Bounding training data reconstruction in DP-SGD

Authors: Jamie Hayes, Borja Balle, Saeed Mahloujifar

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

Reproducibility Variable Result LLM Response
Research Type Experimental We now evaluate both our upper bounds for reconstruction success and our empirical privacy attacks (which gives us lower bounds on reconstruction success). We show that our attack has a success probability nearly identical to the bound given by our theory.
Researcher Affiliation Collaboration Jamie Hayes Google Deep Mind jamhay@google.com Saeed Mahloujifar Meta AI saeedm@meta.com Borja Balle Google Deep Mind bballe@google.com
Pseudocode Yes Algorithm 1 Estimating γ; Algorithm 2 Prior-aware attack; Algorithm 3 Improved prior-aware attack
Open Source Code No The paper does not provide an explicit statement or link indicating that the source code for the described methodology is publicly available.
Open Datasets Yes We now compare the success of model-based and gradient-based reconstruction attacks against classification models trained with DP-SGD on MNIST and CIFAR-10.
Dataset Splits No The paper specifies training dataset sizes ('training set size is |D{z }| = 1, 000' for MNIST and 'training set size is |D{z }| = 500' for CIFAR-10) but does not provide explicit training, validation, and test splits (e.g., percentages or sample counts for each split).
Hardware Specification Yes Setting T = 1 and using a 2.3 GHz 8-Core Intel Core i9 CPU it takes 0.002s to estimate with 10,000 samples.
Software Dependencies No The paper mentions using ML models like 'MLP' and 'Wide Res Net model' but does not specify software dependencies with version numbers (e.g., 'PyTorch 1.9' or 'TensorFlow 2.x').
Experiment Setup Yes We refer to Appendix A for experimental details. For each ϵ, we select the learning rate by sweeping over a range of values between 0.001 and 100; we do not use any momentum in optimization. We set C = 0.1, δ = 10 5 and adjust the noise scale σ for a given target ϵ. Appendix A includes Table 1: Hyperparameter settings for each experiment.