Deep neural networks (DNNs) have made remarkable progress in recent years and are widely applied across many fields. Trained DNNs are valuable assets due to their dependence on large volumes of quality data, expensive computational resources, and the development of sophisticated architectures. However, their increasing vulnerability to intellectual property (IP) infringement, including unauthorized use, replication, and redistribution, underscores the critical need for adequate copyright protection and integrity verification. DNN model watermarking has emerged as a promising solution to these challenges by embedding imperceptible identifiers into models. This survey provides a concise yet comprehensive state-of-the-art review of watermarking techniques focusing on DNN models and generative models, especially for Generative Adversarial Networks (GANs), and Diffusion models. We propose a taxonomy that analyzes existing methods by considering watermark-embedding strategy, robustness to various attacks, and impact on model performance. Specifically, we explore watermarking methods for generative models, including GANs and Diffusion models, and highlight the unique challenges posed by their complex output generation processes. Additionally, we provide a comprehensive comparison of evaluation metrics for existing watermarking methods. Furthermore, we identify challenges in watermarking techniques for model protection and integrity verification. Finally, we outline future directions, including universal watermarking frameworks, certified watermarking techniques, provenance for collective model training, robust watermarking strategies for LLMs, and a unified evaluation protocol and benchmark. A comprehensive review of watermarking techniques for DNNs and generative models, including GANs and Diffusion models. A structured taxonomy for existing watermarking methods along with comparison of evaluation practices and metrics. The outlined research gaps and future directions to promote the next generation of trustworthy AI ecosystems.
Liu et al. (Thu,) studied this question.