DWA-Watermarking Embedding Mechanism for Deep Learning Models

The black-box watermarking for deep learning models utilizes an overfitting trigger set to embed watermarking, and researchers focus on constructing the trigger set. However, this article does not focus on the form of trigger set construction, but rather on the training process of the trigger set. The proposed DWA-Watermarking embedding mechanism takes this into account. The idea of multitask learning is introduced in the watermarking embedding process, and the dynamic weight averaging algorithm is used for the first time to balance the training of the watermarking embedding task and the original task during the training process by dynamically adjusting the weight of the watermarking loss function. The proposed algorithm is tested on three datasets and three mainstream deep learning models. DWA-Watermarking embedding mechanism accelerates the speed of watermarking embedding and improves the extraction rate of the watermarking in the model