Deep learning watermarking techniques

Abstract

The growing prevalence and usage of Deep Learning models in critical applications has emphasised the need to protect such intellectual property from unauthorised usage, tampering or redistribution. This paper explores the usage of watermarks in the outputs of the model as a mechanism to claim ownership of the model. Specifically, the paper will benchmark recent watermarking techniques- through producing and analyzing its performance across a variety of adversarial attack categories, which include Destructive, Constructive and Regenerative attacks, as well as against combined multimodal attacks. Metrics such as bit accuracy and Frechet Inception Distance (FID), together with its robustness to adversarial attacks are utilised to evaluate these techniques comprehensively. The results will provide critical insights into the strengths and limitations of the current approaches, while contributing to the ongoing development of resilient watermarking techniques to ensure the traceability and security of Deep Learning Models.