CITED

This repository provides the official implementation of the paper: "CITED: A Decision Boundary-Aware Signature for GNNs Towards Model Extraction Defense"

Citation

If you find this work helpful, please consider citing our paper. We sincerely appreciate your support and interest in our work.

@misc{shen2026cited,
      title={CITED: A Decision Boundary-Aware Signature for GNNs Towards Model Extraction Defense}, 
      author={Bolin Shen and Md Shamim Seraj and Zhan Cheng and Shayok Chakraborty and Yushun Dong},
      year={2026},
      eprint={2602.20418},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2602.20418}, 
}

Quick Start

Using uv to install the environment

uv sync

Run Experiments

Evaluation on Original Task

Before running the script, please manually update:

• exp1.py: adjust experiment settings (e.g., dataset, model, evaluation level).
• utils/config.py: modify necessary hyperparameters for training and finetuning.

Ensure the configurations reflect your desired setup.

Execute the following command:

python exp1.py

Output Example:

========== Trial 1 ==========
model load:  ./output/target/Pubmed_GraphSAGE.pth
[Defense] Test Acc: 0.7900 | F1: 0.7801 | AUROC: 0.9260
[Defense]>>>>>>>>>>>>>>>>>>>>>Signature Acc: 1.0000
Defense Model saved to ./output/defense/Pubmed_GraphSAGE_CITED.pth

========== Trial 2 ==========
model load:  ./output/target/Pubmed_GraphSAGE.pth
[Defense] Test Acc: 0.7940 | F1: 0.7844 | AUROC: 0.9270
[Defense]>>>>>>>>>>>>>>>>>>>>>Signature Acc: 1.0000
Defense Model saved to ./output/defense/Pubmed_GraphSAGE_CITED.pth

========== Trial 3 ==========
model load:  ./output/target/Pubmed_GraphSAGE.pth
[Defense] Test Acc: 0.7910 | F1: 0.7811 | AUROC: 0.9269
[Defense]>>>>>>>>>>>>>>>>>>>>>Signature Acc: 1.0000
Defense Model saved to ./output/defense/Pubmed_GraphSAGE_CITED.pth

========== Summary ==========
ACC: 79.17 ± 0.17
F1-score: 78.19 ± 0.18
AUROC: 92.66 ± 0.05

Results will be saved automatically to the results directory.

Performance of Ownership Verification

Before running the script, please manually update:

• exp2.py: adjust experiment settings (e.g., dataset, model, evaluation level).
• utils/config.py: modify necessary hyperparameters for training and finetuning.

Ensure the configurations reflect your desired setup.

Execute the following command:

python exp2.py --defense CITED --data cora --device 0

Output Example:

[Result - Trial 1] ARUC = 0.3993, AUC = 0.7733
[Result - Trial 2] ARUC = 0.4780, AUC = 0.9733
[Result - Trial 3] ARUC = 0.3460, AUC = 0.8933
Experiment Configuration: level=label | variant_num=15 | model_name=gcn2 | hidden_dim=128 | train_epochs=200 | lr=0.001 | weight_decay=1e-05 | query_ratio=0.5 | fixed_seed=42 | device=cuda:0 | defense_name=CITED | cited_boundary_ratio=0.05 | cited_signature_ratio=0.1 | finetune_epochs=20 | ds_name=cora | threshold=0.2
========== Summary ==========
ARUC: 0.4078 ± 0.0542
AUC : 0.8800 ± 0.0822
[Saved] Result saved to: ./results/Res_CITED_gcn2_cora_label.npz

Results will be saved automatically to the results directory.

Ablation Study

Before running the script, please manually update:

• exp4.py: adjust experiment settings (e.g., dataset, model, evaluation level).
• utils/config.py: modify necessary hyperparameters for training and finetuning.

Ensure the configurations reflect your desired setup.

Execute the following command:

python exp4.py --defense CITED --data cora --device 0

Output Example:

[Result - Trial 1] ARUC = 0.5780, AUC = 0.9200
[Result - Trial 2] ARUC = 0.6960, AUC = 1.0000
[Result - Trial 3] ARUC = 0.4980, AUC = 0.8800
Experiment Configuration: level=label | variant_num=5 | model_name=gcn | hidden_dim=128 | train_epochs=200 | finetune_epochs=20 | lr=0.001 | weight_decay=1e-05 | query_ratio=0.5 | cited_choice=all | fixed_seed=42 | device=cuda:0 | defense_name=CITED | cited_boundary_ratio=0.05 | cited_signature_ratio=0.1 | ds_name=cora | threshold=0.2
========== Summary - all ==========
ARUC: 0.5907 ± 0.0813
AUC : 0.9333 ± 0.0499
[Saved] Result saved to: ./results/Res_exp4_CITED_gcn_cora_label_all.npz

Results will be saved automatically to the results directory.

Visualization

Visualize ownership indicators degradation

python viz_wm.py

Visualize efficiency

python viz_effi.py

Visualize ARUC

python viz_aruc.py