Adversarially robust multimedia watermarking via data-centric optimization

Digital watermarking plays a crucial role in protecting intellectual property rights and ensuring content authenticity across diverse multimedia formats. While existing watermarking techniques rely on extractors for detection, these extractors are vulnerable to adversarial attacks during transmission. Current model-level defenses through adversarial training require modifications to existing extractors, often leading to compromised extraction performance. To address these limitations, we propose Data-centric Adversarially Robust waTermarking (DART), a novel paradigm that incorporates defense mechanisms during the data creation process itself, enabling robust watermarking without modifying existing extractors. DART employs Implicit Neural Representations (INRs) as information carriers, reformulating watermarking from discrete pixel optimization to continuous function learning. This continuous and differentiable representation naturally accommodates various media types, including images and videos, providing a unified framework for multimedia watermarking. Extensive experiments demonstrate that our method outperforms baseline methods under the evaluated white-box and black-box attack scenarios while maintaining high perceptual quality and extraction accuracy.