FedMark: Privacy-Preserving Federated Learning-Based Watermarking for Large-Scale Image Datasets
Himanshu Kumar Singh, Kedar Nath Singh, Amit Kumar Singh · IEEE Transactions on Consumer Electronics · 2024
With the accelerated advancement of consumer devices and multimedia editing software, the manipulation and sharing of digital images have become ubiquitous. While these functions enhance user convenience in image editing, they also face more threats, such as data leakage and information theft. Deep learning-based watermarking provides a unique method of digital-image protection. However, it is challenging for existing approaches to provide an effective solution for privacy, generalisation, and scalability at the same time. This study proposes a federated learning-based watermarking framework, called FedMark, to improve the robustness and imperceptibility of watermarks in large-scale image datasets. It enables collaborative model training across distributed consumer devices while maintaining data privacy and model generalisation and scalability across diverse datasets. Empirical validation across multiple datasets shows that FedMark consistently outperforms existing methods with significantly improvement of 36.8% in terms of robustness and 48.2% in terms of imperceptibility while ensuring reversibility and maintaining stringent security standards. With its combination of federated learning and advanced watermarking techniques, FedMark is a promising step towards a secure, privacy-preserving future for digital-image watermarking.