LabelDP-Pro: Learning with Label Differential Privacy via Projections

Authors: Badih Ghazi, Yangsibo Huang, Pritish Kamath, Ravi Kumar, Pasin Manurangsi, Chiyuan Zhang

ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We show that such projection-based algorithms can be made practical and that they improve on the state-of-the-art for label DP training in the high-privacy regime. We complement our empirical evaluation with theoretical results shedding light on the efficacy of our method through the lens of bias-variance trade-offs.
Researcher Affiliation Collaboration Badih Ghazi Google Research Yangsibo Huang Princeton University Pritish Kamath Google Research Ravi Kumar Google Research Pasin Manurangsi Google Research Chiyuan Zhang Google Research
Pseudocode Yes Algorithm 1: Label DP-Pro
Open Source Code No The paper does not provide a specific link or explicit statement about releasing the source code for the methodology described.
Open Datasets Yes We compare our Label DP-Pro on four benchmark datasets commonly used in Label DP learning evaluations: MNIST (Le Cun et al., 1998), k MNIST (Clanuwat et al., 2018), Fashion MNIST (Xiao et al., 2017), and CIFAR-10 (Krizhevsky et al., 2009).
Dataset Splits No The paper does not provide explicit train/validation/test split percentages or absolute counts for all datasets. For LP-2ST baseline, it mentions: 'We use 40% of the data for the first stage and 60% for the second.' but this is not a general validation split.
Hardware Specification Yes Results for MNIST, CIFAR-10, and CIFAR10-Self SL are reported on a single NVIDIA Tesla-P100 GPU, while results for Criteo Attribution are reported on a single NVIDIA A100 GPU.
Software Dependencies No The paper mentions 'all implemented in JAX' and refers to PyTorch and TensorFlow for autodiff primitives, but does not provide specific version numbers for these software dependencies.
Experiment Setup Yes Table 8: Hyper-parameters for our experimental results presented in the main paper. Optimizer: SGD, RMSprop; Learning rate: various values; # Epochs: various values; Batch size: various values; Clipping norm (DP-SGD & Label DP-Pro): various values.