Do Modern Post-Hoc Watermarking Methods Beat Broken-Arrows? Rigorous Comparison Says Not Always

arXiv·medium signal

Gesny and Giboulot conduct the first rigorous comparison of modern neural-network-based image watermarking against classic methods like Broken-Arrows, specifically in real-world scenarios prioritizing robustness and security over benchmark metrics. The study finds modern methods don't consistently outperform classic approaches under adversarial conditions — timely context as SynthID adoption goes cross-industry and the watermarking arms race intensifies.