After 15 years of intense activity, research in media watermarking has seen a declining interest due to the relative maturity of the field. In the last few years, though, the new opportunities offered by Deep Learning (DL) technology and the necessity  of protecting the IPR associated to Deep Neural Networks (DNNs) is revitalising the interest in watermarking technology. On one hand, DNN models can be used for traditional media watermarking (and data hiding in general) offering possible advantages due to the superior capabilities of deep networks to analyze the content of media assets and process them even in the absence of precise statistical models. On the other hand, and even more interestingly, watermarking DNN models for ownership verification or to trace the usage of pre-trained models, presents opportunities and challenges that were not present in the case of media watermarking. As a matter of fact, the three basic requirements of any watermarking system, namely, unobtrusiveness, robustness and even payload, assume a different meaning in the DNN case. For this reason, new techniques have been proposed in the last few years, revealing an extremely rich, and somewhat unexpected, variety of interconnections with some of the hottest topics in DL research, including, adversarial examples, DNN backdoors, identification of synthetic media produced by Generative Adversarial Networks (GAN). It is the goal of this SS to present a snapshot of the current research in DNN watermarking, helping researchers to spot the most interesting open issues and opportunities associated to the field, possibly treasuring on the wealth of theoretical and practical insights stemming from traditional media watermarking.

A non-comprehensive list of topics addressed by the Special Session include: