Confidential-DPproof: Confidential Proof of Differentially Private Training

Authors: Ali Shahin Shamsabadi, Gefei Tan, Tudor Ioan Cebere, Aurélien Bellet, Hamed Haddadi, Nicolas Papernot, Xiao Wang, Adrian Weller

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

Reproducibility Variable Result LLM Response
Research Type Experimental In experiments on CIFAR-10, Confidential DPproof trains a model achieving state-of-the-art 91% test accuracy with a certified privacy guarantee of (ε = 0.55, δ = 10 5)-DP in approximately 100 hours. and 4 EXPERIMENTAL EVALUATION
Researcher Affiliation Collaboration 1 Brave Software, 2 Northwestern University, 3 Inria, Univ Lille, 4 Inria, Univ Montpellier, 5 Imperial College London, 6 Vector Institute, 7 University of Toronto, 8 University of Cambridge, 9 The Alan Turing Institute
Pseudocode Yes Algorithm 1: Differentially Private Stochastic Gradient Descent (DP-SGD), Algorithm 2: Unbiased randomness seed commitment, Algorithm 3: Zero Knowledge Proof of DP-SGD Training
Open Source Code Yes We have described our framework in details and our code is available at https://github.com/brave-experiments/Confidential-DPproof and https://github.com/ Gefei-Tan/Confidential-DPProof-zk.
Open Datasets Yes We consider two common datasets (see Appendix A), for DP-SGD training benchmarking Tramer & Boneh (2021); Papernot et al. (2021); De et al. (2022); Shamsabadi & Papernot (2023): CIFAR-10 and MNIST.
Dataset Splits No The paper mentions training iterations and evaluates performance on a test set but does not explicitly state the use or size of a validation dataset split.
Hardware Specification Yes Our experiment is conducted on two Amazon Ec2 m1.xlarge machines (ARM machines), representing the prover and auditor. and across three different machines with different CPUs: ARM, intel, and AMD (m1.xlarge, m7i.2xlarge, m7a.2xlarge, all with 16GB of RAM)
Software Dependencies No The paper mentions EMP-toolkit (Wang et al., 2016) and a Python implementation (Tramer & Boneh, 2021) but does not provide specific version numbers for Python or any other software libraries.
Experiment Setup Yes Prover and auditor agree on the DP-SGD training algorithm and specific values for its hyperparameters including minibatch size, noise multiplier, clipping norm, learning rate, number of iterations and loss function. and Table 2 details: Clipping norm 0.1, Noise multiplier 3 for CIFAR-10 and Clipping norm 0.1, Noise multiplier 3.32 for MNIST.