Weight grouping method for generation management of watermarks for white-box DNN models

DNN models have become increasingly popular in recent years, and developers' rights need to be protected due to the various costs involved in creating DNN models. In this context, research on embedding watermarks to assert the copyright of DNN models has attracted attention. Furthermore, it is well known that secondary use of DNN models can be used to create high-performance models specialized for specific tasks, and that secondary use is readily available. In the case of legitimate secondary use, it is necessary to protect the rights of both the original model developer and the derived model developer. To achieve copyright protection of DNN models over multiple generations, this study proposes a method for embedding multiple watermarks. When multiple watermarks are embedded sequentially, there is a concern about interference between watermarks. The proposed method avoids this problem by grouping the DNN weight parameters used for watermarking. The proposed grouping also makes it possible to represent more information as watermark than in our previous study. Furthermore, embedding the watermark using the proposed method has a small impact on the original task.