Rethinking Label Poisoning for GNNs: Pitfalls and Attacks
Authors: Vijay Lingam, Mohammad Sadegh Akhondzadeh, Aleksandar Bojchevski
ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We aim to fill this gap by conducting a comprehensive robustness analysis of GNNs under label poisoning, making four key contributions. |
| Researcher Affiliation | Academia | 1 CISPA Helmholtz Center for Information Security 2 University of Cologne |
| Pseudocode | No | The paper describes methods and outlines steps, but it does not include any clearly labeled pseudocode blocks or algorithm sections. |
| Open Source Code | Yes | Furthermore, we make our code publicly available.1https://github.com/Vijay Lingam95/Rethinking Label Poisoning For GNNs |
| Open Datasets | Yes | For the Cora-ML dataset these budgets correspond to [7, 14, 21, 28, 42] training labels out of 2810 in expectation. ... Specifically, we evaluate on the Cora Full dataset (Bojchevski & Günnemann, 2018) that contains 70 output classes, and the OGBN-Ar Xiv dataset (Hu et al., 2020) a medium sized graph with 169K nodes. ... In Table 7, we tabulate dataset statistics. |
| Dataset Splits | Yes | We remedy these pitfalls by using a validation and train set of the same size, and by reporting the mean and standard deviation across 10 splits (more details in B.8). ... In Table 7, we tabulate dataset statistics. We additionally include the train/val/test split statistics for the default and the proposed CV setting. |
| Hardware Specification | Yes | The majority of our experiments are run on an Nvidia k80 GPU with 24GB memory, and the remaining experiments were run using an A100 GPU with 40GB of memory. ... Table 13: Timing analysis. Dataset K-80 GPU A100 GPU |
| Software Dependencies | No | We use Optuna (Akiba et al., 2019) to optimize the hyper-parameters search, and set the number of trials to 20. For learning we use the Adam optimizer (Kingma & Ba, 2015), with 1000 maximum number of epochs and an early stopping patience of 100. While specific software (Optuna, Adam) are mentioned, their version numbers are not provided, preventing full reproducibility of software dependencies. |
| Experiment Setup | Yes | The hyper-parameter ranges we explore for both the attack and defense phases are included in the initial section of B. ... Specifically, for the shared general hyper-parameters we sweep the following ranges: [0.1, 0.01, 0.05, 0.08] for the learning rate, [0.0, 0.005, 0.0005, 0.00005] for weight decay, and [0.3, 0.5, 0.7] for dropout. We fix the hidden dimensions to 64... |