Reverse-SynthID exposes SynthID spectral fingerprint

This is the kind of adversarial research watermarking schemes need, because fixed, learnable patterns are exactly what motivated reverse engineers will hunt for. The repo does not prove watermarking is dead, but it does argue that model-level fingerprints can be profiled, detected, and partially suppressed in practice.

• –The README claims 90% detection accuracy and a frequency-domain fingerprint extracted without proprietary access, which is the core technical story here
• –The most important implication is architectural: if SynthID relies on stable carrier frequencies and phase templates, attackers can build codebooks instead of brute-forcing edits
• –Google’s own SynthID positioning stresses robustness to common edits, so a public spectral-analysis repo directly pressures those claims and future watermark designs
• –Because the work is open source on GitHub, other researchers can now reproduce, challenge, or extend the findings instead of treating watermark security as a black box