over the comparable past 24 hours 0%