Unveiling Adversarially Robust Graph Lottery Tickets
Authors: Subhajit Dutta Chowdhury, Zhiyu Ni, Qingyuan Peng, Souvik Kundu, Pierluigi Nuzzo
TMLR 2024 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Evaluations conducted on various benchmarks, considering attacks such as PGD, Meta Attack, PR-BCD, GR-BCD, and adaptive attack, demonstrate that ARGS can significantly improve the robustness of the generated GLTs, even when subjected to high levels of sparsity. Our proposal is evaluated across various GNN architectures on both homophilic (Cora, citeseer, Pub Med, OGBN-Ar Xiv, OGBN-Products) and heterophilic (Chameleon, Squirrel) graphs attacked by the PGD attack, the Meta Attack, the PR-BCD attack (Geisler et al., 2021), and the GR-BCD attack, for the node classification task. |
| Researcher Affiliation | Collaboration | Subhajit Dutta Chowdhury EMAIL Department of Electrical and Computer Engineering University of Southern California, Los Angeles Zhiyu Ni EMAIL Department of Electrical and Computer Engineering University of Southern California, Los Angeles Qingyuan Peng EMAIL Department of Electrical and Computer Engineering University of Southern California, Los Angeles Souvik Kundu EMAIL Intel Labs, San Diego Pierluigi Nuzzo EMAIL Department of Electrical Engineering and Computer Sciences, University of California, Berkeley Department of Electrical and Computer Engineering, University of Southern California, Los Angeles |
| Pseudocode | Yes | Algorithm 1 Adversarially Robust Graph Sparsification |
| Open Source Code | No | The paper mentions using third-party libraries like "Deep Robust" (Li et al., 2020) and "Pytorch-Geometric" (Fey & Lenssen, 2019) to perform attacks, but there is no explicit statement or link indicating that the authors' implementation code for the ARGS methodology itself is openly available. |
| Open Datasets | Yes | We use seven benchmark datasets, namely, Cora, Citeseer, Pub Med, OGBN-ar Xiv, OGBN-Produts, Chameleon, and Squirrel, to evaluate the efficacy of ARGS. Details about the datasets are summarized in Table 6. For OGBN-Ar Xiv and OGBN-Products we follow the data split setting of the Open Graph Benchmark (OGB) (Hu et al., 2020). |
| Dataset Splits | Yes | In the case of Cora, Citeseer, and Pub Med, 10% of the data constitutes the train set, 10% of the data constitutes the validation set, while the test set is the remaining 80%. For Chameleon and Squirrel, we keep the same data split settings as Chien et al. (2020). For OGBN-Ar Xiv and OGBN-Products we follow the data split setting of the Open Graph Benchmark (OGB) (Hu et al., 2020). |
| Hardware Specification | Yes | All the experiments are conducted on an NVIDIA Tesla A100 (80-GB GPU). |
| Software Dependencies | No | The paper mentions using "Deep Robust" (Li et al., 2020) and "Pytorch-Geometric" (Fey & Lenssen, 2019) but does not provide specific version numbers for these or any other software dependencies crucial for replication. |
| Experiment Setup | Yes | The graph sparsity pg and model sparsity pθ are 5% and 20% unless otherwise stated. The value of β is chosen from {0.01, 0.1, 1, 10} while the value of α, γ, η, and ζ is 1 by default. We use the Adam optimizer for training the GNNs. In each pruning round, the number of epochs to update the masks is by default 200, using early stopping. The 2-layer MLP used for predicting the pseudo-labels of the test nodes has by default hidden dimension of 1024 unless otherwise mentioned. For GCN, the values of λ1 and λ2 are both 10^-2 for Cora and Citeseer, while for Pub Med they are 10^-6 and 10^-3, respectively. The value of the learning rate is 8*10^-3 and that of the weight decay is 8*10^-5 for the Cora dataset. For Citeseer and Pub Med, the learning rate is 10^-2 and the weight decay is 5*10^-4. For the different datasets, we use a dropout of 0.5. |