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 [1].
Private Set Generation with Discriminative Information
Authors: Dingfan Chen, Raouf Kerkouche, Mario Fritz
NeurIPS 2022 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We first compare private set generation (PSG) with existing DP generative models on standard classification benchmarks including MNIST [17] and Fashion MNIST [39]. |
| Researcher Affiliation | Academia | Dingfan Chen Raouf Kerkouche Mario Fritz CISPA Helmholtz Center for Information Security EMAIL |
| Pseudocode | Yes | Algorithm 1: Private Set Generation (PSG) |
| Open Source Code | Yes | Our code has been open-sourced to facilitate research in the related field. |
| Open Datasets | Yes | We first compare private set generation (PSG) with existing DP generative models on standard classification benchmarks including MNIST [17] and Fashion MNIST [39]. |
| Dataset Splits | Yes | number of samples per class (spc) {10, 20} (Section 5, Setup) and full corresponds to 6000 samples per class (Table 1 caption). |
| Hardware Specification | Yes | We also acknowledge Max Planck Institute for Informatics and "Helmholtz AI computing resources" (HAICORE) for providing computing resources. (Acknowledgments) and Did you include the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)? [Yes] In supplementary (Checklist, point 3d) |
| Software Dependencies | No | The paper describes the architecture (Conv Net, Instance Normalization, Re LU, Avg Pooling, FC layer) and initialization methods (Kaiming initialization) used. However, it does not provide specific version numbers for any software dependencies or libraries (e.g., PyTorch, TensorFlow, Python version). |
| Experiment Setup | Yes | We list below the default hyperparameters used for the main experiments and refer to the supplementary material for more details: Clipping bound C = 0.1, R =1000 for ε = 10 (and 200 for ε = 1), number of samples per class (spc) {10, 20}, K = 10, T = 10 for spc=10 (and =20 for spc=20). |