FedGMark: Certifiably Robust Watermarking for Federated Graph Learning
Authors: Yuxin Yang, Qiang Li, Yuan Hong, Binghui Wang
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments validate the promising empirical and provable watermarking performance of Fed GMark. We evaluate Fed GMark on four real-world graph datasets (MUTAG, PROTEINS, DD, and COLLAB) and three Fed GL models including Fed-GIN, Fed-GSAGE, and Fed-GCN, whose base GL models are GIN [Xu et al., 2019], GSAGE [Hamilton et al., 2017], and GCN [Kipf and Welling, 2017], respectively. Extensive experimental results show Fed GMark achieves high main accuracy and watermark accuracy under no attacks and watermark removal attacks, high certified watermark accuracy, and significantly outperforms the existing method. |
| Researcher Affiliation | Academia | Yuxin Yang1,2 Qiang Li1 Yuan Hong3 Binghui Wang2 1College of Computer Science and Technology, Jilin University, Changchun, Jilin, China 2Department of Computer Science, Illinois Institute of Technology, Chicago, Illinois, USA 3School of Computing, University of Connecticut, Storrs, Connecticut, USA |
| Pseudocode | Yes | Algorithm 1 The training process of Fed GMark |
| Open Source Code | Yes | Source code is available at: https://github.com/Yuxin104/Fed GMark. |
| Open Datasets | Yes | We evaluate our Fed GMark on four real-world graph datasets for graph classification: MUTAG [Debnath et al., 1991], PROTEINS [Borgwardt et al., 2005], DD [Dobson and Doig, 2003], and COLLAB [Yanardag and Vishwanathan, 2015]. Details about the statistics of those datasets are shown in Table 6 in Appendix C. |
| Dataset Splits | Yes | Table 6: Statistics of datasets. Datasets #Graphs #Classes Avg. #Node Avg. # Edge #Training graphs #Testing graphs 1 2 3 MUTAG 188 2 17.93 19.80 83 42 42 21 PROTEINS 1110 2 37.72 70.35 440 300 220 150 DD 950 2 208.3 518.76 330 303 165 152 COLLAB 4981 3 73.49 2336.66 517 1589 1215 258 794 608 |
| Hardware Specification | Yes | We implement our method using one NVIDIA Ge Force GTX 1080 Ti GPU. |
| Software Dependencies | No | The paper does not provide specific software dependencies with version numbers (e.g., Python, PyTorch, CUDA versions) for its core methodology. |
| Experiment Setup | Yes | Parameter setting. We implement our method using one NVIDIA Ge Force GTX 1080 Ti GPU. In Fed GL, we use T = 40 clients in total and train the model 200 iterations. The server randomly selects 50% clients in each iteration. We define the target label of watermarking graphs as 1, and each participating client randomly selects 10% graphs with labels not 1 as the watermarking graphs. We extensively validate the effectiveness and robustness of the proposed watermarking method with the following hyperparameters details: the number of submodels S = {4, 8, 16}, the number watermarked clients Tw = {5, 10, 20} (both S and Tw are halved on MUTAG due to less data), the watermarked nodes nw = {3, 4, 5}, and the number of perturbed layers r = {1, , 5} in the layer-perturbation attack. By default, we set S = 4, Tw = 10, nw = 4, r = 1. While studying the impact of a hyperparameter, we fix the others as the default value. |