A Randomized Approach to Tight Privacy Accounting

Authors: Jiachen (Tianhao) Wang, Saeed Mahloujifar, Tong Wu, Ruoxi Jia, Prateek Mittal

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental An empirical evaluation shows the proposed EVR paradigm improves the utility-privacy tradeoff for privacy-preserving machine learning. (Abstract) / 6 Numerical Experiments: In this section, we conduct numerical experiments to illustrate (1) EVR paradigm with MC verifiers enables a tighter privacy analysis, and (2) MC accountant achieves state-of-the-art performance in privacy parameter estimation.
Researcher Affiliation Academia Jiachen T. Wang Princeton University tianhaowang@princeton.edu Saeed Mahloujifar Princeton University sfar@princeton.edu Tong Wu Princeton University tongwu@princeton.edu Ruoxi Jia Virginia Tech ruoxijia@vt.edu Prateek Mittal Princeton University pmittal@princeton.edu
Pseudocode Yes Algorithm 1 Estimate-Verify-Release (EVR) Framework / Algorithm 2 DPV(M, ε, δest) with Simple MC Estimator and Offset Parameter . / Algorithm 3 MC Accountant for εY (δ).
Open Source Code No The paper does not include any explicit statements or links indicating that the source code for the described methodology is open-source or publicly available.
Open Datasets Yes To further underscore the superiority of the EVR paradigm in practical applications, we illustrate the privacy-utility tradeoff curve when finetuning on CIFAR100 dataset with DP-SGD. (Section 6.1) / Image Net-pretrained BEi T [4] on CIFAR100 (Figure 4, H.2.1, H.2.2)
Dataset Splits No The paper mentions using CIFAR100 and ImageNet for training but does not explicitly state the specific training/validation/test splits (e.g., percentages or sample counts) needed for reproduction. It refers to following existing training procedures but does not detail the splits.
Hardware Specification Yes The runtime is estimated on an NVIDIA A100-SXM4-80GB GPU. (Figure 3 (b) caption) / when using a NVIDIA A100-SXM4-80GB GPU (Section H.1)
Software Dependencies No Section H.1 mentions 'Py Torch s CUDA functionality' and 'torch.randn' but does not specify exact version numbers for PyTorch, CUDA, or other key software components.
Experiment Setup Yes For DP-GD training, we set σ as 28.914, clipping norm as 1, learning rate as 2, and we train for at most 60 iterations, and we only finetune the last layer on CIFAR-100. (Section H.2.1) / For DP-SGD training, we set σ as 5.971, clipping norm as 1, learning rate as 0.2, momentum as 0.9, batch size as 4096, and we train for at most 360 iterations (30 epochs). (Section H.2.2) / For Figure 1 & Figure 3, the FFT-based method has hyperparameter being set as εerror = 10 3, δerror = 10 10. (Section H.2.1)