Abstract
The rise of artificial intelligence (AI) and machine learning (ML) has prompted concerns regarding the intellectual property (IP) protection of neural networks (NNs). A proposed solution is watermarking, which incorporates a unique identifier into a NN. However, the effectiveness of watermarking methods in enhancing privacy and ownership secrecy remains questionable. This study intended to evaluate the efficacy of watermarking techniques for enhancing the security and ownership protection (SOP) of NNs. An exhaustive search of scholarly databases for peer-reviewed journal articles and conference proceedings was conducted in accordance with PRISMA standards. Eligible papers evaluated the efficacy of watermarking techniques used to protect NNs. Twenty research articles using various watermarking techniques, including digital watermarking (DW), reversible watermarking (REW), and robust watermarking (ROW), were analyzed. Various performance indicators, such as detection rate (DR), robustness, and distortion, were employed to evaluate the applicability of each method. The results demonstrated that watermarking techniques effectively protected the intellectual property of NNs with minimal impact on performance. However, the need for specialized apparatus and the difficulty of incorporating watermarks into deep neural networks (DNN) hampered their implementation. To improve the practicability and effectiveness of watermarking techniques, additional research is required. Researchers, professionals, and policymakers should consider watermarking to safeguard the intellectual property of NNs in a variety of domains, including finance, healthcare, and national security.
