Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

MT-Mark: Rethinking Image Watermarking via Mutual-Teacher Collaboration with Adaptive Feature Modulation

Fei Ge, Ying Huang, Jie Liu, Guixuan Zhang, Zhi Zeng, Shuwu Zhang, Hu Guan

TL;DR

<3-5 sentence high-level summary> MT-Mark tackles robustness in image watermarking by replacing the traditional fixed embedder–extractor pipeline with an explicitly collaborative framework. It introduces a Collaborative Interaction Mechanism (CIM) that enables bidirectional, mutual-teacher communication between embedding and extraction, guided by Adaptive Feature Modulation Modules (AFMM) for content-aware feature regulation. The approach shifts robustness learning from distortion enumeration to coordinated representation learning, achieving superior watermark extraction accuracy and perceptual quality on real and AI-generated datasets. Ablation studies confirm the necessity of CIM and AFMM, showing strong generalization to unseen distortions with minimal distortion augmentation.

Abstract

Existing deep image watermarking methods follow a fixed embedding-distortion-extraction pipeline, where the embedder and extractor are weakly coupled through a final loss and optimized in isolation. This design lacks explicit collaboration, leaving no structured mechanism for the embedder to incorporate decoding-aware cues or for the extractor to guide embedding during training. To address this architectural limitation, we rethink deep image watermarking by reformulating embedding and extraction as explicitly collaborative components. To realize this reformulation, we introduce a Collaborative Interaction Mechanism (CIM) that establishes direct, bidirectional communication between the embedder and extractor, enabling a mutual-teacher training paradigm and coordinated optimization. Built upon this explicitly collaborative architecture, we further propose an Adaptive Feature Modulation Module (AFMM) to support effective interaction. AFMM enables content-aware feature regulation by decoupling modulation structure and strength, guiding watermark embedding toward stable image features while suppressing host interference during extraction. Under CIM, the AFMMs on both sides form a closed-loop collaboration that aligns embedding behavior with extraction objectives. This architecture-level redesign changes how robustness is learned in watermarking systems. Rather than relying on exhaustive distortion simulation, robustness emerges from coordinated representation learning between embedding and extraction. Experiments on real-world and AI-generated datasets demonstrate that the proposed method consistently outperforms state-of-the-art approaches in watermark extraction accuracy while maintaining high perceptual quality, showing strong robustness and generalization.

MT-Mark: Rethinking Image Watermarking via Mutual-Teacher Collaboration with Adaptive Feature Modulation

TL;DR

<3-5 sentence high-level summary> MT-Mark tackles robustness in image watermarking by replacing the traditional fixed embedder–extractor pipeline with an explicitly collaborative framework. It introduces a Collaborative Interaction Mechanism (CIM) that enables bidirectional, mutual-teacher communication between embedding and extraction, guided by Adaptive Feature Modulation Modules (AFMM) for content-aware feature regulation. The approach shifts robustness learning from distortion enumeration to coordinated representation learning, achieving superior watermark extraction accuracy and perceptual quality on real and AI-generated datasets. Ablation studies confirm the necessity of CIM and AFMM, showing strong generalization to unseen distortions with minimal distortion augmentation.

Abstract

Existing deep image watermarking methods follow a fixed embedding-distortion-extraction pipeline, where the embedder and extractor are weakly coupled through a final loss and optimized in isolation. This design lacks explicit collaboration, leaving no structured mechanism for the embedder to incorporate decoding-aware cues or for the extractor to guide embedding during training. To address this architectural limitation, we rethink deep image watermarking by reformulating embedding and extraction as explicitly collaborative components. To realize this reformulation, we introduce a Collaborative Interaction Mechanism (CIM) that establishes direct, bidirectional communication between the embedder and extractor, enabling a mutual-teacher training paradigm and coordinated optimization. Built upon this explicitly collaborative architecture, we further propose an Adaptive Feature Modulation Module (AFMM) to support effective interaction. AFMM enables content-aware feature regulation by decoupling modulation structure and strength, guiding watermark embedding toward stable image features while suppressing host interference during extraction. Under CIM, the AFMMs on both sides form a closed-loop collaboration that aligns embedding behavior with extraction objectives. This architecture-level redesign changes how robustness is learned in watermarking systems. Rather than relying on exhaustive distortion simulation, robustness emerges from coordinated representation learning between embedding and extraction. Experiments on real-world and AI-generated datasets demonstrate that the proposed method consistently outperforms state-of-the-art approaches in watermark extraction accuracy while maintaining high perceptual quality, showing strong robustness and generalization.
Paper Structure (15 sections, 19 equations, 5 figures, 5 tables)

This paper contains 15 sections, 19 equations, 5 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methodology
  4. Overview
  5. Collaborative Interaction Mechanism
  6. Adaptive Feature Modulation Module
  7. Watermark Embedder
  8. Distortion Layer
  9. Watermark Extractor
  10. Experiment and Evaluation
  11. Implementation Details and Metrics
  12. Perceptual Quality
  13. Robustness and Generalization
  14. Ablation Study
  15. Conclusion

Figures (5)

  • Figure 1: The structures of different image watermarking architectures. (a) A standard embedding–distortion–extraction pipeline. (b) MT-Mark, an explicitly collaborative architecture centered on the Collaborative Interaction Mechanism (CIM), which establishes mutual-teacher coordination between watermark embedding and extraction; Adaptive Feature Modulation Modules (AFMMs) provide the feature-level modulation interface for this collaboration. Black arrows denote data flow, and blue arrows denote gradient flow.
  • Figure 2: Overall framework of the proposed watermarking method MT-Mark. The top-left illustrates the overall pipeline, including watermark embedder, distortions, and extractor. The bottom-left presents detailed structures of key components, including the AFMM module and distortions layer. The right part shows the detailed structure of the CIM.
  • Figure 3: Visual comparison on a DIV2K sample. Top: original and watermarked images (CVVDP score above each). Middle: difference maps. Bottom: CVVDP heatmaps (blue/green = imperceptible, pink = visible).
  • Figure 4: Qualitative visualization of watermark-focused regions under geometric transformations. (a) Original watermarked image. (b) Watermarked image after horizontal flip (HFlip). (c) Watermarked image after a composite geometric transformation consisting of HFlip followed by an x-axis shear of $30^\circ$. Despite significant spatial reconfiguration, the network consistently focuses on regions carrying watermark information, enabling reliable watermark extraction.
  • Figure 5: Visual ablation study on the CIM.