Spread-Transform Dither Modulation Watermarking of Deep Neural Network

DNN watermarking is receiving an increasing attention as a suitable mean to protect the Intellectual Property Rights associated models. Several methods proposed so far are inspired popular Spread Spectrum (SS) paradigm according which watermark bits embedded into projection of weights model onto pseudorandom sequence. In this paper, we propose new algorithm that leverages on with side information decrease obtrusiveness and increase its payload. particular, scheme exploits main ideas ST-DM (Spread Transform Dither Modulation) improve performance recently based conventional SS. The experiments carried out by applying different models, demonstrate capability provide higher payload lower impact network accuracy than baseline method SS, while retaining satisfactory level robustness.