Masking models and watermark undetection
Arnaud Robert, Justin Picard · Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE · 2001
Masking models are mostly used in data compression algorithms and serve to shape the quantization noise. They were introduced in watermarking as to indicate the regions where the watermark could be introduced without perceptible artifacts. This allowed to embed more watermark energy, for a given absolute distortion constraint, than if no mask is used. Yet, little attention has been paid to the consequences of using these masks with respect to detection performance. In this work, it is shown that blind use of masking models facilitates the attacker's role, and eventually results in severe decreases of detection statistic at the detector, even for reasonable attack distortions.