Differentially Private Sharpness-Aware Training

Authors: Jinseong Park, Hoki Kim, Yujin Choi, Jaewook Lee

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

Reproducibility Variable Result LLM Response
Research Type Experimental Our experimental results demonstrate that the proposed method improves the performance of deep learning models with DP from both scratch and finetuning. Code is available at https://github. com/jinseong P/DPSAT.
Researcher Affiliation Academia 1Department of Industrial Engineering, Seoul National University, 1 Gwanak-ro, Gwanak-gu, Seoul 08826, South Korea. 2Institute of Engineering Research, Seoul National University, 1 Gwanak-ro, Gwanak-gu, Seoul 08826, South Korea.
Pseudocode Yes Algorithm 1 DP-SAT Input: Initial parameter w0, learning rate η, radius ρ, clipping threshold C, variance σ2 from Proposition 2.2, and small τ to prevent zero division. Output: Final parameter w T .
Open Source Code Yes Code is available at https://github. com/jinseong P/DPSAT.
Open Datasets Yes For the empirical results trained from scratch, we evaluate the performance of our method on three commonly used benchmarks for differentially private deep learning: MNIST, Fashion MNIST, CIFAR-10, and SVHN.
Dataset Splits Yes The training data for each dataset was partitioned into training and test sets with a ratio of 0.8:0.2, and the test accuracy was averaged over 5 different random seeds for each dataset.
Hardware Specification Yes All experiments are conducted using the Py Torch-based libraries (Kim, 2020; Yousefpour et al., 2021) with Python on four NVIDIA Ge Force RTX 3090 GPUs.
Software Dependencies No The paper mentions 'Py Torch-based libraries' and 'Python' but does not provide specific version numbers for any software dependencies, which is required for reproducibility.
Experiment Setup Yes We use SGD as a base optimizer with a momentum of 0.9 and a learning rate of 2.0, without any learning rate decay, as mentioned in (Cheng et al., 2022). We conducted a hyperparameter search on ρ = {0.005, 0.01, 0.02, 0.03, 0.05, 0.1}, and the privacy broken probability δ = 10 5 in DP training. Table 5. Hyperparameters for training on MNIST, Fashion MNIST, CIFAR-10, and SVHN.