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..
Multi-Class Support Vector Machine with Differential Privacy
Authors: Jinseong Park, Yujin Choi, Jaewook Lee
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Empirical results demonstrate that our approach surpasses existing DP-SVM methods in multi-class scenarios. |
| Researcher Affiliation | Academia | Jinseong Park1, Yujin Choi2, Jaewook Lee2* 1Korea Institute for Advanced Study 2Seoul National University |
| Pseudocode | No | The paper describes the weight perturbation and gradient perturbation methods using mathematical formulations and descriptive text, but does not include any explicitly labeled 'Pseudocode' or 'Algorithm' blocks. |
| Open Source Code | Yes | Code implementation: https://github.com/Jinseong P/private_multiclass_svm |
| Open Datasets | Yes | We used multi-class classification datasets from the University of California at Irvine (UCI) repository [27] for various data types: Cornell (CS web pages), Dermatology (clinical skin records), HHAR (wearable activity sensors), ISOLET (spoken alphabet), USPS (hand-written digits), and Vehicle (vehicle silhouettes). |
| Dataset Splits | No | The paper uses multi-class classification datasets but does not explicitly state the training, validation, and test splits (e.g., percentages, sample counts, or specific cross-validation strategy) used for the experiments. |
| Hardware Specification | Yes | All experiments were run on an Intel(R) Xeon(R) CPU E5-2680 v3 @ 2.50GHz and a single NVIDIA Ge Force RTX 4090. |
| Software Dependencies | No | We utilized the SVM packages in Sklearn [28] for weight perturbation, and the Opacus [29] for gradient descent methods based on Pytorch [30]. The paper mentions software packages but does not provide specific version numbers for them. |
| Experiment Setup | Yes | For privacy budget, we fixed δ = 10 5 on various ϵ. We reported the mean and standard deviation on each setting, where we used 20 runs for weight perturbation and 5 runs for gradient perturbations. We performed a grid search on each method to find the well-performing one on ϵ = 4, and used the obtained parameters for each model on other epsilons. We searched on C/n for weight perturbation, and learning rate ηt, gradient steps T, and fixed the clipping R = 1 for gradient methods. We further utilize the min-max scaler for weight perturbation to bound the input sensitivity to 1 and thus calculate the sensitivity of w easily. We utilize a Poisson sub-sampling batch size of 128 for gradient methods. |