Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Data-Free Model Extraction for Black-box Recommender Systems via Graph Convolutions
Authors: Zeyu Wang, Yidan Song, Shihao Qin, Shanqing Yu, Yujin Huang, Qi Xuan, Xin Zheng
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results on various datasets and victim models demonstrate the superiority of our attack in data-free scenarios (e.g., surpassing PTQ data-require methods with 17.4% improvement on Light GCN). Code is available: https://github.com/Vencent-Won/DBGRME.git. (...) 5 Experiments: In this section, we conduct a comprehensive experimental evaluation of the proposed DBGRME to validate its effectiveness in black-box recommender model extraction and its generalization performance in a data-free scenario. The experiments consist of four key aspects: First, we perform a comparative study to assess the extraction performance of DBGRME across different victim models and datasets, comparing it with existing methods. Second, we conduct a query budget analysis to investigate the impact of query limitations on the performance of the surrogate model. Third, we performed an ablation study to analyze the contribution of each key module. Finally, we further analyze the behavior of the generator to gain deeper insights into its role in model extraction. |
| Researcher Affiliation | Academia | 1Zhejiang University of Technology 2Griffith University 3Binjiang Institute of Artificial Intelligence 4The University of Melbourne EMAIL,EMAIL, EMAIL |
| Pseudocode | Yes | A Algorithm of DBGRME Algorithm 1 Alogrithm of DBGRME Input: Query budget Q, iteration times of generator iter_G, iteration times of surrogate model iter_C, the synthesized user number n , victim model fθ Output: surrogate model fδ |
| Open Source Code | Yes | Code is available: https://github.com/Vencent-Won/DBGRME.git. |
| Open Datasets | Yes | 5.1 Experimental Setup Datasets. We use four popular recommendation datasets to evaluate our methods: ML-100K, ML1M [27], Yelp [28] and Gowalla [29]. |
| Dataset Splits | Yes | We follow the preprocessing in [26] to process the rating data into implicit feedbacks, and split them into 8:2 as training set and testing set for victim model training. Also, the testing set will be used to test the model extraction attack performance. |
| Hardware Specification | Yes | All experiments are implemented in Py Torch2 on an Ubuntu Server equipped with Intel(R) Core(TM)594 i7-8700K CPU, and 2 NVIDIA Ge Force GTX 1080 Ti (with 11GB memory each). |
| Software Dependencies | No | All experiments are implemented in Py Torch2 on an Ubuntu Server equipped with Intel(R) Core(TM)594 i7-8700K CPU, and 2 NVIDIA Ge Force GTX 1080 Ti (with 11GB memory each). (...) 2https://pytorch.org/ |
| Experiment Setup | Yes | B.3 Patameter Setting The parameters of recommender model extraction involve four parts: victim model, generator, surrogate model, and training. We report the detailed parameter space in Table 6. (...) Table 7: Hyperparameter of victim models. (...) Table 8: Hyperparameters space considered for the DBGRME selection. |