Unified Mechanism-Specific Amplification by Subsampling and Group Privacy Amplification

Authors: Jan Schuchardt, Mihail Stoian, Arthur Kosmala, Stephan Günnemann

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental evaluation demonstrates that our tight mechanism-specific guarantees outperform both tight mechanism-agnostic bounds and classic group privacy results.
Researcher Affiliation Academia Jan Schuchardt1, Mihail Stoian2 , Arthur Kosmala1 , Stephan Günnemann1 {j.schuchardt, a.kosmala, s.guennemann}@tum.de, mihail.stoian@utn.de 1Dept. of Computer Science & Munich Data Science Institute, Technical University of Munich 2Dept. of Engineering, University of Technology Nuremberg
Pseudocode No The paper describes algorithms and procedures using prose and mathematical formulations but does not contain any structured pseudocode blocks or figures explicitly labeled 'Algorithm' or 'Pseudocode'.
Open Source Code Yes An implementation will be made available at https://cs.cit.tum.de/daml/group-amplification.
Open Datasets Yes We train a convolutional neural network (2 convolution layers with kernel sizes 3 and 32 / 64 channels, followed by two linear layers with hidden dimension 128) for image classification on MNIST (55000 training, 5000 validation, 10000 test samples).
Dataset Splits Yes We train a convolutional neural network (2 convolution layers with kernel sizes 3 and 32 / 64 channels, followed by two linear layers with hidden dimension 128) for image classification on MNIST (55000 training, 5000 validation, 10000 test samples).
Hardware Specification Yes We conduct all experiments on a set of Xeon E5-2630 v4 CPUs @ 2.2 GHz.
Software Dependencies Yes To perform high-precision quadrature for RDP guarantees, we use the tanh-sinh quadrature implementation from the mpmath library (version 1.3.0.). For PLD accounting and evaluation of ADP guarantees via bisection, we use and extend the dp_accounting library [45] (commit 0b109e959470c43e9f177d5411603b70a56cdc7a)...For conversion from RDP to ADP guarantees, we use the get_privacy_spent method implemented in the Opacus library [73] (version 1.4.1).
Experiment Setup Yes Further details on the experimental setup are provided in Appendix C...We set the gradient clipping norm of DP-SGD [6] to C = 10 4, the Gaussian noise standard deviation to 0.6 C, and the subsampling rate to r = 64 / 55000. The optimizer is ADAM with learning rate 1e 3.