Ownership protection of deep learning models using watermarking

Abstract

Deep learning (DL) models have achieved remarkable success in the multimedia domain. However, the potential misuse of these powerful models poses significant challenges in many security-sensitive domains. To address this issue, digital watermarking schemes have emerged, aiming to embed watermark information into DL models to prove copyright ownership. In this paper, we propose a copyright protection scheme for DL models to resolve ownership conflicts and enhance overall system security. Initially, we generate two different watermarks using chaotic and Gold sequences. The generated watermarks are then embedded into selected layers of the DL model using redundant discrete wavelet transform and singular value decomposition. After the watermarking process, we shuffle the model’s weights before sharing it with a third party. On the receiver’s side, the inverse of the embedding process, followed by watermark verification, ensures secure access to authentic and unaltered model data. Ownership analysis shows that the proposed scheme is robust against pruning and fine-tuning attacks. Experimental results further validate the effectiveness of the proposed scheme compared to other competitive approaches.