Bayesian Estimation of Differential Privacy

Authors: Santiago Zanella-Beguelin, Lukas Wutschitz, Shruti Tople, Ahmed Salem, Victor Rühle, Andrew Paverd, Mohammad Naseri, Boris Köpf, Daniel Jones

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

Reproducibility Variable Result LLM Response
Research Type Experimental We implement an end-to-end system for privacy estimation that integrates our approach and state-of-the-art membership inference attacks, and evaluate it on text and vision classification tasks.
Researcher Affiliation Collaboration 1Microsoft, Cambridge, UK 2University College London, London, UK.
Pseudocode Yes Algorithm 3 in Appendix A.1 shows pseudocode for this challenge point selection procedure. (...) A.1. Omitted algorithms Algorithm 3: Select Worst
Open Source Code Yes The implementation of the core Bayesian estimation method used to produce the results reported in the paper can be found at https://aka.ms/privacy-estimates.
Open Datasets Yes We evaluate our system for privacy estimation on text (SST-2) and vision (CIFAR-10) classification tasks. (...) CIFAR-10 (Krizhevsky, 2009) (...) SST 2 (Socher et al., 2013)
Dataset Splits No CIFAR-10 (...) consists of 60 000 labeled (50 000 training, 10 000 test) labeled images. (...) SST 2 (...) consisting of 67 349 training samples and 1821 test samples.
Hardware Specification No We implement the system as an Azure ML pipeline, allowing for an efficient utilization of large GPU clusters, but the system can make use of more modest resources and its design is generic enough to be ported to any other ML framework.
Software Dependencies No The resulting integral in Equation (3) cannot be expressed in analytical form so we approximate it numerically using SciPy s dblquad, based on QUADPACK s qagse. (...) We implement the end-to-end pipeline depicted in Figure 6 in Azure ML. (...) our original implementation used open-source libraries (Ray Tune).
Experiment Setup Yes We use a 4-layer CNN with 974 K parameters and tanh activations with average pooling and max pooling units, which we train for 50 epochs. (...) We fine-tune a RoBERTa base model with a classification head for 3 epochs (Liu et al., 2021).