Researchers introduce ScrambleMark to evade speech watermarks on three platforms

Researchers propose ScrambleMark, a zero-query, watermark-agnostic black-box attack that evades modern speech watermarking by disrupting harmonic structures. The method combines a deterministic perturbation discovery engine with a generative randomized scrambling model. Testing across three commercial platforms showed that ScrambleMark achieves nearly 97 percent evasion while maintaining high speech naturalness and withstanding adaptive recovery. Open article →