Drgw Achieves Robust Graph Watermarking Via Disentangled Representations for Data Provenance

Researchers are increasingly focused on protecting the intellectual property of structured data, vital to countless web applications, and a new framework called DRGW offers a significant leap forward in this area. Jiasen Li, Yanwei Liu and Zhuoyi Shang, from the Institute of Information Engineering, Chinese Academy of Sciences, alongside Xiaoyan Gu and Weiping Wang, present a novel watermarking technique that overcomes limitations of existing methods by learning disentangled representations. Unlike previous approaches which often compromise watermark transparency and robustness through entangled data or uncontrollable discretisation, DRGW employs an adversarially trained encoder to create invariant structural representations and a statistically independent watermark carrier, ensuring both resilience to attacks and minimal impact on data utility. This innovative approach, coupled with an awareness-based invertible function and a structure-aware editor, demonstrably outperforms current techniques across multiple benchmark datasets, promising a new standard for robust and transparent graph watermarking.

DRGW employs a disentangled representation learning approach, meticulously separating structural information from the watermark carrier, thereby enhancing both robustness and transparency. This innovative network avoids the information loss inherent in traditional discretization processes, a common weakness in previous latent-space watermarking techniques.

Furthermore, the study unveils a structure-aware editor that intelligently resolves latent modifications into discrete graph edits, bolstering robustness against structural perturbations. This editor ensures that even if the graph undergoes changes, the embedded watermark remains resilient and verifiable. Experiments conducted on diverse benchmark datasets consistently demonstrate the superior effectiveness of DRGW compared to existing methods, showcasing its ability to maintain both watermark detectability and graph utility. The research establishes a new paradigm for graph watermarking, moving beyond fragile structure-based approaches and entangled latent representations.
This work opens exciting possibilities for secure data sharing in numerous web applications, including social network analysis, recommendation systems, and knowledge discovery. By providing a reliable mechanism for provenance tracking and copyright protection, DRGW fosters trust in data exchange and mitigates the risks of intellectual property infringement. The team’s innovative disentangled representation learning approach not only enhances watermark robustness but also preserves the functional integrity of the graph, enabling its continued use in downstream tasks without performance degradation.

The team measured Area Under the Curve (AUC) values to quantify performance, achieving consistently high results, averaging 0.998 or higher on clean, unattacked graphs across six categories: Social Networks, Academic Networks, Knowledge Graphs, E-commerce & Recommendation, Web Graphs, and Road Networks. These measurements confirm the framework’s ability to embed and detect watermarks with near-perfect accuracy in the absence of adversarial attacks. Scientists recorded AUC scores of 0.999 on the Academic Networks category, and 0.997 on Web Graphs, showcasing the framework’s consistent performance across varying graph topologies. Furthermore, the structure-aware editor successfully translates this high-fidelity signal into a detectable pattern within the discrete graph structure, as evidenced by the high AUC values.

The framework’s design ensures both robustness and transparency of watermarks, a significant advancement over previous approaches. Tests prove DRGW’s resilience against various attacks, including Edge Flips, Node Deletions, and adversarial perturbations. Specifically, the team measured performance under adversarial attacks, achieving an AUC of 0.958 on Social Networks and 0.964 on Web Graphs, demonstrating strong resistance to malicious modifications. Data shows that DRGW outperforms baseline methods, including Towards and KGMark, in most categories, particularly under structural perturbations. The decision rule, grounded in Neyman-Pearson testing, controls the false positive rate, ensuring reliable watermark detection with a pre-defined error rate β.

Additionally, researchers assessed transparency through Structural Fidelity and Functional Fidelity metrics. Measurements confirm low Flipped Edges percentages, below 0.835 across all categories, indicating minimal structural distortion caused by the watermarking process. The team also evaluated performance drop on link prediction tasks, recording values below 30% in most categories, and achieved high Cosine Similarity of node embeddings, demonstrating preservation of graph functionality. These results highlight DRGW’s ability to embed watermarks without significantly compromising the integrity or utility of the underlying graph data, opening possibilities for secure data sharing and provenance tracking.

Disentangled Representations Safeguard Graph Data Provenance by enabling

This innovative approach utilises disentangled representation learning to address key challenges in existing watermarking methods, namely information entanglement and uncontrollable discretisation. DRGW’s design prioritises graph topology, enabling robustness even with heterogeneous data such as knowledge graphs. The authors acknowledge a limitation in explicitly modelling rich node and edge semantics, suggesting future work could focus on type-aware disentanglement to further enhance performance in complex environments. This research establishes a foundation for securing graph data assets with potential applications in areas like social network analysis and recommendation systems, offering a secure solution that preserves structural and functional integrity under adversarial conditions.