A Black-Box Watermarking Modulation for Semantic Segmentation Models

Abstract

The capability of clearly identifying the origin of a ML model is an important element of trustworthy AI. First standardisation reports highlight the necessity of providing ML traceability, while pointing out that existing tools for Digital Right Management are not sufficient in the context of ML. Watermarking has been explored as a possible answer for this need, and has been implemented for image classification models, but there remains a substantial research gap in its application to other tasks such as object detection or semantic segmentation, which remains largely unexplored. In this paper, we propose a novel black-box watermarking technique specifically designed for semantic segmentation. Our contributions include a novel watermarking method links visual data to text semantics and provide comparative analysis of the effect of fine-tuning techniques on watermark detectability. Finally, we highlight several regulatory recommendations on how to design watermarking techniques for segmentation purposes.