@article{CRSIJ26000216,
                  author = {Agbagbo Princewill, Daniel Ekpah},
                  title = {Watermarking and intellectual property protection in neural networks},
                  journal = {Cosmo Research and Science International Journal},
                  year = {2025},
                  volume = {2},
                  number = {1},
                  pages = {22-42},
                  issn = {3108-1584},
                  url = {https://cosmorsij.com/published/CRSIJ26000216.pdf},
                  abstract = {The rapid commercialization of deep neural networks has intensified concerns regarding intellectual property theft, unauthorized model redistribution, and illegal replication of artificial intelligence systems. As trained neural network models require significant computational resources, large datasets, and expert knowledge, effective ownership protection mechanisms have become increasingly important. This study examines major neural network watermarking and fingerprinting techniques, including white-box watermarking, black-box watermarking, parameter-based fingerprinting, behavioural fingerprinting, and adversarial fingerprinting, with the aim of evaluating their effectiveness in protecting AI intellectual property. The study analyses these techniques based on model accuracy, watermark detection rate, robustness against attacks, traceability efficiency, and computational overhead. The findings reveal that white-box watermarking achieved a high watermark detection rate of 96.5%, demonstrating strong ownership verification capability due to the embedding of watermark information directly into internal model parameters. However, this approach introduced moderate computational overhead during training and verification. Black-box watermarking exhibited slightly lower robustness of 84.7% but maintained lower computational complexity, making it highly suitable for commercial AI APIs and cloud-based machine learning services where internal model access is restricted. The results further indicate that fingerprinting techniques significantly improve the ability to trace unauthorized model redistribution while preserving high predictive performance. Parameter-based fingerprinting achieved a fingerprint detection rate of 95.4% and strong traceability performance, although its resistance to collusion attacks was lower than adversarial approaches. Behavioural fingerprinting maintained the highest model accuracy of 98.1% with relatively low computational overhead, demonstrating strong suitability for black-box AI environments. Adversarial fingerprinting achieved the highest robustness against collusion attacks at 91.4% and the highest traceability efficiency of 94.1%, indicating superior resistance against sophisticated attacks designed to remove or conceal ownership information. Nevertheless, adversarial fingerprinting incurred higher computational costs due to the complexity of generating adversarial examples. The study also examines major threat models and attacks against watermarking systems, including fine-tuning attacks, pruning attacks, model extraction, collusion attacks, trigger inversion attacks, overwriting attacks, and adversarial evasion attacks. Despite minor reductions in model accuracy, the findings demonstrate that watermarking and fingerprinting techniques substantially enhance intellectual property protection in neural networks. The study concludes that fingerprinting and white-box watermarking provide the strongest protection for high-security AI applications, while black-box methods offer practical deployment advantages for cloud-based systems. Overall, neural network watermarking and fingerprinting represent critical solutions for safeguarding artificial intelligence assets in modern machine learning environments.},
                  keywords = {},
                  month = {July}
        }