It's not trivial hiding information by slightly biasing the distribution of a generative model. The information hidden is added in a semantic meaningful way. Th

It's not trivial hiding information by slightly biasing the distribution of a generative model. The information hidden is added in a semantic meaningful way. This was not true with previous methods that ignored the semantics of the content in which they tried to hide information.