FGRW: Fine-Grained Reversible Watermarking Based on Distribution-Adaptive Contrastive Augmentation Across Diverse Domains

Watermarking-based ownership protection is widely adopted to control usage rights and implement integrity verification by discovering specific properties or clues inherited from the corresponding models (<italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">i.e.</i>, owned models or released models). However, the cross-domain fine-tuning operations on pre-trained models to enable these models to adapt downstream tasks, which may compromise watermarking stability and accuracy. To address this issue, we propose a fine-grained reversible model watermarking scheme that can automatically adjust the synthetic samples generated from the contrastive augmentation network and the modification operations performed by the watermark embedding network. Specifically, the source and target data are exploited to generate the trigger and task samples by out-of-distribution (OOD) awareness, which can promote the robustness of embedded watermark and maintain the performance of task-related model. Subsequently, the representation-oriented regularization constraint is designed to precisely match suitable modification spaces, while the embedding loss is fed back to the augmentation network to guide the sample generation. Extensive experiments show that our scheme has superior performance through fine-grained authorization and reversible integrity verification. Furthermore, our results reveal there exists a trade-off between generating task-related samples for training a clean ideal model and trigger samples for obtaining a watermarked model.